A molten iron intermodal transportation state intelligent detection method based on multi-source data fusion

Abstract

The present application relates to the technical field of multi-source data fusion, and particularly relates to a molten metal combined transport state intelligent detection method based on multi-source data fusion, which obtains at least two monitoring data sequences in the transport process of molten metal combined transport, extracts features of each monitoring data sequence through data level fusion and feature level fusion to obtain at least two feature sequences in the process of multi-source data fusion of all monitoring data sequences, classifies all feature sequences, acquires the best attention mechanism according to the category to which each feature sequence belongs, and performs decision level fusion on all feature sequences according to the best attention mechanism of each feature sequence to obtain the transport state detection result of molten metal combined transport. Through the analysis of the adaptive attention mechanism, the robustness of the transport state detection based on multi-source data fusion technology in the molten metal combined transport scene is improved.

Description

Technical Field

[0001] This invention relates to the field of multi-source data fusion technology, and in particular to an intelligent detection method for the transportation status of rail-water intermodal transport based on multi-source data fusion. Background Technology

[0002] Intermodal transport is a multimodal transport model that organically combines rail transport with waterway transport (such as inland waterway shipping, coastal shipping, or ocean shipping). By optimizing transport organization and connecting different modes of transport, it achieves efficient logistics services with "one bill of lading to the end and seamless connection" for goods. Its core lies in integrating the advantages of rail and water transport, reducing logistics costs, improving transport efficiency, and reducing environmental impact.

[0003] Traditional rail-water intermodal transport monitoring relies on single data sources, such as railway dispatching systems or port loading and unloading records, resulting in fragmented information. Multi-source data fusion technology, however, integrates multi-dimensional information such as railway train timetables, port container stacking status, ship AIS positioning, meteorological data, and customs clearance progress. This allows for the construction of a full-chain data network covering "port-railway-waterway-terminal." For example, Lianyungang Port, by integrating data from cargo owners, carriers, and customs, has achieved an information interoperability rate exceeding 90%, with annual data exchange covering over 3 million TEUs of user business and over 120 million tons of freight, completely eliminating information blind spots inherent in traditional methods. Furthermore, multi-source data fusion technology addresses the issues of poor timeliness and low prediction accuracy inherent in traditional monitoring methods, truly driving the transformation of rail-water intermodal transport towards intelligence, greening, and economic efficiency, providing core support for building a modern integrated transportation system.

[0004] Multi-source data fusion technology is structured in three levels: data-level fusion, feature-level fusion, and decision-level fusion. Decision-level fusion, based on data-level and feature-level fusion, combines data sources at a higher level to make decisions and output results. It directly targets specific decision objectives, generating more accurate and comprehensive final conclusions by comprehensively analyzing the independent decision results from various data sources, providing direct evidence for command and control, situation assessment, and system performance optimization. Attention mechanism processing is the most crucial operational step in decision-level fusion. The attention mechanism weights the decisions from different data sources, assigning different attention scores based on their importance or relevance. These attention scores are then used to weight and combine the decision results from different data sources to form the final decision.

[0005] However, the existing attention mechanism has problems such as computational redundancy, inefficient resource allocation and insufficient model robustness compared to the rail-water intermodal transport scenario. For example, in the current scenario, when processing multi-source heterogeneous data such as port status, train timetables and weather data at the same time, the attention mechanism (fully connected mechanism) is prone to serious problems such as computational resource overload. Summary of the Invention

[0006] In view of this, embodiments of the present invention provide an intelligent detection method for rail-water intermodal transport status based on multi-source data fusion, in order to solve the problem of how to optimize the attention mechanism processing in the decision-level fusion in the rail-water intermodal transport scenario, so as to improve the accuracy of rail-water intermodal transport status detection based on multi-source data fusion.

[0007] This invention provides an intelligent detection method for rail-water intermodal transport status based on multi-source data fusion, the method comprising the following steps:

[0008] During the rail-water intermodal transport process, monitoring data from at least two dimensions are collected, resulting in at least two monitoring data sequences. Each dimension corresponds to one monitoring data sequence. In the process of multi-source data fusion of all monitoring data sequences, features are extracted from each monitoring data sequence through data-level fusion and feature-level fusion, resulting in at least two feature sequences.

[0009] Based on the length, variance, and information entropy of each feature sequence, all feature sequences are divided into a first category and a second category. Based on the sparsity and periodicity of each feature sequence in the second category, the feature sequences in the second category are further divided into a third category and a fourth category.

[0010] Based on the category to which each feature sequence belongs, an optimal attention mechanism is adaptively obtained. The categories include a first category, a third category, and a fourth category. Based on the optimal attention mechanism for each feature sequence, an attention score for each feature sequence is obtained. Based on the attention scores of each feature sequence, decision-level fusion is performed on all feature sequences to obtain the transportation status detection result of rail-water intermodal transport.

[0011] Preferably, the step of dividing all feature sequences into a first category and a second category based on the length, variance, and information entropy of each feature sequence includes:

[0012] For any feature sequence, a first feature evaluation value is obtained based on the length, variance, and information entropy of the feature sequence. If the first feature evaluation value of the feature sequence is less than a preset first feature evaluation value threshold, the feature sequence is classified into a first category. If the first feature evaluation value of the feature sequence is greater than or equal to the preset first feature evaluation value threshold, the feature sequence is classified into a second category.

[0013] Preferably, obtaining the first feature evaluation value of any feature sequence based on its length, variance, and information entropy includes:

[0014] The length feature value is obtained by taking the negative of the length of any feature sequence as the independent variable of an exponential function with the natural constant as the base; the variance of any feature sequence is normalized to obtain a first feature value; the information entropy of any feature sequence is normalized to obtain a second feature value; the feature fluctuation value is obtained based on the mean of the first feature value and the second feature value; and the mean of the length feature value and the feature fluctuation value is used as the first feature evaluation value of any feature sequence.

[0015] Preferably, the step of dividing the feature sequences in the second category into a third category and a fourth category based on the sparsity and periodicity of each feature sequence in the second category includes:

[0016] For any feature sequence in the second category, a second feature evaluation value is obtained based on the number of zero elements in the feature sequence and the periodicity of the feature sequence. If the second feature evaluation value of the feature sequence is greater than or equal to a preset second feature evaluation value threshold, the feature sequence is classified into a third category; if the second feature evaluation value of the feature sequence is less than the preset second feature evaluation value threshold, the feature sequence is classified into a fourth category.

[0017] Preferably, obtaining the second feature evaluation value of any feature sequence based on the number of zero elements in any feature sequence and the periodicity of any feature sequence includes:

[0018] The number of elements with a value of 0 in any given feature sequence is counted to obtain the corresponding percentage. This percentage is then normalized to obtain a normalized percentage. The autocorrelation coefficient of any given feature sequence is obtained. The square root of the total number of elements in any given feature sequence is then obtained. The ratio between the upper quantile of the standard normal distribution and the square root result is then calculated. The difference between the autocorrelation coefficient and the ratio is then normalized to obtain a normalized difference. Finally, the normalized percentage and the normalized difference are weighted and summed to obtain the second feature evaluation value of any given feature sequence.

[0019] Preferably, the adaptive attention mechanism based on the category to which each feature sequence belongs includes:

[0020] If any feature sequence belongs to the first or third category, the best attention mechanism is set to sparse attention mechanism; if any feature sequence belongs to the fourth category, the best attention mechanism is set to fully connected attention mechanism.

[0021] The beneficial effects of the embodiments of the present invention compared with the prior art are as follows:

[0022] This invention collects monitoring data from at least two dimensions during the rail-water intermodal transport process, resulting in at least two monitoring data sequences, with one monitoring data sequence corresponding to one dimension. During multi-source data fusion of all monitoring data sequences, feature extraction is performed on each monitoring data sequence through data-level fusion and feature-level fusion, yielding at least two feature sequences. Based on the length, variance, and information entropy of each feature sequence, all feature sequences are divided into a first category and a second category. Based on the sparsity and periodicity of each feature sequence in the second category, the feature sequences in the second category are further divided into a third category and a fourth category. An optimal attention mechanism is adaptively acquired based on the category to which each feature sequence belongs. The categories include the first, third, and fourth categories. An attention score is obtained for each feature sequence based on the optimal attention mechanism. Decision-level fusion is then performed on all feature sequences based on their attention scores to obtain the rail-water intermodal transport status detection result. Among them, by adopting an adaptive processing method of "sparse attention mechanism + fully connected attention mechanism", the optimal attention mechanism is obtained adaptively for each feature sequence in the decision-level fusion process. This can enhance the efficiency of obtaining attention scores by uniformly using the fully connected attention mechanism, reduce the impact of computing resource overload, and improve the robustness of transportation status detection based on multi-source data fusion technology in the rail-water intermodal transport scenario. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart of a method for intelligent detection of rail-water intermodal transport status based on multi-source data fusion, provided in Embodiment 1 of the present invention. Detailed Implementation

[0025] Embodiments of this disclosure are described in detail below, with examples of these embodiments illustrated in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this disclosure, and should not be construed as limiting it.

[0026] It should be noted that the terms "first," "second," etc., used in this disclosure and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure.

[0027] To illustrate the technical solution of the present invention, specific embodiments are described below.

[0028] See Figure 1 This is a flowchart of a method for intelligent detection of rail-water intermodal transport status based on multi-source data fusion, as provided in Embodiment 1 of the present invention. Figure 1 As shown, the method may include:

[0029] Step S101: During the rail-water intermodal transport process, at least two dimensions of monitoring data are collected, resulting in at least two monitoring data sequences. Each dimension corresponds to one monitoring data sequence. During the multi-source data fusion of all monitoring data sequences, features are extracted from each monitoring data sequence through data-level fusion and feature-level fusion to obtain at least two feature sequences.

[0030] The monitoring data involved in rail-water intermodal transport is diverse, including railway train timetables, port container stacking status, ship AIS positioning, meteorological data, customs clearance progress, and other multi-dimensional information. To avoid relying on a single data source for rail-water intermodal transport status monitoring, which leads to information fragmentation, existing technologies acquire multi-dimensional monitoring data during rail-water intermodal transport and integrate this data using multi-source data fusion technology to improve the accuracy of transport status monitoring. However, in decision-level fusion within multi-source data fusion technology, existing attention mechanisms suffer from computational redundancy, inefficient resource allocation, and insufficient model robustness compared to rail-water intermodal transport scenarios. Therefore, in this embodiment of the invention, multi-dimensional monitoring data is analyzed to employ the optimal attention mechanism for each dimension, thereby improving the effectiveness of decision-level fusion.

[0031] Multi-source data fusion technology is known to be divided into a three-level architecture: data-level fusion, feature-level fusion, and decision-level fusion. Decision-level fusion, based on the first two levels of fusion processing, combines data sources at a higher level to make decisions and output results. Therefore, in this embodiment of the invention, firstly, during the rail-water intermodal transport process, monitoring data from at least two dimensions are collected, resulting in at least two monitoring data sequences. Each dimension corresponds to one monitoring data sequence. The specific dimensions are not limited and can be set according to requirements. Preferably, this embodiment selects monitoring data from four dimensions: railway train timetables, port container stacking status, ship AIS positioning, and meteorological data. Then, data-level fusion is performed on all monitoring data sequences to solve the problem of spatiotemporal alignment of heterogeneous data. Data-level fusion aligns monitoring data from different sensors and different stages in both time and space, eliminating data silos. The specific operation is as follows:

[0032] (1) Time synchronization

[0033] Interpolation: For data with inconsistent sampling frequencies (e.g., ship AIS updates every 30 seconds, port cameras capture one frame every 1 second), linear interpolation or spline interpolation is used to generate data with a unified timestamp.

[0034] Time window alignment: Divide multi-source data into fixed time windows (e.g., 1 minute) to ensure that data within the same window participates in the fusion.

[0035] (2) Spatial alignment

[0036] Coordinate transformation: Converting data from different coordinate systems (such as WGS84 coordinates from a ship's GPS or local coordinates from a port yard) into a unified coordinate system. For example, a seven-parameter transformation model can be used to convert WGS84 coordinates into local port coordinates.

[0037] Spatial interpolation: For data with inconsistent spatial resolution (such as wind speed data from a weather station covering a range of 10 kilometers, while GPS data from ships are discrete points), inverse distance weighted interpolation (IDW) is used to generate a continuous spatial wind speed distribution map.

[0038] (3) Data association

[0039] Entity matching: Linking data from different stages using unique identifiers (such as container IDs or ship MMSI codes). For example, linking RFID data of containers transported by rail with camera image data from port yards using container IDs.

[0040] Trajectory matching: For moving targets (such as ships and trains), trajectory similarity algorithms (such as Dynamic Time Warping DTW) are used to match trajectory data from different sensors to eliminate positioning errors.

[0041] Following data-level fusion, feature-level fusion is performed on all feature sequences to extract multi-dimensional state representations. This process extracts key features reflecting the transportation state from the aligned data obtained after the data-level fusion, providing input for decision-level fusion. The specific operations are as follows:

[0042] (1) Temporal feature extraction

[0043] For example, the LSTM model: input the time series of ship position, speed, and acceleration, and capture time series features such as navigation stability (e.g., lateral acceleration variance) and arrival time prediction.

[0044] (2) Spatial feature extraction

[0045] CNN model: Input camera images of the port yard, and extract spatial features of whether the containers are stacked in a standardized manner (such as whether they are tilted or exceed the yard boundary) through convolutional layers.

[0046] Graph Neural Networks: Constructing railway tracks, waterways, and port nodes into a graph structure, inputting equipment status (such as track vibration) and environmental data (such as water level), and extracting network-level transportation status features.

[0047] (3) Statistical feature extraction

[0048] Random forest features: The random forest model is used to select features that have a significant impact on transportation status (such as ship load, wind speed, and yard occupancy rate).

[0049] Principal Component Analysis (PCA): Reduces the dimensionality of high-dimensional features (such as 10-dimensional sensor data) and retains the first three principal components (cumulative variance contribution rate > 85%).

[0050] At this point, the feature sequences for each dimension have been extracted. Each dimension corresponds to a monitoring data sequence, which is the input data for decision-level fusion. It should be noted that data-level fusion and feature-level fusion are existing technologies and will not be elaborated here.

[0051] Step S102: Based on the length, variance and information entropy of each feature sequence, divide all feature sequences into a first category and a second category. Based on the sparsity and periodicity of each feature sequence in the second category, divide the feature sequences in the second category into a third category and a fourth category.

[0052] After obtaining several feature sequences, these sequences are used as input data for decision-level fusion to arrive at specific decisions, such as path planning, resource allocation, or state detection. In decision-level fusion, attention mechanisms significantly improve the accuracy, robustness, and interpretability of the final decision through dynamic weight allocation, cross-modal semantic alignment, and anti-interference capabilities. It has become a core technology for multi-source data decision fusion in complex scenarios. Traditionally, fully connected attention mechanisms are used directly. However, in rail-water intermodal transport scenarios, dozens of features such as port status, train timetables, weather data, and equipment status are processed simultaneously. In this case, fully connected attention mechanisms generate a massive attention matrix, leading to a surge in computational resource consumption (the computational principle of fully connected attention mechanisms is to calculate the similarity between all input elements, causing computational complexity to increase quadratically with the input size). Furthermore, high computational latency can cause decision lag, affecting transportation efficiency. In addition, traditional attention mechanisms also suffer from a lack of targeted resource allocation and robustness deficiencies.

[0053] Based on the above analysis, the embodiments of the present invention use an adaptive processing method of "sparse attention mechanism + fully connected attention mechanism" to effectively solve the problems existing in the above methods. However, in the scenario of rail-water intermodal transport, decision-level fusion needs to process multi-source heterogeneous data (such as port equipment status, train timetable, ship dynamics, weather, etc.) and needs to take into account both real-time performance and accuracy. Therefore, the adaptive selection of sparse attention mechanism and fully connected attention mechanism needs to be dynamically adjusted based on task requirements, data characteristics and computing resources. Thus, the embodiments of the present invention make an adaptive selection based on sequence characteristics.

[0054] First, based on the sequence feature quantization classification criteria for each feature sequence, and then by dividing each feature sequence into multiple categories, the optimal attention mechanism for each feature sequence is determined based on its category. The method for dividing each feature sequence into multiple categories is as follows:

[0055] Sequence length is a key factor in determining the choice of attention mechanism: using a fully connected attention mechanism on long sequences leads to computational explosion, making a sparse attention mechanism more suitable; while using a fully connected attention mechanism on short sequences can more effectively capture global dependencies. Besides sequence length, this embodiment of the invention believes that statistical indicators such as variance and sparsity can also provide powerful quantitative classification standards for the adaptive selection of the optimal attention mechanism. Therefore, for any feature sequence, based on the length, variance, and information entropy of that feature sequence, a first feature evaluation value is obtained. The specific method for obtaining this value is as follows:

[0056] The length feature value is obtained by taking the negative of the length of any feature sequence as the independent variable of an exponential function with the natural constant as the base. The calculation formula is as follows: , Represents the length characteristic value. This represents an exponential function with the natural constant as its base. This represents the length of any feature sequence. The longer the sequence, the more suitable it is for using sparse attention mechanisms.

[0057] The variance of any feature sequence is normalized to obtain a first feature value. The information entropy of any feature sequence is normalized to obtain a second feature value. The feature fluctuation value is obtained based on the mean of the first and second feature values, and its calculation formula is as follows: , Indicates the characteristic fluctuation value. The variance of any characteristic sequence is represented. The information entropy of any feature sequence, The normalization function is represented by . It is worth noting that high variance and high information entropy can reflect the non-stationary mutation characteristics and chaotic fluctuation characteristics of any feature sequence. Norm is a proportional normalization function. When the result is closer to 1, it indicates that the fluctuation characteristics are particularly obvious. At this time, a fully connected attention mechanism is needed to capture such fluctuations. When it is closer to 0, it indicates that the fluctuation characteristics are not obvious. At this time, a sparse attention mechanism can be used to focus on key points.

[0058] Furthermore, the average of the feature length feature value and the feature fluctuation value is used as the first feature evaluation value of any feature sequence, and its calculation formula is as follows: , This represents the first feature evaluation value of any feature sequence.

[0059] For a binary classification calculation result, its value satisfies the range [0, 1]. When it approaches 0, it indicates that any feature sequence is relatively long and has inconspicuous fluctuation characteristics. In this case, it is more inclined to use the sparse attention mechanism for processing. Conversely, it is more inclined to use the fully connected attention mechanism for processing. Therefore, in this embodiment of the invention, the first feature evaluation value threshold for binary classification is set to 0.5. If the first feature evaluation value of any feature sequence is less than the preset first feature evaluation value threshold, it is considered that any feature sequence is more inclined to be processed by the sparse attention mechanism, and thus the any feature sequence is classified into the first category. If the first feature evaluation value of any feature sequence is greater than or equal to the preset first feature evaluation value threshold, it is considered that any feature sequence is more inclined to be processed by the fully connected attention mechanism, and thus the any feature sequence is classified into the second category.

[0060] Similarly, a first feature evaluation value is obtained for each feature sequence, and all feature sequences are divided into a first category and a second category. At this point, the feature sequences in the second category exhibit more pronounced fluctuations. However, due to the periodicity and sparsity of the sequences, frequent fluctuations also occur. Theoretically, this category is more suited to sparse attention mechanisms. Based on the analysis of the first feature evaluation value, it is easier to categorize these sequences into the second category, which tends to use fully connected attention mechanisms. This leads to a significant bias in the selection of attention mechanisms. Therefore, this embodiment of the invention constructs a second feature evaluation value based on the sparsity and periodicity of each feature sequence in the second category to distinguish frequent fluctuations caused by the periodicity and sparsity of the sequences, thereby making the selection of attention mechanisms more rigorous.

[0061] For any feature sequence in the second category, count the number of elements with a value of 0 in the feature sequence to obtain the corresponding proportion. Normalize the proportion to obtain a normalized proportion. Obtain the autocorrelation coefficient of the feature sequence. Obtain the square root of the total number of elements in the feature sequence. Obtain the ratio between the upper quantile of the standard normal distribution and the square root result. Normalize the difference between the autocorrelation coefficient and the ratio to obtain a normalized difference. Perform a weighted summation of the normalized proportion and the normalized difference to obtain the second feature evaluation value of the feature sequence.

[0062] The formula for calculating the second feature evaluation value is as follows:

[0063]

[0064] in, This represents the second feature evaluation value of any feature sequence in the second category. This represents the number of elements that are 0 in any given characteristic sequence. Represents the total number of elements in any characteristic sequence. Indicates the first weight. Indicates the second weight. The autocorrelation coefficient represents the autocorrelation coefficient of any characteristic sequence. This represents the upper quantile of the standard normal distribution.

[0065] It should be noted that if most elements in the feature sequence are zero (such as outlier markers in sensor data), the sparse attention mechanism can directly skip zero-value calculations, making it significantly more efficient than the fully connected attention mechanism. The larger the value, the stronger the advantage of using the sparse attention mechanism on it; if there exists a positive integer T (candidate period length) such that the autocorrelation coefficient at lag K=T satisfies the inequality: Then it can be determined that the characteristic sequence has periodicity. It is the upper quantile of the standard normal distribution, for (i.e., 95% confidence level), 0.975 ≈ 1.96 (which can be expressed as the critical value of the 95% confidence interval under the standard normal distribution). In periodicity judgment, the 95% confidence level is a commonly used choice, therefore it is assumed to be 1.96. , The larger the value, the more obvious the periodicity of any feature sequence, and the stronger the advantage of using sparse attention mechanism on it.

[0066] Preferably, in the embodiments of the present invention, the following settings are provided: No restrictions are imposed here.

[0067] Similarly, for each feature sequence in the second category, the evaluation value of the second feature is obtained. The larger the value, the more likely the corresponding feature sequence is to use the sparse attention mechanism. Therefore, based on the second feature evaluation value of each feature sequence in the second category, the sequence is reclassified: the second feature evaluation value threshold is set to 0.6. If the second feature evaluation value of any feature sequence is greater than or equal to the preset second feature evaluation value threshold, the fully connected attention mechanism determined in the first classification is modified to a sparse attention mechanism, thereby classifying the feature sequence into the third category; if the second feature evaluation value of any feature sequence is less than the preset second feature evaluation value threshold, the fully connected attention mechanism is more likely to be used for processing, thereby classifying the feature sequence into the fourth category.

[0068] At this point, all feature sequences are divided into three categories: Category 1, Category 3, and Category 4.

[0069] Step S103: Based on the category to which each feature sequence belongs, adaptively obtain the best attention mechanism. The categories include the first category, the third category, and the fourth category. Based on the best attention mechanism for each feature sequence, obtain the attention score for each feature sequence. Perform decision-level fusion on all feature sequences based on the attention scores of each feature sequence to obtain the transportation status detection result of rail-water intermodal transport.

[0070] After classifying the feature sequences into categories through the above steps, the optimal attention mechanism for each feature sequence can be adaptively determined based on its category. Specifically: if any feature sequence belongs to the first or third category, the optimal attention mechanism is set to sparse attention mechanism; if any feature sequence belongs to the fourth category, the optimal attention mechanism is set to fully connected attention mechanism.

[0071] The calculation formula for the fully connected attention mechanism is as follows:

[0072]

[0073] in, This represents the vector to be queried. Represents the key vector. Represents a value vector. The dimension of the key vector. This represents the normalization function.

[0074] It should be noted that the calculation scope of this mechanism is all queries. and bond ( ) Calculate the global attention score (i.e., the interaction of each location with all other locations), which is characterized by capturing global dependencies but has a high computational cost.

[0075] The sparse attention mechanism restricts the computational scope to a specific region by introducing a sparse pattern matrix M, and its calculation formula is as follows:

[0076]

[0077] Here, M is a binary matrix with elements of 0 or 1, used to mask irrelevant positions. This indicates element-wise multiplication.

[0078] It should be noted that its characteristic is to reduce the amount of computation by limiting the scope of interaction (such as local windows, random sampling, and Top-k important positions).

[0079] Since both fully connected attention mechanisms and sparse attention mechanisms are existing technologies, they will not be elaborated upon here. Instead, in the decision-level fusion process, the attention mechanism for each feature sequence is set to the optimal attention mechanism, resulting in an attention score for each feature sequence. Then, based on the attention scores of each feature sequence, decision-level fusion is performed on all feature sequences to obtain the transportation status detection result for rail-water intermodal transport. The general process of decision-level fusion is as follows:

[0080] (1) Feature-level fusion produces multi-source feature vectors (i.e., feature sequences) (such as train density, port turnover rate); (2) Obtain the attention score of each feature vector according to the fully connected attention mechanism; (3) Perform weighted averaging or selective fusion on the decision results of each feature vector according to the attention score, and output the final decision (such as resource allocation scheme, route planning).

[0081] It is worth noting that, according to existing theories, the core function of the attention mechanism is to assess the importance or relevance of each data source based on the result of feature-level fusion, calculate the attention score, and then use the attention score to weight and combine the decision results of each data source to achieve global optimization and balance of multi-source information, thereby improving the accuracy and adaptability of the final decision. Therefore, the focus of this embodiment is on step (2) in the decision-level fusion process, which obtains the optimal attention mechanism for each feature sequence to calculate a more suitable attention score. The remaining processes are existing technologies and will not be described in detail here.

[0082] For example: Suppose that by performing multi-source data fusion analysis on the feature sequences of train density, port turnover rate, and weather interference, and using the best attention mechanism of each feature sequence to process the output: the delay probability is 0.78. At this time, 0.78 is greater than the preset threshold of 0.7, triggering an early warning. The decision-level fusion output of the rail-water intermodal transport status detection result is: the port A train is expected to be delayed, and it is recommended to adjust the route.

[0083] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention 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 the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A multi-source data fusion-based intelligent detection method for the transportation status of intermodal transportation of molten iron, characterized in that, The method comprises: In the transportation process of the combined transport of molten iron, monitoring data of at least two dimensions is collected, at least two monitoring data sequences are correspondingly obtained, one dimension corresponds to one monitoring data sequence, in the process of multi-source data fusion of all monitoring data sequences, feature extraction is performed on each monitoring data sequence through data level fusion and feature level fusion, and at least two feature sequences are obtained; According to the length, variance and information entropy of each feature sequence, all feature sequences are divided into a first category and a second category, and according to the sparsity and periodicity of each feature sequence in the second category, the feature sequences in the second category are divided into a third category and a fourth category; According to the category to which each feature sequence belongs, an optimal attention mechanism is adaptively obtained, the category includes the first category, the third category and the fourth category, according to the optimal attention mechanism of each feature sequence, an attention score of each feature sequence is obtained, and all feature sequences are fused at a decision level according to the attention score of each feature sequence, so as to obtain a transportation state detection result of the combined transport of molten iron. The method comprises: For any feature sequence, a first feature evaluation value of the any feature sequence is obtained according to the length, variance and information entropy of the any feature sequence, if the first feature evaluation value of the any feature sequence is less than a preset first feature evaluation value threshold, the any feature sequence is divided into the first category, and if the first feature evaluation value of the any feature sequence is greater than or equal to the preset first feature evaluation value threshold, the any feature sequence is divided into the second category. The method comprises: For any feature sequence in the second category, a second feature evaluation value of the any feature sequence is obtained according to the number of elements being 0 in the any feature sequence and the periodicity of the any feature sequence, if the second feature evaluation value of the any feature sequence is greater than or equal to a preset second feature evaluation value threshold, the any feature sequence is divided into the third category, and if the second feature evaluation value of the any feature sequence is less than the preset second feature evaluation value threshold, the any feature sequence is divided into the fourth category. 2.The intelligent detection method for the intermodal transportation status based on multi-source data fusion according to claim 1, characterized in that, The method comprises: The length of the any feature sequence is taken as an independent variable of an exponential function with a natural constant as a base number to obtain a length feature value, the variance of the any feature sequence is normalized to obtain a first feature value, the information entropy of the any feature sequence is normalized to obtain a second feature value, and the mean value of the first feature value and the second feature value is taken as a feature fluctuation value, and the mean value of the length feature value and the feature fluctuation value is taken as the first feature evaluation value of the any feature sequence. 3.The intelligent detection method of the intermodal transportation status based on multi-source data fusion according to claim 1, characterized in that, The second feature evaluation value of the any feature sequence is obtained according to the number of elements being 0 in the any feature sequence and the periodicity of the any feature sequence, and the method comprises the following steps: The number of elements being 0 in the any feature sequence is counted to obtain a corresponding number proportion, the number proportion is normalized to obtain a normalized number proportion, the autocorrelation coefficient of the any feature sequence is obtained, the square root result of the total number of elements of the any feature sequence is obtained to obtain a ratio between the upper side quantile of the standard normal distribution and the square root result, the difference between the autocorrelation coefficient and the ratio is normalized to obtain a normalized difference, and the normalized number proportion and the normalized difference are weighted and summed to obtain the second feature evaluation value of the any feature sequence. 4.The intelligent detection method for the intermodal transportation status based on multi-source data fusion according to claim 1, characterized in that, The best attention mechanism is adaptively obtained according to the category to which each feature sequence belongs, and the method comprises the following steps: If the any feature sequence belongs to the first category or the third category, the best attention mechanism is set as a sparse attention mechanism; if the any feature sequence belongs to the fourth category, the best attention mechanism is set as a full connection attention mechanism.

Images