A port loading and unloading efficiency prediction method based on deep learning and internet of things data fusion

Abstract

The application discloses a port loading and unloading efficiency prediction method based on deep learning and Internet of Things data fusion, and relates to the technical field of intelligent logistics, and comprises the following steps: S1, through deploying an Internet of Things terminal and a data transmission channel, synchronously collecting equipment operation state data, container attribute data, yard layout data, meteorological environment data and operation scheduling data; S2, pre-processing the collected data; S3, constructing a CNN-LSTM fusion model based on an attention mechanism and training; the model dynamic adaptation module is arranged, the CNN convolution kernel size and the LSTM hidden layer neuron quantity can be automatically adjusted, different scenes such as a container port and a bulk cargo port can be adapted, for different scales and different equipment configurations of the port, without additional customization and development, the existing Internet of Things terminal and the scheduling system can be directly connected, and the technical landing threshold and cost can be greatly reduced.

Description

Technical Field

[0001] This invention relates to the field of intelligent logistics technology, and in particular to a method for predicting port loading and unloading efficiency based on the fusion of deep learning and Internet of Things data. Background Technology

[0002] As a core hub of the global supply chain, the efficiency of port loading and unloading directly determines the efficiency and cost of international trade. According to statistics from the United Nations Conference on Trade and Development, more than 80% of global goods trade relies on maritime transport, and every 10% increase in loading and unloading efficiency can reduce the cost of the entire supply chain by 3%-5%.

[0003] With the advancement of "smart port" construction, the throughput of domestic ports continues to grow (reaching 340 million TEUs in 2024), but significant technical bottlenecks remain in predicting loading and unloading efficiency. These bottlenecks include: 1. Insufficient data fusion: Focusing primarily on single-type data and failing to integrate multi-source heterogeneous data from "equipment, cargo, environment, and scheduling," resulting in a one-sided prediction dimension; 2. Model structural defects: Using pure LSTM or pure CNN models fails to capture joint "temporal-spatial" features simultaneously and does not strengthen the weight of key factors (such as strong winds); weak generalization ability: Training for specific ports or cargo types leads to an accuracy drop of over 40% when switching scenarios, lacking engineering practicality; 3. Insufficient real-time performance: Complex models (such as Transformer) have inference latency of 50-100 milliseconds, failing to meet the real-time scheduling needs of ports. Ports urgently need accurate prediction tools to support scheduling optimization (increasing berth utilization by 10%), equipment maintenance (reducing failure rate by 5%-8%), and cost control (reducing energy consumption by 40%), while also adapting to dynamic scenarios such as cargo type changes and severe weather. In addition, different ports have different equipment types and cargo structures, so the prediction method needs to be adaptable to different scenarios to avoid a significant drop in accuracy when switching scenarios. To this end, a port loading and unloading efficiency prediction method based on the fusion of deep learning and IoT data is proposed. Summary of the Invention

[0004] The purpose of this invention is to solve the problems in the prior art by proposing a method for predicting port loading and unloading efficiency based on the fusion of deep learning and Internet of Things data.

[0005] A method for predicting port loading and unloading efficiency based on the fusion of deep learning and IoT data includes the following steps: S1. By deploying IoT terminals and data transmission channels, synchronously collect equipment operating status data, container attribute data, yard layout data, meteorological environment data, and operation scheduling data; S2. Preprocess the collected data; S3. Construct and train a CNN-LSTM fusion model based on the attention mechanism; S4. Input the preprocessed data into the trained CNN-LSTM fusion model and output the efficiency prediction results of port loading and unloading operations. S5. Based on the deviation between efficiency prediction results and actual operation data, the model parameters or data preprocessing strategies are dynamically adjusted through automatic data feedback and manual intervention verification to form a closed-loop optimization mechanism.

[0006] Preferably, in step S1, the IoT terminal includes: a quay crane sensor, a gantry crane sensor, a container RFID reader, a yard positioning module, a weather monitoring instrument, and a truck dispatching terminal; The equipment operating status data includes load value, rotation speed, operation frequency and fault warning signal, and the collection frequency is 1 time / minute; the container attribute data includes container type, weight, dangerous goods class and destination information; the operation scheduling data includes berth occupancy status, truck matching plan and operation team schedule.

[0007] Preferably, in step S2, the preprocessing of the collected data specifically includes: (1) Data cleaning Box plot method is used to identify and remove outliers, missing values ​​are filled in based on improved KNN algorithm, and duplicate records with identical fields are identified and deleted based on the uniqueness constraints of timestamp and device ID to remove duplicate values. (2) Normalization All numerical data were normalized using the Z-Score standardization formula and mapped to the interval [-1, 1]. (3) Feature alignment Based on the job timestamp, data from different collection frequencies are uniformly synchronized to a minute-level time granularity; (4) Feature fusion The sliding window method is used to extract time-series features from dynamic data, and one-hot encoding is used to convert static features into numerical features. The time-series features and static features are then concatenated to form a joint feature matrix. (5) Attention mechanism weighted fusion processing, By calculating the Pearson correlation coefficient between each feature and loading / unloading efficiency, feature weights are dynamically allocated.

[0008] Preferably, in step S3, the constructed CNN-LSTM fusion model specifically includes: The CNN module, consisting of three convolutional layers and three pooling layers, is used to extract spatially correlated features from a multi-dimensional joint feature set. The LSTM module, including two hidden layers, is designed to capture the temporal evolution of loading and unloading efficiency over time. The attention mechanism layer, embedded between the CNN module and the LSTM module, weights the spatial features extracted by the CNN before inputting them into the LSTM module.

[0009] Preferably, in step S3, the training of the CNN-LSTM fusion model specifically includes: The model training uses the Adam optimizer, with mean absolute error as the loss function. It combines mean squared error to suppress extreme value interference and uses Dropout layer and L2 regularization to avoid overfitting. By using transfer learning techniques, a model that has already been trained on operational data from large hub ports is used as a pre-trained model. The parameters of the first three network layers are frozen, and only the parameters of the fully connected layers are fine-tuned.

[0010] Preferably, step S4 further includes a model dynamic adaptation module, which automatically adjusts the CNN convolution kernel size and the number of neurons in the LSTM hidden layer according to the port type, cargo type changes and weather level, so that the CNN-LSTM fusion model can adapt to different port scenarios.

[0011] Preferably, the data transmission channel uses 5G communication technology combined with edge computing nodes. The edge computing nodes can filter the collected raw data in real time, remove invalid and redundant data, and then upload the valid feature data to the cloud model.

[0012] Preferably, the method further includes step S5: real-time monitoring of the deviation between the efficiency prediction result and the actual operation data, and dynamic adjustment of the parameters of the fused deep learning model or the preprocessing strategy in step S2 based on the deviation through automatic data feedback and manual intervention verification, to form a closed-loop optimization mechanism.

[0013] Preferably, in step S4, the efficiency prediction results include the single-box operation time, the total batch operation time, and the efficiency bottleneck node.

[0014] Compared with existing technologies, the advantages of this invention are: 1. This invention employs a CNN-LSTM fusion model based on an attention mechanism, which achieves both the collaborative extraction of spatial and temporal features and enhances the contribution of core influencing factors, resulting in single-sample inference time ≤22ms. This comprehensively meets the "minute-level response" hard real-time requirement for port scheduling. Whether it's a sudden change in cargo flow, a shift in weather conditions, or adjustments to operational plans, it can quickly output the latest prediction results, helping dispatchers adjust their plans in a timely manner.

[0015] 2. This invention is equipped with a model dynamic adaptation module, which can adapt to different scenarios such as container ports and bulk cargo ports by automatically adjusting the size of the CNN convolution kernel and the number of neurons in the LSTM hidden layer. For ports of different sizes and with different equipment configurations, no additional customized development is required. It can be deployed directly by connecting to existing IoT terminals and scheduling systems, which greatly reduces the threshold and cost of technology implementation.

[0016] 3. This invention proposes an improved KNN missing value imputation algorithm that integrates time and feature correlation. Combined with techniques such as box plot outlier handling and timestamp feature alignment, a dedicated preprocessing process for multi-source heterogeneous port data is constructed. The data preprocessing accuracy is improved by more than 20% compared with traditional methods, providing a reliable guarantee for the quality of model input. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the process of the present invention.

[0018] Figure 2 This is a cross-sectional view of the structure of multi-source data acquisition in this invention.

[0019] Figure 3 This is a schematic diagram of the CNN-LSTM fusion model in this invention. Detailed Implementation

[0020] To facilitate understanding of this application and to make the aforementioned objectives, features, and advantages of this application more apparent, a detailed description of specific embodiments of this application is provided below in conjunction with the accompanying drawings. Numerous specific details are set forth in the following description to provide a thorough understanding of this application, and preferred embodiments are shown in the accompanying drawings. However, this application can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of this application. This application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of this application; therefore, this application is not limited to the specific embodiments disclosed below. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified. In the description of this application, "several" means at least one, such as one, two, etc., unless otherwise explicitly specified. It should be noted that when an element is referred to as being "fixed to" another element, it can be directly attached to the other element or there may be an intervening element. When an element is referred to as being "connected to" another element, it can be directly connected to the other element or there may be an intervening element. The terms "vertical," "horizontal," "left," "right," and similar expressions used herein are for illustrative purposes only and do not represent the only possible implementations. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is only for describing particular implementations and is not intended to limit the scope of this application. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0021] Reference Figure 1 As shown, a method for predicting port loading and unloading efficiency based on the fusion of deep learning and IoT data includes the following steps: S1. By deploying IoT terminals and data transmission channels, synchronously collect equipment operating status data, container attribute data, yard layout data, meteorological environment data, and operation scheduling data; In the multi-source data acquisition phase, high-precision sensors are installed on core loading and unloading equipment such as quay cranes and gantry cranes to collect real-time data on equipment operating status, including load values, rotation speed, operation frequency, and fault warning signals. The sampling frequency is set to once per minute to balance data real-time performance and storage costs. RFID readers are deployed at key nodes in container flow to accurately acquire container attribute data such as container type, weight, dangerous goods class, and destination. GNSS and UWB-based positioning modules are deployed in the yard area to collect data on yard layout and container stacking locations. Multi-dimensional meteorological monitoring instruments are set up in the port operation area to capture meteorological environmental data such as wind speed, rainfall, visibility, and temperature. At the same time, the port dispatch system is connected via an interface to synchronously acquire operation scheduling data such as berth occupancy, truck matching plans, and work shift schedules.

[0022] like Figure 2 As shown, the collected multi-source heterogeneous data is categorized into "sensor data → device bus data → external file data → real-time streaming data," and data reception and parsing are completed through corresponding interfaces (I2C / SPI, Ethernet, USB / SD card, WiFi / Bluetooth) and protocols (TCP / IP, MQTT), covering more than 90% of the core influencing factors of loading and unloading efficiency and solving the problem of one-dimensional data. To ensure the real-time performance and security of data transmission, the transmission link adopts a "5G base station + edge computing node" architecture. The edge node has a built-in real-time filtering unit, and the edge computing node performs pre-processing (filtering, deduplication, and preliminary anomaly screening), and completes the qualification verification through the "data quality judgment module." Qualified data is stored in a local dual-buffer unit (SD card + memory), and unqualified data is discarded and anomaly logs are recorded. This can remove invalid data such as offline noise from sensors (data validity rate is improved to 98%). The data is uploaded through SSL / TLS encryption protocol, and the transmission latency is controlled within 10 milliseconds, which fully meets the needs of real-time port scheduling. At the same time, the local caching unit ensures data integrity in network outage scenarios, laying a high-quality data foundation for subsequent accurate prediction.

[0023] S2. Preprocess the collected data; Data preprocessing is a key prerequisite for improving the prediction accuracy of the model. This invention designs a hierarchical processing flow for multi-source heterogeneous data, which first involves data cleaning. Box plot analysis is used to identify and remove outliers in data. Applicable scenarios include sudden anomalies (such as jumps caused by sensor malfunctions) in equipment sensor data (e.g., crane load, quay crane operation frequency) and environmental data (e.g., wind speed). The algorithm steps and formulas are as follows: First, for single-dimensional data (e.g., the "crane load" sequence)... Calculate the quartiles, lower quartile This refers to the value at the 25th percentile after the data is sorted in ascending order.

[0024] Upper Quartile The value at the 75th position, that is:

[0025] in, This represents the floor function. The number of data samples is given; then the interquartile range (IQR) is calculated using the following formula:

[0026] This is used to measure the dispersion of the data; finally, an outlier threshold is defined, with a lower limit of [value missing]. The upper limit is That is, the outlier determination criteria are:

[0027] Data that meets this condition is removed. For abnormal segments of more than 10 minutes caused by continuous sensor failure, they are marked as "missing values" and proceed to the next step of processing. For missing values, an improved KNN algorithm (K=5-8) is used for imputation. This is applicable to scenarios where missing values ​​are caused by sensor offlineness or data transmission delays (single missing value or consecutive short-term missing values, not exceeding 30 minutes). Algorithm improvements: Traditional KNN relies solely on historical data of the same dimension; this method integrates "temporal correlation" and "feature correlation." The core calculation formula is as follows: First, the time of the target missing value is... Select the first hour ( minutes to A set of non-missing data points (minutes) As a candidate set, among Corresponding time ( ) and eigenvalues ; Calculate candidate data and The "time distance" of a moment After normalizing to the [0,1] interval using min-max, the formula is:

[0028] in, , These represent the maximum and minimum temporal distances of the candidate sets; calculate the "feature distance". ,in, For the same equipment The mean of the three normal data points before time step 1, also normalized, yields:

[0029] The formula for calculating the overall distance weight is as follows: (Weights were determined through 5-fold cross-validation optimization); the 5 nearest neighbors with the smallest weights (K=5) were selected, and the missing values ​​were filled using a weighted average. The formula for filling missing values ​​is:

[0030] Where i, j∈K selected nearest neighbors, this formula makes the data of the closer neighbors contribute higher weight, improving the filling accuracy. Based on the uniqueness constraints of timestamp and device ID, it identifies and deletes all duplicate records with exactly the same fields to remove duplicate values.

[0031] After data cleaning, normalization is performed using the Z-Score normalization formula to normalize all numerical data. This is applicable to scenarios where different units of measurement (e.g., "cargo weight" in tons, "wind speed" in m / s, "operation frequency" in times / hour) are standardized, eliminating the impact of numerical differences on model training. The core calculation formula is the Z-Score normalization formula:

[0032] in, These are the original eigenvalues. The mean of this feature in the training set ( , =1 to N, where N is the number of samples in the training set). The standard deviation of this feature in the training set (

[0033] This formula maps all numerical data to the interval [-1, 1]. Constraints: To avoid the influence of extreme values, a truncation rule is set for the normalized result, i.e., when... If the value is greater than 3, it will be forcibly set to 3. When the value is less than -3, it is forcibly set to -3 to reduce the interference of abnormal noise on subsequent model training.

[0034] To achieve time synchronization of multi-source data, this invention uses the operation timestamp as a benchmark for feature alignment, unifies the synchronization of data from different collection frequencies to a minute-level time granularity, and finally constructs a joint feature set with dimensions of [time step, feature dimension] by extracting time-series features such as the equipment operation trend over the past hour and static features such as cargo type.

[0035] For dynamic data (such as equipment operating status and environmental data), a sliding window method is used, with a window size of 60 (corresponding to 1 hour and a step size of 1 minute). Statistical features within the window are extracted: time-domain features include the mean ( ),variance( ), maximum value ( ), minimum value ( ), slope ( (Reflecting the trend of change within 1 hour); frequency domain features are extracted using Fast Fourier Transform (FFT), with the formula as follows:

[0036] The top three frequency components with the largest amplitudes are selected as periodic features (such as the load fluctuation period of a crane).

[0037] Static feature processing: One-hot encoding is performed on cargo attributes (e.g., type: container / bulk cargo / liquid) and berth attributes (e.g., berth class: 100,000-ton / 200,000-ton). For example, "container" is encoded as [1,0,0], and "100,000-ton berth" is encoded as [1,0]. Combined with features such as "truck matching rate 85%" in the scheduling data, all static features are finally concatenated into an 8-dimensional vector. Fusion method: The temporal feature matrix (60×12) and the static feature vector (1×8) are concatenated column-wise to form a joint feature matrix of dimension 60×20, which is used as the model input.

[0038] The attention mechanism layer calculates the Pearson correlation coefficient between each feature and loading / unloading efficiency, dynamically allocates feature weights, and controls the weight ratio within the range of 0.05-0.4. This effectively enhances the feature contribution of key influencing factors such as crane load and strong winds, and selects the core feature vectors.

[0039] S3. Construct and train a CNN-LSTM fusion model based on the attention mechanism; like Figure 3 As shown, the CNN-LSTM fusion model fully leverages the advantages of both networks. The CNN module has three convolutional layers and two pooling layers, with convolutional kernels of 3×3 size and numbers of 64, 128, and 256 respectively. Max pooling is used for dimensionality reduction to efficiently extract spatial correlation features from multi-source features, such as the spatial matching relationship between equipment operating status and yard layout. The LSTM module has two hidden layers, each with 256 neurons. Dropout technology (dropout rate of 0.2) is used to prevent overfitting and accurately capture the temporal evolution of loading and unloading efficiency over time, such as the fluctuation characteristics of the operation rhythm at different times. An attention mechanism layer is embedded between the CNN and LSTM modules, weighting the spatial features extracted by the CNN before inputting them into the LSTM module, making the model more focused on key information.

[0040] To address the issues of insufficient training data and slow convergence speed, this invention introduces transfer learning technology. It uses a model that has already been trained on operational data from large hub ports as a pre-trained model, freezes the parameters of the first three network layers, and only fine-tunes the parameters of the fully connected layers, thereby reducing the number of model training iterations by 40% and improving the prediction accuracy in small sample scenarios.

[0041] During model training, the Adam optimizer was used with an initial learning rate of 0.001, which was decayed to 0.8 times the previous learning rate every 20 iterations. The mean absolute error (MAE) was used as the core loss function, combined with mean squared error (MSE) to suppress extreme value interference. Overfitting was avoided by using Dropout layer and L2 regularization. Training was stopped when the loss value of the validation set fluctuated less than 0.005 for 10 consecutive rounds to ensure model convergence and good generalization performance.

[0042] S4. Input the preprocessed data into the trained CNN-LSTM fusion model and output the efficiency prediction results of port loading and unloading operations. In the prediction output and dynamic adaptation stage, the trained model can receive preprocessed real-time fusion features and output multi-dimensional prediction results such as single container operation time, total batch operation time, and efficiency bottleneck nodes. The efficiency bottleneck nodes are derived by inversely inferring the feature contribution of the model output. When the weight of a certain feature is ≥0.3 and the time consumption of the corresponding operation deviates from the average by more than 20%, it is marked as a bottleneck node, providing a clear direction for the port to optimize its operation process in a targeted manner.

[0043] S5. Based on the discrepancy between efficiency prediction results and actual operational data, the model parameters or data preprocessing strategies are dynamically adjusted through automatic data feedback and manual intervention verification to form a closed-loop optimization mechanism. Considering the differences in various port scenarios, this invention specifically designs a dynamic model adaptation module. This module can automatically adjust the CNN convolutional kernel size (3×3 / 5×5) and the number of LSTM hidden layer neurons (128 / 256) based on port type (container port / bulk cargo port), cargo type changes, and weather level. The adaptation adjustment time is ≤10 minutes, and the model prediction accuracy remains at R after adjustment. 2 With a resolution of ≥0.90, the model can be quickly adapted to different scenarios without reconstruction, greatly improving the practicality and robustness of the technology.

[0044] Example 1. Experimental data: Six months of actual operation data from a coastal hub port were collected, including operation records of 10 cranes and 5 berths, totaling more than 100,000 samples as the dataset, which were divided into training set and test set in a 7:3 ratio.

[0045] 2. Comparison Objects: Traditional machine learning models: Linear Regression (LR), Support Vector Machine (SVM), Random Forest (RF); Single deep learning model: pure convolutional neural network (pure CNN), pure long short-term memory network (pure LSTM); A CNN-LSTM fusion model without attention mechanism; The model of this invention is a CNN-LSTM fusion model based on the attention mechanism (CNN-Attention-LSTM).

[0046] 3. Evaluation indicators: Mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) are used. 2 As a core evaluation indicator, the accuracy of prediction is quantified.

[0047] 4. Experimental steps: 4.1 Based on the same dataset, the comparison model and the model of this invention were trained separately.

[0048] 4.2 Validate the prediction results of each model on the test set and obtain statistical index values.

[0049] 4.3 Simulate different scenarios (such as severe weather, sudden changes in cargo type) to verify the scenario adaptability of the model of this invention.

[0050] 5. Expected Results: The MAE and RMSE of the model in this invention are reduced by more than 30% compared with the traditional model, and R... 2 ≥0.92, real-time prediction latency ≤5 seconds, scene adaptation adjustment time ≤10 minutes.

[0051] 6. Experimental Results and Analysis 6.1 Experimental Results To evaluate the predictive performance of the models under normal working conditions, the metrics of each model were compared on the test set, and the results are shown in Table 1: Table 1 Performance comparison of each model on the test set

[0052] 6.2 Results Analysis The model of this invention significantly outperforms all comparable models in terms of MAE, RMSE, and R². Specifically, the MAE is reduced by approximately 30.4% compared to the best comparable model (CNN-LSTM without attention), and the R² is significantly higher. 2 The score was improved to 0.93, meeting the expected target. Traditional machine learning models (LR, SVM, RF) perform poorly because they cannot effectively capture temporal and spatial features. Pure CNN and pure LSTM are good at spatial and temporal feature extraction, respectively, but their effectiveness is limited when used alone.

[0053] By introducing an attention mechanism, the model can focus more on key features (such as wind speed and peak load), further improving prediction accuracy.

[0054] 7. Scene Adaptability Verification To verify the generalization ability of the model of this invention, tests were conducted under simulated severe weather (wind speed ≥ 8 m / s) and sudden changes in cargo type (container → bulk cargo). The results are shown in the table below: Table 2 Model performance after scene adaptation (R 2 )

[0055] The model of this invention can adapt to sudden scene changes within 10 minutes through a dynamic parameter adjustment module, and the adjusted R... 2 It remains above 0.90, demonstrating good generalization ability and engineering practicality.

[0056] As is known from common technical knowledge, this invention can be implemented through other embodiments that do not depart from its spirit or essential characteristics. Therefore, the disclosed embodiments described above are merely illustrative in all respects and are not the only ones. All modifications within the scope of this invention or its equivalents are included in this invention.

Claims

1. A method for predicting port loading and unloading efficiency based on the fusion of deep learning and IoT data, characterized in that: Includes the following steps: S1. By deploying IoT terminals and data transmission channels, synchronously collect equipment operating status data, container attribute data, yard layout data, meteorological environment data, and operation scheduling data; S2. Preprocess the collected data; S3. Construct and train a CNN-LSTM fusion model based on the attention mechanism; S4. Input the preprocessed data into the trained CNN-LSTM fusion model and output the efficiency prediction results of port loading and unloading operations. S5. Based on the deviation between efficiency prediction results and actual operation data, the model parameters or data preprocessing strategies are dynamically adjusted through automatic data feedback and manual intervention verification to form a closed-loop optimization mechanism.

2. The port loading and unloading efficiency prediction method based on deep learning and IoT data fusion as described in claim 1, characterized in that: In step S1, the IoT terminal includes: quay crane sensor, gantry crane sensor, container RFID reader, yard positioning module, weather monitoring instrument and truck dispatching terminal; The equipment operating status data includes load value, rotation speed, operation frequency and fault warning signal, and the collection frequency is 1 time / minute; the container attribute data includes container type, weight, dangerous goods class and destination information; the operation scheduling data includes berth occupancy status, truck matching plan and operation team schedule.

3. The port loading and unloading efficiency prediction method based on deep learning and IoT data fusion as described in claim 1, characterized in that: In step S2, the preprocessing of the collected data specifically includes: (1) Data cleaning Box plot method is used to identify and remove outliers, missing values ​​are filled in based on improved KNN algorithm, and duplicate records with identical fields are identified and deleted based on the uniqueness constraints of timestamp and device ID to remove duplicate values. (2) Normalization All numerical data were normalized using the Z-Score standardization formula and mapped to the interval [-1, 1]. (3) Feature alignment Based on the job timestamp, data from different collection frequencies are uniformly synchronized to a minute-level time granularity; (4) Feature fusion The sliding window method is used to extract time-series features from dynamic data, and one-hot encoding is used to convert static features into numerical features. The time-series features and static features are then concatenated to form a joint feature matrix. (5) Attention mechanism weighted fusion processing By calculating the Pearson correlation coefficient between each feature and loading / unloading efficiency, feature weights are dynamically allocated.

4. The port loading and unloading efficiency prediction method based on deep learning and IoT data fusion as described in claim 1, characterized in that: In step S3, the constructed CNN-LSTM fusion model specifically includes: The CNN module, consisting of three convolutional layers and three pooling layers, is used to extract spatially correlated features from a multi-dimensional joint feature set. The LSTM module, including two hidden layers, is designed to capture the temporal evolution of loading and unloading efficiency over time. The attention mechanism layer, embedded between the CNN module and the LSTM module, weights the spatial features extracted by the CNN before inputting them into the LSTM module.

5. The port loading and unloading efficiency prediction method based on deep learning and IoT data fusion as described in claim 1, characterized in that: In step S3, the training of the CNN-LSTM fusion model specifically includes: The model training uses the Adam optimizer, with mean absolute error as the loss function. It combines mean squared error to suppress extreme value interference and uses Dropout layer and L2 regularization to avoid overfitting. By using transfer learning techniques, a model that has already been trained on operational data from large hub ports is used as a pre-trained model. The parameters of the first three network layers are frozen, and only the parameters of the fully connected layers are fine-tuned.

6. The port loading and unloading efficiency prediction method based on deep learning and IoT data fusion as described in claim 1, characterized in that: Step S4 also includes a model dynamic adaptation module, which automatically adjusts the CNN convolution kernel size and the number of neurons in the LSTM hidden layer according to the port type, cargo type changes and weather level, so that the CNN-LSTM fusion model can be adapted to different port scenarios.

7. The port loading and unloading efficiency prediction method based on deep learning and IoT data fusion as described in claim 1, characterized in that: The data transmission channel uses 5G communication technology combined with edge computing nodes. The edge computing nodes can filter the collected raw data in real time, remove invalid and redundant data, and then upload the valid feature data to the cloud model.

8. The port loading and unloading efficiency prediction method based on deep learning and IoT data fusion as described in claim 1, characterized in that: It also includes step S5: real-time monitoring of the deviation between the efficiency prediction results and the actual operation data, and dynamic adjustment of the parameters of the fused deep learning model or the preprocessing strategy in step S2 based on the deviation through automatic data feedback and manual intervention verification, forming a closed-loop optimization mechanism.

9. The port loading and unloading efficiency prediction method based on deep learning and IoT data fusion as described in claim 1, characterized in that: In step S4, the efficiency prediction results include the single-box operation time, the total batch operation time, and the efficiency bottleneck node.

Images