A cloud computing-based sea cucumber futures transaction data processing method and system

Abstract

The application provides a kind of based on cloud computing's sea cucumber futures transaction data processing method and system, applied to data processing technical field, by obtaining the multi-source heterogeneous original data of non-traditional information channel, carries out multimodal feature extraction, identifies undefined abnormal co-occurrence mode, constructs multidimensional association structure, and carries out confidence quantization evaluation, finally outputs risk early warning signal, can effectively identify and early warning potential, not being captured by traditional system in non-traditional information channel risk signal, and improve the accuracy and forward-looking advantage of sea cucumber futures transaction decision.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a cloud computing-based method and system for processing sea cucumber futures trading data. Background Technology

[0002] In the sea cucumber futures trading sector, existing cloud-based systems primarily rely on pre-set structured or semi-structured data sources, such as exchange quotes, official meteorological data, and mainstream news reports, for analysis. While these methods are effective when the market is stable, they suffer from serious limitations when faced with potential drastic changes caused by unknown or sudden factors, such as emerging diseases or localized environmental anomalies.

[0003] The core issue is that early warning signals of emergencies often first appear on non-traditional information channels, such as local aquaculture forums and social media platforms like short video platforms, and are presented in unstructured and non-standardized forms, such as short videos described in dialects and colloquial requests for help.

[0004] Existing systems lack the ability to proactively discover and analyze these non-standard information sources. Designed for processing structured data, their built-in text analysis module struggles to understand forum content mixed with dialects and industry jargon, and their image recognition function is ineffective at analyzing blurry, shaky short videos taken by farmers. This information, which contains critical risks, is simply filtered out as noise by the system.

[0005] Therefore, despite significant changes occurring in the aquaculture area that were enough to disrupt market expectations, the cloud system remained completely unaware, continuing to draw the erroneous conclusion that the market was stable based on routine data. It wasn't until weeks later, when the impact of the event expanded, generated a formal report, and was captured by the system, that the market reacted with a delayed but dramatic shift, causing traders who relied on the system for decision-making to suffer substantial losses due to a lack of early warning.

[0006] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0007] In view of the shortcomings of the prior art, this application provides a cloud computing-based method and system for processing sea cucumber futures trading data. It has the advantage of being able to effectively identify and warn of potential risk signals in non-traditional information channels that have not been captured by traditional systems, thereby improving the accuracy and foresight of sea cucumber futures trading decisions.

[0008] Firstly, a cloud computing-based method for processing sea cucumber futures trading data, the method comprising the following steps:

[0009] S1: Obtain multi-source heterogeneous raw data related to sea cucumber farming, wherein the multi-source heterogeneous raw data comes from non-traditional information channels;

[0010] S2: Perform multimodal feature extraction on the multi-source heterogeneous raw data to obtain a feature set containing text semantic features, visual features, and speech semantic features;

[0011] S3: Perform statistical correlation analysis on the feature set to identify undefined abnormal co-occurrence patterns based on the co-occurrence frequency among features;

[0012] S4: Using the aforementioned abnormal co-occurrence pattern as the core node, a multi-dimensional association structure containing geographical, temporal, feature, and impact dimensions is constructed by extracting contextual information from the multi-source heterogeneous original data.

[0013] S5: Quantitatively evaluate the confidence level based on the indicator data of each dimension in the multidimensional correlation structure, and output a risk warning signal for sea cucumber futures trading when the confidence level reaches a preset threshold.

[0014] Furthermore, step S1 includes:

[0015] S11: Deploy a distributed data collection program in the cloud to access the non-traditional information channels by simulating user browsing behavior or calling public interfaces;

[0016] S12: Based on preset aquaculture keywords, download the multi-source heterogeneous raw data containing text, comments, videos, and video comments from the non-traditional information channels.

[0017] Furthermore, step S2 includes:

[0018] S21: Clean and segment the multi-source heterogeneous raw data to extract the text semantic features;

[0019] S22: Perform visual enhancement processing on the image or video data in the multi-source heterogeneous raw data, and extract the visual features;

[0020] S23: Perform speech-to-text processing on the speech data in the multi-source heterogeneous raw data to extract the speech semantic features.

[0021] Furthermore, step S21 includes:

[0022] S211: Identify and remove advertising information, emoticons, and link addresses from the multi-source heterogeneous raw data;

[0023] S212: Load a preset dedicated dictionary and use the dedicated dictionary to segment the cleaned data into words. The dedicated dictionary includes sea cucumber farming industry terms and local dialect words.

[0024] Furthermore, step S22 includes:

[0025] S221: Extract keyframe images from the video data according to a preset time interval;

[0026] S222: Perform histogram equalization processing on the keyframe image or the image data to adjust the brightness and contrast, and perform color correction and sharpening filtering processing to improve the recognizability of the visual features, thereby extracting the visual features.

[0027] Step S23 includes:

[0028] S231: Call the cloud-based speech recognition interface to perform preliminary conversion on the speech data;

[0029] S232: When dialect features are detected in the speech data, a preset dialect speech recognition model is invoked for secondary conversion to extract the speech semantic features.

[0030] Furthermore, step S3 includes:

[0031] S31: The feature set is scanned temporally using a sliding window algorithm to count the co-occurrence frequency of each feature within a preset time window;

[0032] S32: Use an unsupervised clustering algorithm to cluster feature combinations whose co-occurrence frequency exceeds a preset frequency threshold, and identify clusters that do not match the preset risk definition as the abnormal co-occurrence pattern.

[0033] Furthermore, step S4 includes:

[0034] S41: Utilize a graph database to use the abnormal co-occurrence pattern as the core node, and backtrack the multi-source heterogeneous original data to extract the associated geographical locations, time series, and event descriptions;

[0035] S42: Establish topological connections between the core node and the geographical location, the time series and the event description in the graph database to generate the multidimensional association structure.

[0036] Furthermore, step S5 includes:

[0037] S51: Extract the persistence index, regional concentration index, and information consistency index from the multidimensional association structure;

[0038] S52: The confidence level is obtained by weighted summation of the persistence index, the regional concentration index, and the information consistency index;

[0039] S53: When the confidence level reaches the preset threshold, a warning notification containing an abnormal pattern description, the affected area, and the confidence level is pushed to the transaction terminal in real time through the message queue service.

[0040] Furthermore, step S52 includes:

[0041] S521: Identify mutually exclusive feature descriptions for the same geographical dimension and the same time dimension in the multidimensional association structure;

[0042] S522: Extract the source credit score and publication timestamp of the multi-source heterogeneous original data corresponding to the mutually exclusive feature description;

[0043] S523: Calculate the time-decrease factor of the mutually exclusive feature description based on the publication timestamp, and correct the information consistency index in combination with the source credit score;

[0044] S524: The confidence level is obtained by weighting and summing the modified information consistency index with the persistence index and the regional concentration index.

[0045] Secondly, a cloud-based sea cucumber futures trading data processing system, characterized in that the system is used to implement the steps of any of the methods described above, the system comprising:

[0046] The data acquisition module is used to acquire multi-source heterogeneous raw data related to sea cucumber farming, which comes from non-traditional information channels;

[0047] The feature extraction module is used to perform multimodal feature extraction on the multi-source heterogeneous raw data to obtain a feature set containing text semantic features, visual features and speech semantic features;

[0048] The pattern recognition module is used to perform statistical correlation analysis on the feature set and identify undefined abnormal co-occurrence patterns based on the co-occurrence frequency between features;

[0049] The structure building module is used to construct a multi-dimensional association structure containing geographical, temporal, feature, and impact dimensions by extracting contextual information from the multi-source heterogeneous raw data, with the abnormal co-occurrence pattern as the core node.

[0050] The early warning output module is used to perform confidence quantification assessment based on the indicator data of each dimension in the multidimensional correlation structure, and output a risk warning signal for sea cucumber futures trading when the confidence level reaches a preset threshold.

[0051] Beneficial effects: The cloud computing-based sea cucumber futures trading data processing method and system proposed in this application acquires multi-source heterogeneous raw data from non-traditional information channels, extracts multi-modal features, identifies undefined abnormal co-occurrence patterns, constructs a multi-dimensional correlation structure, and performs quantitative confidence assessment, ultimately outputting risk warning signals. This method can effectively identify and warn of potential risk signals in non-traditional information channels that have not been captured by traditional systems, thereby improving the accuracy and foresight of sea cucumber futures trading decisions. Attached Figure Description

[0052] Figure 1 A flowchart illustrating a cloud computing-based sea cucumber futures trading data processing method proposed in this application.

[0053] Figure 2 This is a structural diagram of a cloud computing-based sea cucumber futures trading data processing system proposed in this application.

[0054] Figure 3 This is a schematic diagram of a cloud computing-based sea cucumber futures trading data processing system proposed in this application.

[0055] Labeling Explanation: 201, Data Acquisition Module; 202, Feature Extraction Module; 203, Pattern Recognition Module; 204, Structure Construction Module; 205, Early Warning Output Module. Detailed Implementation

[0056] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. The components of the embodiments of this application described and marked in the accompanying drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0057] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0058] Traditional sea cucumber futures trading systems primarily rely on the analysis of structured or semi-structured data sources, such as exchange market data, official meteorological data, and mainstream news reports. While these methods are effective when the market is stable, their early warning capabilities are severely lacking when faced with potential upheavals caused by unknown or sudden factors, such as outbreaks of new sea cucumber diseases or abnormal local marine environments. The core issue is that early warning signals for these emergencies often first appear in non-traditional information channels, such as local aquaculture forums and social media platforms like short video platforms. This information is presented in an unstructured and non-standardized form, and existing systems lack the ability to proactively discover and analyze these non-standard information sources. This results in traders who rely on the system for decision-making suffering significant losses due to the lack of early warnings.

[0059] To solve the above problem, please refer to Figure 1 This application proposes a cloud computing-based method for processing sea cucumber futures trading data, the method comprising the following steps:

[0060] S1: Obtain multi-source heterogeneous raw data related to sea cucumber farming. The multi-source heterogeneous raw data comes from non-traditional information channels.

[0061] S2: Perform multimodal feature extraction on multi-source heterogeneous raw data to obtain a feature set containing text semantic features, visual features, and speech semantic features;

[0062] S3: Perform statistical correlation analysis on the feature set and identify undefined abnormal co-occurrence patterns based on the co-occurrence frequency among features;

[0063] S4: Taking the anomaly co-occurrence pattern as the core node, a multi-dimensional association structure containing geographical, temporal, feature, and impact dimensions is constructed by extracting contextual information from multi-source heterogeneous raw data.

[0064] S5: Quantitatively evaluate the confidence level based on the indicator data of each dimension in the multidimensional correlation structure, and output a risk warning signal for sea cucumber futures trading when the confidence level reaches a preset threshold.

[0065] This application acquires multi-source heterogeneous raw data from non-traditional information channels, extracts multimodal features and identifies abnormal patterns, constructs a multi-dimensional correlation structure and conducts confidence assessment, and finally outputs a risk warning signal. This effectively makes up for the shortcomings of existing systems in processing unstructured and non-standardized data, and significantly improves the timeliness and accuracy of risk warnings for sea cucumber futures trading.

[0066] Specifically, in step S1, multi-source heterogeneous raw data refers to raw information from different channels and in different formats, such as text, images, videos, and audio. This data may contain unstructured or semi-structured content, and it originates from non-traditional information channels. For example, web crawling technology can be used to periodically or in real-time crawl publicly available data related to sea cucumber farming from various social media platforms, professional forums, short video platforms, and other non-traditional information channels.

[0067] In step S2, Natural Language Processing (NLP) techniques can be used for word segmentation, part-of-speech tagging, and named entity recognition to extract semantic features of the text, such as keywords and topic orientation. For image or video data, computer vision techniques, such as image recognition, object detection, and scene classification, can be used to extract visual features, such as the health status of sea cucumbers and abnormal conditions in the aquaculture environment. For speech data, speech recognition technology can be used to convert it into text, and then text semantic features can be extracted, or the acoustic features of the speech, such as tone and speech rate, can be directly extracted to identify emotion or abnormal information.

[0068] In step S3, an abnormal co-occurrence pattern refers to the unexpected high frequency of certain features occurring together within a specific time or spatial range, indicating potential risks or events. Unsupervised clustering algorithms from data mining can be used to analyze the frequency of different features appearing simultaneously in the data. When the co-occurrence frequency of certain feature combinations is significantly higher than the historical average or a preset threshold, and these combinations do not conform to known normal patterns or preset risk definitions, they are identified as abnormal co-occurrence patterns. For example, in a specific area, textual features related to sea cucumber diseases and visual features of water turbidity co-occurring in large quantities within a short period constitute an abnormal co-occurrence pattern.

[0069] In step S4, once an abnormal co-occurrence pattern is identified, the system will backtrack to the original data to extract geographical location information related to the pattern, such as the breeding area, time information, such as the time of the event, specific feature descriptions, such as disease type, environmental indicators, and potential impact information, such as farmers' concerns and market sentiment. This information is organized into a multi-dimensional association graph, where the abnormal co-occurrence pattern serves as the central node, and other dimensional information serves as its associated attributes or nodes.

[0070] In step S5, various indicators in the multidimensional correlation structure can be quantified, such as the duration of the abnormal pattern, the geographical scope involved, the breadth of related information dissemination, and the consistency between different information sources. A comprehensive confidence score is calculated through weighted averages. When this score reaches a preset risk threshold, the system generates and outputs a risk warning signal. This signal may include a detailed description of the abnormal pattern, the potential affected area, the confidence level, and other information, and is pushed to relevant traders or decision-makers in real time.

[0071] The cloud computing-based sea cucumber futures trading data processing method proposed in this application effectively solves the limitations of existing technologies in processing data from non-traditional information channels through a series of closely collaborative steps.

[0072] Furthermore, step S1 includes:

[0073] S11: Deploy distributed data collection programs in the cloud to access non-traditional information channels by simulating user browsing behavior or calling public interfaces;

[0074] S12: Based on preset aquaculture keywords, download multi-source heterogeneous raw data containing text, comments, videos, and video comments from non-traditional information channels.

[0075] Deploying a distributed data collection program in the cloud means building a network of multiple independent collection units. These units can be deployed on different cloud server instances, and can even simulate access requests from different geographical locations. This distributed architecture not only improves the efficiency and concurrency of data collection, but also effectively circumvents anti-scraping strategies employed by some information platforms based on access frequency or source IP address.

[0076] To access non-traditional information channels, data collection programs can employ two parallel technical approaches.

[0077] The first approach is to simulate user browsing behavior. In this implementation, each distributed data scraping program launches a headless browser instance, such as using an automated testing framework like Selenium or Puppeteer to control a browser that doesn't display a graphical user interface. The scraping program writes scripts to simulate the actions of a real user, such as opening the target forum's homepage, logging in with a username and password, browsing the post list, clicking to enter the post details page, scrolling the page to load more comments, and even simulating the behavior of swiping up and down to watch videos on a short video platform. This method is effective in acquiring content generated through dynamically loaded techniques, and because its behavioral patterns are highly similar to those of real users, it is more difficult for target websites to identify and block it.

[0078] The second approach is to call public APIs. Some technically open forums or social media platforms may offer public Application Programming Interfaces (APIs). Data scraping programs can directly call these APIs via HTTP requests, passing parameters according to the format defined in the API documentation, such as keywords and time ranges, to directly obtain structured data, typically in JSON or XML format. The advantages of this method are high data acquisition efficiency, low network overhead, and well-formatted data, eliminating the need for complex page parsing. However, not all non-traditional information channels provide such APIs, and these APIs usually have access frequency restrictions.

[0079] Therefore, a robust data collection strategy combines these two approaches. It prioritizes calling public APIs; if the APIs are unavailable or cannot meet the requirements, it supplements this by launching a program that simulates user browsing behavior.

[0080] After accessing these non-traditional information channels, the data collection program filters and downloads content based on preset aquaculture keywords. These keywords constitute a vocabulary set highly relevant to sea cucumber farming. For example, core terms may include sea cucumber, spiny sea cucumber, and sea cucumber seedlings; status description terms may include disease, mortality, reduced yield, loss of appetite, and slow growth; environmental description terms may include water temperature, salinity, red tide, and water turbidity. The data collection program retrieves these keywords from page content or API returned data, and once a match is found, it downloads the relevant data units completely.

[0081] The downloaded data is multi-source and heterogeneous, specifically including text, such as the main post and all replies to forum threads; comments, specifically user comments below short videos; the video itself, i.e., MP4 or other video files; and video comments, which are the same concept as the aforementioned comments but emphasize their direct relevance to the video content. The reason for downloading both the main content and comments is that the comment section is often a crucial place for information to ferment and cross-verify. An anomaly posted by one user may be confirmed, supplemented, or refuted by multiple other users in the comment section, providing extremely valuable contextual information.

[0082] After acquiring this multi-source heterogeneous raw data containing text, comments, videos, and video comments, multimodal feature extraction needs to be performed on this raw data. Further, step S2 includes:

[0083] S21: Clean and segment the multi-source heterogeneous raw data to extract text semantic features;

[0084] S22: Perform visual enhancement processing on image or video data in multi-source heterogeneous raw data and extract visual features;

[0085] S23: Perform speech-to-text processing on the speech data in the multi-source heterogeneous raw data and extract speech semantic features.

[0086] First, the extraction of semantic features from text is a process of transforming unstructured text into structured information. The first step in this process is to clean and segment the raw data.

[0087] Specifically, step S21 includes:

[0088] S211: Identify and remove advertising information, emoticons, and link addresses from heterogeneous multi-source raw data;

[0089] S212: Load the preset dedicated dictionary and use the dedicated dictionary to segment the cleaned data into words. The dedicated dictionary includes sea cucumber farming industry terms and local dialect words.

[0090] The data cleaning process identifies and removes advertising information, emoticons, and links from the raw data. Advertising information typically appears in fixed patterns and can be identified and removed using regular expressions or keyword matching. Furthermore, both text-based emoticons, such as "^_^", and Unicode-encoded emojis can interfere with subsequent semantic analysis, requiring identification and removal using a specialized emoticon library. Web link addresses, i.e., URLs, also lack direct semantic information and need to be removed.

[0091] After cleaning, the data enters the word segmentation stage. Here, a pre-set specialized dictionary is loaded, and the cleaned data is segmented into words using this dictionary. A specialized dictionary is needed because general-purpose Chinese word segmentation tools encounter difficulties when handling specific industry sectors. This specialized dictionary contains two key categories of vocabulary: one is sea cucumber farming industry terminology, such as sea cucumber seedlings, enclosure farming, bottom seeding, skinning, and trawling, ensuring that these professional terms are segmented as complete units, rather than being incorrectly split into individual Chinese characters. The other category is local dialect vocabulary, because farmers use a lot of local colloquialisms and dialects when communicating; for example, certain regions may use specific words to describe a particular abnormal state of sea cucumbers. Adding these dialect words to the dictionary can greatly improve the accuracy of word segmentation. For example, for the sentence "The sea cucumber seedlings in my pond haven't been eating lately, and the water is a bit reddish," the word segmentation tool loaded with the specialized dictionary can accurately segment meaningful word sequences such as "my family," "pond," "sea cucumber seedlings," "lately," "not eating," "water," "a bit," and "reddish," which constitute the initial semantic features of the text.

[0092] Secondly, the extraction of visual features mainly targets images or videos in the original data. Specifically, step S22 includes:

[0093] S221: Extract keyframe images from video data according to a preset time interval;

[0094] S222: Perform histogram equalization on keyframe images or image data to adjust brightness and contrast, and perform color correction and sharpening filtering to improve the recognizability of visual features, thereby extracting visual features.

[0095] The first step in this process is to extract keyframe images from the video data at preset time intervals. Directly processing the video stream is computationally very expensive, and there is significant information redundancy between adjacent frames. Therefore, by setting a time interval, such as extracting one or two frames per second, the amount of data processing can be greatly reduced while preserving the core visual information of the video.

[0096] After obtaining keyframe images or raw static image data, a series of visual enhancement processes are required to improve the recognizability of visual features. Images or videos taken by fish farmers with their mobile phones are often of poor quality, potentially suffering from issues such as low light, color casts due to underwater shooting, and blurry images due to camera shake. To address this, various image processing techniques are employed for optimization. For example, histogram equalization is performed, a technique that redistributes the pixel grayscale of an image, effectively adjusting its brightness and contrast, making details in overexposed or dark scenes clearer.

[0097] Next, color correction is performed, especially for underwater images. Due to the absorption and scattering of light by water, images often appear bluish or greenish. Color correction algorithms can compensate for this color cast and restore the true colors of objects. Finally, sharpening filtering is applied, such as using the Laplacian operator or Unsharp Mask (USM) sharpening algorithm, to enhance the edges and details of objects in the image, making blurry images clearer. After this series of enhancement processes, the recognizability of visual elements such as the sea cucumber's body shape, surface spots, and water color is significantly improved, facilitating the subsequent extraction of specific visual features such as color histograms, texture features, and edge contours.

[0098] Visual enhancement processing is not merely simple image retouching, but a specialized optimization targeting the unique characteristics of sea cucumber farming environments, such as turbid water and uneven lighting. During histogram equalization, an adaptive histogram equalization algorithm with limited contrast is employed. The keyframe image is divided into multiple non-overlapping sub-regions, and the histogram of each sub-region is calculated and transformed. By limiting the amplification rate of local contrast, the common noise amplification problem in underwater images is suppressed. In the color correction stage, the gray-world hypothesis algorithm is used to estimate the ambient color of underwater lighting. By compensating for the attenuation of the red channel, the true colors of the sea cucumber and its environment, which may appear greenish or bluish due to seawater absorption, are restored. Sharpening filtering utilizes a Laplacian operator to perform convolution operations on the enhanced image, strengthening the edge features between the sea cucumber's surface texture and the background, thereby resulting in more spatially discriminative visual feature vectors extracted subsequently.

[0099] Finally, the extraction of speech semantic features mainly targets audio streams in short videos or user-uploaded pure voice messages. Specifically, step S23 includes:

[0100] S231: Call the cloud-based speech recognition interface to perform preliminary conversion of the speech data;

[0101] S232: When dialect features are detected in the speech data, a preset dialect speech recognition model is called to perform a secondary conversion in order to extract the speech semantic features.

[0102] The core of this process is converting speech into text. Specifically, it first calls a cloud-based speech recognition API to perform preliminary conversion of the speech data. These cloud services, such as the speech recognition APIs provided by Alibaba Cloud, Tencent Cloud, or iFlytek, typically have a high accuracy rate in recognizing Mandarin.

[0103] However, considering that dialects are commonly used in communication among farmers, relying solely on a general speech recognition interface is insufficient. Therefore, when dialect features are detected in the speech data, a pre-defined dialect speech recognition model is invoked for secondary conversion. Dialect feature recognition can be achieved in various ways. For example, if the confidence score of the initially converted text is generally low, or if the text contains a large number of illogical word combinations, these may indicate that the input speech is in a dialect.

[0104] Once dialect characteristics are confirmed, and even specific dialect types are identified, such as Jiaodong dialect in Shandong or Minnan dialect in Fujian, a speech recognition model specifically trained for that dialect is invoked. These dialect models utilize a large amount of specific dialect data during training, thus achieving a better fit between the acoustic features and language model of that dialect, significantly improving the accuracy of speech-to-text conversion. The converted text constitutes the speech-semantic features, which can be incorporated into the aforementioned set of text semantic features for unified analysis.

[0105] After completing multimodal feature extraction and obtaining a feature set containing textual, visual, and phonological semantic features, the next crucial step is to perform statistical correlation analysis on these features to identify undefined anomalous co-occurrence patterns. This process does not rely on any pre-defined risk list but is purely data-driven, discovering new and potential threat signals from massive amounts of information.

[0106] Furthermore, step S3 includes:

[0107] S31: Use the sliding window algorithm to perform a time-series scan of the feature set and count the co-occurrence frequency of each feature within a preset time window;

[0108] S32: Use an unsupervised clustering algorithm to cluster feature combinations whose co-occurrence frequency exceeds a preset frequency threshold, and identify clusters that do not match the preset risk definition as abnormal co-occurrence patterns.

[0109] Specifically, this step includes using a sliding window algorithm to perform a time-series scan of the feature set and to count the co-occurrence frequency of each feature within a preset time window; and using an unsupervised clustering algorithm to cluster feature combinations whose co-occurrence frequency exceeds a preset frequency threshold, identifying clusters that do not match the preset risk definition as abnormal co-occurrence patterns.

[0110] The sliding window algorithm is used to capture the dynamic changes in feature co-occurrence over time. Imagine a fixed-size window continuously sliding forward across a constantly flowing stream of feature data. This window can be a time window, such as the last 24 hours, or a data volume window, such as the last 1000 pieces of information. At each time point, only the data within the current window is analyzed. Within the window, the frequency of co-occurrence of different features is counted. For example, in text data, the number of times the words "sprout" and "white spot" appear together in the same post or comment is counted. In multimodal data, the number of times the text feature "growth slow" and the visual feature "body surface attachment" co-occur in the same short video unit is counted. Through this temporal scanning method, the co-occurrence intensity of various feature combinations can be dynamically tracked.

[0111] When the co-occurrence frequency of certain feature combinations increases significantly within a short period of time and exceeds a preset frequency threshold, these combinations are considered candidate patterns worthy of attention. The frequency threshold can be set as an absolute value, such as co-occurring more than 50 times in 24 hours; or as a relative value, such as the current co-occurrence frequency being more than 5 times the average frequency over the past month.

[0112] For feature combinations whose co-occurrence frequency exceeds a threshold, unsupervised clustering algorithms are further used for analysis. Unsupervised clustering algorithms, such as DBSCAN or K-Means, aim to automatically divide data points into different clusters based on their inherent similarity, without requiring any pre-labeling of the data. Here, each frequently co-occurring feature combination can be considered a data point. Through clustering, combinations that are similar in feature composition can be grouped together.

[0113] For example, the feature combinations of sea cucumber seedlings with white spots indicating non-feeding, sea cucumbers with white spots on their bodies indicating poor vitality, and small sea cucumbers with white spots indicating slow growth all point to similar anomalies both semantically and visually. Unsupervised clustering algorithms can group these similar combinations into a single cluster.

[0114] After clustering is complete, the most crucial step is to match these clusters with a pre-defined risk definition library. This risk definition library stores all known and clearly defined risk events and their corresponding characteristic patterns. For example, the characteristics corresponding to enteritis are filthy stool and skin peeling, while the characteristics corresponding to abnormal water temperature are water temperature above 25 degrees Celsius and floating.

[0115] If a cluster successfully matches a known risk in the risk definition library, it is classified as a known risk event. However, if a cluster has a high density of data points, indicating that the pattern is frequent and the characteristics are stable, but no match can be found in the risk definition library, then this cluster will be ultimately identified as an undefined anomalous co-occurrence pattern. This signifies that a potential, novel risk signal has been successfully discovered from massive noise. Unsupervised clustering here specifically refers to density-based clustering processes used to uncover potential risks without prior risk labels.

[0116] High-frequency co-occurring feature combinations within a preset time window are mapped to a multi-dimensional Euclidean space. Each feature combination forms spatial coordinates based on its frequency of occurrence, associated keyword weights, and geographical distribution. A density clustering algorithm with noise reduction capabilities is employed, setting neighborhood radius thresholds and minimum point count thresholds. Feature clusters are identified by scanning the distribution density of spatial points. When a newly generated feature cluster forms high-density connections in space, and the cosine similarity between its centroid and any feature vector in the pre-stored known risk feature fingerprint database is below a preset threshold, the cluster is determined to be an undefined anomalous co-occurrence pattern. This method can identify novel disaster-causing factors, such as novel microbial attachments or abnormal growth caused by local marine trace element variations.

[0117] Once this undefined anomalous co-occurrence pattern is identified, it is insufficient to merely know the pattern itself; it is also necessary to understand its context, scope, and potential impact. Therefore, further, step S4 includes:

[0118] S41: Utilize graph databases to identify co-occurrence patterns of anomalies as core nodes, and backtrack multi-source heterogeneous original data to extract associated geographical locations, time series, and event descriptions;

[0119] S42: Establish topological connections between core nodes and geographical locations, time series, and event descriptions in the graph database to generate a multidimensional association structure.

[0120] Graph databases, such as Neo4j, are a data storage method particularly well-suited for representing and processing complex relationships. In this scenario, when an abnormal co-occurrence pattern, such as ginseng seedlings – white spots – slow growth, is identified, a core node representing that pattern will be created in the graph database.

[0121] Next, it will backtrack all the original data containing this abnormal co-occurrence pattern, such as forum posts, short videos, and comments, and extract contextual information from them to create new nodes and establish connections with the core nodes.

[0122] At the geographic level, named entity recognition technology from natural language processing is used to automatically identify place names from text content, such as Rongcheng in Weihai, Ningde in Fujian, and the aquaculture area of ​​Changhai County. Simultaneously, short videos or images are checked for geolocation tags. Each identified geographic location is created as a geographic node in the graph database, and a connection is established with the core abnormal pattern nodes to indicate where the occurrence occurred.

[0123] In terms of time, the publication timestamp of each relevant piece of raw data is extracted. These timestamps are created as time nodes in the graph database and connected to the anomalous pattern nodes to represent their occurrence. By aggregating all the time nodes, a complete time series can be formed, allowing analysis of the frequency of the anomalous pattern over time to determine whether it is in its nascent, explosive, or declining stage.

[0124] At the feature dimension, other high-frequency feature descriptions accompanying the abnormal co-occurrence pattern will be further extracted. For example, in the text, words such as turbid water and increased attachment to net cages may be frequently mentioned; visually, the visual feature of shrinking net cages may appear alongside white spots. These supplementary descriptions will be created as feature nodes, and connections representing features will be established with the core nodes.

[0125] In terms of impact, words mentioning negative consequences in the text, such as reduced production, losses, and increased mortality, will be identified. These words will be created as impact nodes, and connections will be established with core nodes to represent potential consequences, used to initially assess the potential harm of the anomalous event.

[0126] By continuously creating new nodes and connections in the graph database, a multidimensional relational structure around anomaly co-occurrence patterns is gradually generated. This structure is not static but continuously grows and evolves as new data is continuously added. For example, when a new post mentions the same anomaly pattern in different geographical locations, a new geographical node and corresponding connection are added to the graph database. This approach effectively organizes previously isolated and scattered weak signals into a structured and visualized knowledge graph, providing a solid foundation for subsequent risk quantification assessment.

[0127] Furthermore, step S5 includes:

[0128] S51: Extract persistence indicators, regional concentration indicators, and information consistency indicators from the multidimensional association structure;

[0129] S52: The confidence level is obtained by weighting and summing the persistence index, the regional concentration index, and the information consistency index.

[0130] S53: When the confidence level reaches the preset threshold, a warning notification containing an abnormal pattern description, the affected area, and the confidence level is pushed to the transaction terminal in real time through the message queue service.

[0131] First, three core quantitative indicators need to be extracted from the multidimensional relational structure constructed from the graph database.

[0132] The persistence metric measures the stability and trend of anomalous patterns over time. This metric can consist of multiple sub-dimensions, such as the total duration of the anomalous pattern since its initial discovery, the number of associated time points, and the growth rate of newly emerging time points in the recent period. A pattern with a long duration and an accelerating frequency of recent occurrences will have a high persistence metric score.

[0133] The geographic concentration index measures the spatial distribution and impact of anomalous patterns. This index can also be composed of multiple sub-dimensions, such as the total number of independent geographic nodes associated with the anomalous pattern, and whether these nodes are concentrated in a few recognized major sea cucumber producing areas. If a large amount of information points to the same core aquaculture area, its geographic concentration index score will be significantly higher than when the information sources are geographically dispersed.

[0134] The Information Consistency Index (ICI) measures the strength and diversity of evidence supporting an anomalous pattern. Its evaluation dimensions include the number of information sources (i.e., how many independent users or accounts reported the anomalous pattern); the degree of cross-validation between information modalities (e.g., whether text descriptions in forums corroborate visual content in short videos); and the reliability of the information sources. An anomalous pattern repeatedly mentioned from multiple different sources and through various modalities will score highly on the ICI.

[0135] After extracting the quantitative values ​​of these three core indicators, they are weighted and summed to calculate a comprehensive confidence score. The calculation formula can be expressed as: Confidence Score = w1 × Persistence Index + w2 × Geographic Concentration Index + w3 × Information Consistency Index. Here, w1, w2, and w3 are preset weight coefficients, and their sum is 1. These weights can be adjusted based on expert experience or historical data backtesting. For example, in some scenarios, geographic concentration may be more important than persistence.

[0136] However, when processing information from non-traditional channels, inconsistencies and even contradictions often arise. To improve the accuracy of confidence assessment, the calculation of the information consistency index needs more refined adjustments. Specifically, step S52 includes:

[0137] S521: Identify mutually exclusive feature descriptions for the same geographic dimension and the same time dimension in a multidimensional association structure;

[0138] S522: Extract the source credit score and publication timestamp of the multi-source heterogeneous raw data corresponding to mutually exclusive feature descriptions;

[0139] S523: Calculate the time-decrease factor of mutually exclusive feature descriptions based on the publication timestamp, and correct the information consistency index by combining the source credit score;

[0140] S524: The confidence level is obtained by weighting and summing the modified information consistency index with the persistence index and the regional concentration index.

[0141] For example, in a multidimensional association structure, it's possible to find two mutually exclusive feature descriptions within the same geographical node and similar time frame. One post from user A describes ginseng seedlings showing white spots and large-scale death; the other comment from user B describes their ginseng seedlings as growing well without any abnormalities. These two pieces of information constitute mutually exclusive feature descriptions.

[0142] At this point, more metadata about these two raw data entries will be extracted. First is the source credit score, which is calculated based on factors such as the quality of a user's historical posts, account registration time, and whether they are a verified user. A seasoned farmer's credit score will be significantly higher than that of a newly registered anonymous account. Second is the posting timestamp.

[0143] Next, a time-depreciation factor is calculated based on the publication timestamp. The value of information diminishes over time; information published a day ago is obviously more valuable than information published a week ago. The time-depreciation factor can be designed as an exponential decay function: Time-depreciation factor = e^(-k*Δt), where e is the base of the natural logarithm; k is the decay coefficient, which can be set empirically, for example, 0.1, representing the rate at which the information's value decays; and Δt is the time difference between the current moment and the information's publication timestamp, in hours. The larger the time difference, the smaller the time-depreciation factor.

[0144] Finally, the source credit score is multiplied by the time-decrease factor, and the result is the adjusted weight, which can be used to adjust the information consistency index. When calculating the information consistency index, the influence of information supporting and denying risk needs to be considered comprehensively. Specifically, the adjusted weight of each piece of information can be regarded as the influence score of that information on the risk judgment. The sum of the adjusted weights of all information pointing to the existence of risk yields the sum of the risk-supporting weights, which reflects the overall influence of the information group supporting the risk; similarly, the sum of the adjusted weights of all information denying the existence of risk yields the sum of the risk-denying weights, which reflects the overall influence of the information group denying the risk.

[0145] Furthermore, the final revised value of the information consistency index can be compared with these two accumulated values ​​using a normalization function, for example:

[0146] The information consistency index is calculated as follows: (Sum of supporting risk weights - Sum of negative risk weights) / (Sum of supporting risk weights + Sum of negative risk weights). This revised information consistency index ranges from -1 to 1, where a positive value indicates that information tends to consistently support the risk, a negative value indicates that information tends to consistently deny the risk, and 0 indicates complete information conflict.

[0147] Finally, the revised information consistency index is weighted and summed with the persistence index and the regional concentration index to obtain the final confidence level.

[0148] This correction mechanism effectively filters out outdated or unreliable information, allowing the information consistency index to more accurately reflect the true situation. The correction of the information consistency index aims to address the problem of mixed true and false information in non-traditional channels. After identifying mutually exclusive feature descriptions for the same geographical location and time period, a dynamic correction model based on source credit score and time-lapse factor is established.

[0149] The source credit score is jointly determined by the historical accuracy of the issuing account's alerts, the account's activity cycle, and the weight of real-name authentication. The timeliness decay factor uses a negative exponential function, with the time difference between the current moment and the information release moment as the independent variable; the larger the time difference, the faster the weight decays. During the correction process, the weights of features that tend to indicate risk in mutually exclusive descriptions are compared with the weights of features that tend to indicate normality. The product of the credit score and the decay factor serves as the voting weight for each party's opinion. If high-credit, high-timeliness information points to risk, even if a low-credit source has a negative description, the information consistency score will still be increased to ensure the sensitivity of the alert, thereby effectively purifying weak signals.

[0150] By using the revised information consistency index and weighting it together with the other two indices, a final, more reliable confidence score is obtained.

[0151] When the final confidence score exceeds a preset threshold, such as 0.75, the abnormal pattern is considered to have evolved from a weak signal into a high-probability risk event. At this point, an early warning mechanism is immediately triggered. The warning notification is pushed in real-time to the trader's trading terminal or monitoring platform via a message queue service, such as Kafka or RabbitMQ, ensuring high throughput and low latency. The notification is clear and concise, including a detailed description of the abnormal pattern, such as suspected novel vitiligo in ginseng seedlings, the potentially affected core areas, and the current quantitative confidence level, such as high (0.85), along with several links to key original information sources for manual verification.

[0152] Please refer to Figure 2 , Figure 3This application also provides a cloud computing-based sea cucumber futures trading data processing system, characterized in that the system is used to implement the steps of any of the above methods, and the system includes:

[0153] Data acquisition module 201 is used to acquire multi-source heterogeneous raw data related to sea cucumber farming. The multi-source heterogeneous raw data comes from non-traditional information channels.

[0154] Feature extraction module 202 is used to perform multimodal feature extraction on multi-source heterogeneous raw data to obtain a feature set containing text semantic features, visual features and speech semantic features;

[0155] The pattern recognition module 203 is used to perform statistical correlation analysis on the feature set and identify undefined abnormal co-occurrence patterns based on the co-occurrence frequency between features.

[0156] The structure building module 204 is used to construct a multi-dimensional association structure with geographic, temporal, feature, and impact dimensions by extracting contextual information from multi-source heterogeneous raw data, with anomaly co-occurrence patterns as the core nodes.

[0157] The early warning output module 205 is used to perform confidence quantification assessment based on indicator data of each dimension in the multidimensional correlation structure, and output a risk warning signal for sea cucumber futures trading when the confidence level reaches a preset threshold.

[0158] Specifically, the data acquisition module 201 can be configured to collect data through various methods. For example, it can use a general-purpose web crawler based on the HTTP / HTTPS protocol to periodically access preset non-traditional information channel websites and download page content. Alternatively, the module can integrate a file upload interface, allowing users to manually upload local data files obtained from non-traditional channels. In another implementation, the data acquisition module 201 can be a simple API client that retrieves pre-processed non-traditional channel data by calling the API interface provided by a third-party data service provider.

[0159] The feature extraction module 202 can be configured to employ basic feature extraction algorithms. For example, for text data, a standard lexical analyzer can be used for word segmentation, and keywords can be extracted as semantic features based on word frequency statistics. For image or video data, basic image processing techniques such as grayscale conversion and edge detection can be used to extract texture or shape features as visual features. For speech data, a general speech recognition service can be called to convert it into text, and then simple text feature extraction can be performed, or acoustic features such as pitch and loudness of the speech can be directly extracted.

[0160] The pattern recognition module 203 can be configured to perform basic statistical analysis. For example, a simple counter can be used to count the number of times different features co-occur within a preset time period and compare it with historical averages. When the co-occurrence frequency exceeds a fixed threshold, it is marked as a potential anomalous pattern.

[0161] The structure building module 204 can be configured to use a traditional relational database to store and manage this related information. For example, multiple tables can be created to store exception patterns, geographic locations, timestamps, feature descriptions, and impact information, and these tables can be linked through foreign keys to form a logical multidimensional relational structure.

[0162] The early warning output module 205 can be configured to perform a simple confidence calculation and notification mechanism. For example, it can perform a simple weighted summation of a few indicators such as the duration of the abnormal pattern and the amount of information involved to obtain a preliminary confidence score. When this score reaches a preset fixed threshold, the system can send an early warning notification containing a brief description of the abnormality to preset recipients via email or SMS service.

[0163] The cloud-based sea cucumber futures trading data processing system proposed in this application represents a significant innovation and advancement compared to existing technologies. Traditional sea cucumber futures trading systems primarily rely on analyzing pre-set structured or semi-structured data sources, such as exchange market data, official meteorological data, and mainstream news reports. These systems suffer from severe deficiencies in their early warning capabilities when faced with unknown or sudden events triggered by non-traditional information channels, failing to effectively capture early risk signals.

[0164] The system of this application, through its data acquisition module 201, can actively acquire and process multi-source heterogeneous raw data from non-traditional information channels, breaking through the limitation of traditional systems that rely solely on regular data. For example, existing systems may not be able to effectively process colloquial discussions in local aquaculture forums or blurry images on short video platforms, while the system of this application can effectively parse these multimodal data through its feature extraction module 202, transforming them into an analyzable feature set.

[0165] Furthermore, the pattern recognition module 203 of this application can identify undefined abnormal co-occurrence patterns from these complex feature sets. This contrasts sharply with existing systems that can only provide early warnings based on preset rules or known patterns, thus enabling the discovery of potential risks that are difficult to capture using traditional methods. The structure construction module 204 then places these abnormal patterns in a multi-dimensional context for comprehensive analysis, while the early warning output module 205 provides quantitative risk assessment and timely early warning push.

[0166] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A cloud computing-based method for processing sea cucumber futures trading data, characterized in that, The method includes the following steps: S1: Obtain multi-source heterogeneous raw data related to sea cucumber farming, wherein the multi-source heterogeneous raw data comes from non-traditional information channels; S2: Perform multimodal feature extraction on the multi-source heterogeneous raw data to obtain a feature set containing text semantic features, visual features, and speech semantic features; S3: Perform statistical correlation analysis on the feature set to identify undefined abnormal co-occurrence patterns based on the co-occurrence frequency among features; S4: Using the aforementioned abnormal co-occurrence pattern as the core node, a multi-dimensional association structure containing geographical, temporal, feature, and impact dimensions is constructed by extracting contextual information from the multi-source heterogeneous original data. S5: Quantitatively evaluate the confidence level based on the indicator data of each dimension in the multidimensional correlation structure, and output a risk warning signal for sea cucumber futures trading when the confidence level reaches a preset threshold.

2. The cloud computing-based sea cucumber futures trading data processing method according to claim 1, characterized in that, Step S1 includes: S11: Deploy a distributed data collection program in the cloud to access the non-traditional information channels by simulating user browsing behavior or calling public interfaces; S12: Based on preset aquaculture keywords, download the multi-source heterogeneous raw data containing text, comments, videos, and video comments from the non-traditional information channels.

3. The cloud computing-based sea cucumber futures trading data processing method according to claim 1, characterized in that, Step S2 includes: S21: Clean and segment the multi-source heterogeneous raw data to extract the text semantic features; S22: Perform visual enhancement processing on the image or video data in the multi-source heterogeneous raw data, and extract the visual features; S23: Perform speech-to-text processing on the speech data in the multi-source heterogeneous raw data to extract the speech semantic features.

4. The cloud computing-based sea cucumber futures trading data processing method according to claim 3, characterized in that, Step S21 includes: S211: Identify and remove advertising information, emoticons, and link addresses from the multi-source heterogeneous raw data; S212: Load a preset dedicated dictionary and use the dedicated dictionary to segment the cleaned data into words. The dedicated dictionary includes sea cucumber farming industry terms and local dialect words.

5. The cloud computing-based sea cucumber futures trading data processing method according to claim 2, characterized in that, Step S22 includes: S221: Extract keyframe images from the video data according to a preset time interval; S222: Perform histogram equalization processing on the keyframe image or the image data to adjust the brightness and contrast, and perform color correction and sharpening filtering processing to improve the recognizability of the visual features, thereby extracting the visual features. Step S23 includes: S231: Call the cloud-based speech recognition interface to perform preliminary conversion on the speech data; S232: When dialect features are detected in the speech data, a preset dialect speech recognition model is invoked for secondary conversion to extract the speech semantic features.

6. The cloud computing-based sea cucumber futures trading data processing method according to claim 1, characterized in that, Step S3 includes: S31: The feature set is scanned temporally using a sliding window algorithm to count the co-occurrence frequency of each feature within a preset time window; S32: Use an unsupervised clustering algorithm to cluster feature combinations whose co-occurrence frequency exceeds a preset frequency threshold, and identify clusters that do not match the preset risk definition as the abnormal co-occurrence pattern.

7. The cloud computing-based sea cucumber futures trading data processing method according to claim 1, characterized in that, Step S4 includes: S41: Utilize a graph database to use the abnormal co-occurrence pattern as the core node, and backtrack the multi-source heterogeneous original data to extract the associated geographical locations, time series, and event descriptions; S42: Establish topological connections between the core node and the geographical location, the time series and the event description in the graph database to generate the multidimensional association structure.

8. The cloud computing-based sea cucumber futures trading data processing method according to claim 1, characterized in that, Step S5 includes: S51: Extract the persistence index, regional concentration index, and information consistency index from the multidimensional association structure; S52: The confidence level is obtained by weighted summation of the persistence index, the regional concentration index, and the information consistency index; S53: When the confidence level reaches the preset threshold, a warning notification containing an abnormal pattern description, the affected area, and the confidence level is pushed to the transaction terminal in real time through the message queue service.

9. A cloud computing-based method for processing sea cucumber futures trading data according to claim 8, characterized in that, Step S52 includes: S521: Identify mutually exclusive feature descriptions for the same geographical dimension and the same time dimension in the multidimensional association structure; S522: Extract the source credit score and publication timestamp of the multi-source heterogeneous original data corresponding to the mutually exclusive feature description; S523: Calculate the time-decrease factor of the mutually exclusive feature description based on the publication timestamp, and correct the information consistency index in combination with the source credit score; S524: The confidence level is obtained by weighting and summing the modified information consistency index with the persistence index and the regional concentration index.

10. A cloud computing-based sea cucumber futures trading data processing system, characterized in that, The system is used to implement the steps of the method according to any one of claims 1-9, the system comprising: The data acquisition module is used to acquire multi-source heterogeneous raw data related to sea cucumber farming, which comes from non-traditional information channels; The feature extraction module is used to perform multimodal feature extraction on the multi-source heterogeneous raw data to obtain a feature set containing text semantic features, visual features and speech semantic features; The pattern recognition module is used to perform statistical correlation analysis on the feature set and identify undefined abnormal co-occurrence patterns based on the co-occurrence frequency between features; The structure building module is used to construct a multi-dimensional association structure containing geographical, temporal, feature, and impact dimensions by extracting contextual information from the multi-source heterogeneous raw data, with the abnormal co-occurrence pattern as the core node. The early warning output module is used to perform confidence quantification assessment based on the indicator data of each dimension in the multidimensional correlation structure, and output a risk warning signal for sea cucumber futures trading when the confidence level reaches a preset threshold.

Images