Breeding feedback big data mining method and system based on artificial intelligence

Abstract

The invention relates to the technical field of artificial intelligence, and relates to a breeding feedback big data mining method and system based on artificial intelligence. According to the method, the target breeding feedback text and the derivative sequence thereof are accurately identified from massive breeding feedback big data, the global breeding feedback text sequence is constructed, key information in the breeding process is effectively integrated, the target breeding feedback mining network is introduced, and the breeding feedback semantic space sequence is set, so that the breeding feedback efficiency is improved. Deep mining and reinforcement of breeding feedback semantics are realized; a target breeding feedback semantic collaboration space is generated by using a semantic space feature collaboration technology, so that the internal relation between texts is strengthened, and the sensitivity and accuracy of abnormal event recognition are also improved; finally, the correlation degree between the target breeding feedback text and the abnormal breeding trigger event is calculated through the target pairing neural network, timely and accurate abnormity early warning and reason analysis are provided for breeding personnel, and rapid response and effective measure taking are facilitated.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and more specifically, to an artificial intelligence-based aquaculture feedback big data mining method and system. Background Art

[0002] With the continuous expansion of the scale of modern aquaculture and the improvement of its intelligence level, the accumulation of data in the aquaculture process has shown explosive growth. These data, especially aquaculture feedback data, contain rich information on aquaculture status, which is of great significance for timely detection of aquaculture problems, prevention of disease outbreaks, and optimization of aquaculture management strategies. However, traditional aquaculture feedback data analysis methods often rely on manual experience and simple statistical methods, which makes it difficult to efficiently extract valuable information from massive and complex data, and are easily affected by subjective factors, resulting in insufficient accuracy and timeliness of the analysis results.

[0003] In the relevant technologies, directly applying existing artificial intelligence technologies to the big data mining of aquaculture feedback still faces many challenges. On the one hand, aquaculture feedback texts are usually unstructured, semantically rich, and closely related to context, requiring effective text processing technology to extract and represent their deep semantic information; on the other hand, abnormal events in the aquaculture process often involve the interaction of multiple factors, and a single data source or analysis dimension is difficult to fully reveal its internal mechanism and scope of influence. Summary of the invention

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present application, the present application embodiment provides a method for mining big data of breeding feedback based on artificial intelligence, the method comprising: Obtain the target breeding feedback text and the collaborative breeding feedback text sequence of the target abnormal breeding trigger event from the breeding feedback big data mining; Determine a derived breeding feedback text sequence of the target breeding feedback text from the collaborative breeding feedback text sequence of the target abnormal breeding trigger event; A global breeding feedback text sequence is generated according to the derived breeding feedback text sequence and the target breeding feedback text, and the global breeding feedback text sequence is loaded into a target breeding feedback mining network; the target breeding feedback mining network includes a set breeding feedback semantic space sequence, the target breeding feedback mining network is generated by knowledge learning based on the collaborative breeding feedback text sequence of the sample abnormal breeding trigger event, the set breeding feedback semantic space in the set breeding feedback semantic space sequence is the target breeding feedback semantic space of the breeding feedback text, and the target breeding feedback semantic space of the breeding feedback text is generated by strengthening the semantic association link of the basic breeding feedback semantic space of the breeding feedback text; Utilizing the target breeding feedback mining network, in the set breeding feedback semantic space sequence, the set breeding feedback semantic space corresponding to each breeding feedback text in the global breeding feedback text sequence is retrieved, and semantic space feature coordination is performed on each retrieved set breeding feedback semantic space to generate a target breeding feedback semantic coordination space of the global breeding feedback text sequence; The breeding feedback semantic space of the target breeding feedback text and the abnormal breeding feature space of the target abnormal breeding trigger event are obtained, and the target breeding feedback semantic collaborative space, the breeding feedback semantic space of the target breeding feedback text, and the abnormal breeding feature space of the target abnormal breeding trigger event are loaded into the target pairing neural network to generate the correlation between the target breeding feedback text and the target abnormal breeding trigger event.

[0005] On the other hand, an embodiment of the present application also provides a breeding feedback mining system, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method. Based on the above aspects, the embodiment of the present application significantly improves the processing efficiency and analysis accuracy of aquaculture feedback data. By accurately identifying the target aquaculture feedback text and its derivative sequence from the massive aquaculture feedback big data, and constructing a global aquaculture feedback text sequence, the key information in the aquaculture process is effectively integrated, providing a comprehensive and in-depth data basis for subsequent analysis. The introduction of the target aquaculture feedback mining network, combined with the setting of the aquaculture feedback semantic space sequence, realizes the deep mining and strengthening of the aquaculture feedback semantics, and significantly enhances the model's ability to understand complex aquaculture situations. Further, the semantic space feature collaborative technology is used to generate the target aquaculture feedback semantic collaborative space, which not only strengthens the intrinsic connection between the texts, but also improves the sensitivity and accuracy of abnormal event identification. Finally, the correlation between the target aquaculture feedback text and the abnormal aquaculture triggering event is calculated through the target pairing neural network, providing timely and accurate abnormal warning and cause analysis for the aquaculture personnel, which helps to respond quickly and take effective measures to reduce aquaculture losses and improve aquaculture benefits. As a result, the aquaculture management level is improved and aquaculture safety is guaranteed. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Figure 1 It is a schematic diagram of the execution flow of the artificial intelligence-based breeding feedback big data mining method provided in the embodiment of the present application.

[0007] Figure 2 It is a schematic diagram of the hardware architecture of the aquaculture feedback mining system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0008] The present application will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of an artificial intelligence-based aquaculture feedback big data mining method provided by an embodiment of the present application. The artificial intelligence-based aquaculture feedback big data mining method is introduced in detail below.

[0009] Step S110, obtaining a target breeding feedback text and a coordinated breeding feedback text sequence of a target abnormal breeding triggering event from the breeding feedback big data.

[0010] In detail, the breeding feedback big data refers to a collection of breeding environment data (such as air temperature, dissolved oxygen, pH value, etc.) collected in real time from multiple farms and feedback text data from breeders, which are usually stored in a big data warehouse for subsequent data mining and analysis. For example, suppose a large pig breeding area has 10 farms, each of which generates hundreds of breeding environment data and dozens of breeding feedback texts every day. These data are transmitted to the server in real time and stored in the big data warehouse. The server can query this warehouse to obtain all breeding feedback data within a specific time period for further analysis.

[0011] The target breeding feedback text refers to the breeding feedback text that is directly related to the event and has important analytical value in the context of a specific abnormal breeding trigger event. These texts usually contain key information, such as symptom descriptions, abnormal phenomena, etc. For example, in the abnormal breeding trigger event of "large-scale death of pigs", the target breeding feedback text may include descriptions such as "a large number of dead pigs were found lying on the ground" and "the activity of the pig herd suddenly decreased". These texts directly point to the core phenomenon of the abnormal event and are therefore selected as target texts for in-depth analysis. The target abnormal breeding trigger event refers to an abnormal event that occurs during the breeding process and may cause breeding problems or losses, such as disease outbreaks, water source deterioration, feed poisoning, etc. These abnormal events are usually identified by abnormal changes in environmental data or specific descriptions of breeding feedback texts. For example, during the breeding process, the server detects a sudden drop in the dissolved oxygen concentration of a farm, accompanied by feedback text from the breeder mentioning "pigs have symptoms of difficulty breathing". Based on this information and the pre-trained anomaly detection model, the server identifies "abnormal dissolved oxygen concentration" as the target abnormal breeding trigger event.

[0012] The collaborative breeding feedback text sequence refers to a collection of all relevant breeding feedback texts collected within a certain period of time before and after the target abnormal breeding trigger event occurs. These texts are arranged in chronological order to reflect the development process and context of the event. For example, after the "mass death of pigs" event occurred, the server retrieved all breeding feedback texts from one week before the event to three days after the event, and arranged them into a collaborative breeding feedback text sequence in chronological order. This sequence contains a series of texts ranging from "decline in pig herd activity" to "discovery of a large number of dead pigs", reflecting the development trajectory of the event.

[0013] That is, in this embodiment, a large pig breeding area has multiple farms, each of which is equipped with an intelligent monitoring system for real-time collection of breeding environment data (such as temperature, dissolved oxygen, pH value, etc.) and recording of feedback texts from breeders. These breeding feedback texts cover various situations in the breeding process, such as the growth status of pigs, disease occurrence, feed feeding effect, etc. The server is responsible for mining valuable information from the breeding feedback big data.

[0014] In detail, the server first collects recent breeding feedback text data from the monitoring system of each farm through an API interface or database query. These breeding feedback text data are stored in the big data warehouse and indexed according to the timestamp and farm identifier. Then, the server uses the pre-trained anomaly detection model to analyze the collected environmental data and identify target abnormal breeding trigger events that may cause breeding problems, such as sudden drop in temperature, abnormal dissolved oxygen concentration, etc. Each target abnormal breeding trigger event is assigned a unique identifier and records the time point and related environmental parameters. For each identified target abnormal breeding trigger event, the server retrieves all breeding feedback texts within a certain period of time before and after the target abnormal breeding trigger event (such as one week before the event to three days after the event) by querying the big data warehouse. These breeding feedback texts are arranged in chronological order to form a collaborative breeding feedback text sequence of the target abnormal breeding trigger event. In the collaborative breeding feedback text sequence, the server can filter out the breeding feedback text directly related to the target abnormal breeding trigger event as the target breeding feedback text according to business rules or keyword matching. For example, if the target abnormal farming trigger event is "large-scale death of pigs", the target farming feedback text may include descriptions such as "decline in pig herd activity" and "dead pigs found".

[0015] Step S120, determining a derived breeding feedback text sequence of the target breeding feedback text from the collaborative breeding feedback text sequence of the target abnormal breeding trigger event.

[0016] In detail, the derived breeding feedback text sequence refers to a text sequence in the collaborative breeding feedback text sequence that is directly or indirectly related to the target breeding feedback text and provides more information about the development process of the target abnormal breeding trigger event. These texts are usually screened out from the collaborative sequence by calculating the semantic relevance. For example, still taking the above explanation as an example, in the collaborative breeding feedback text sequence of "large-scale death of pigs", the server calculated the semantic relevance between each text and the target breeding feedback text (such as "dead pigs found"), and screened out the texts with the highest relevance (such as "decline in pig herd activity" and "abnormal water source detection") as the derived breeding feedback text sequence.

[0017] That is, after determining the target breeding feedback text, the server needs to further analyze the collaborative breeding feedback text sequence to find out other texts that are directly or indirectly related to the target breeding feedback text. These texts may provide more information about the development process of the target abnormal breeding triggering event.

[0018] In detail, for each collaborative farming feedback text in the collaborative farming feedback text sequence, the server first extracts its target farming feedback semantic space, which involves converting the collaborative farming feedback text into a high-dimensional vector representation that captures the key semantic information in the collaborative farming feedback text. The target farming feedback semantic space is the result of strengthening the semantic association links of the basic semantic space, which can more accurately reflect the intrinsic connection between texts.

[0019] Next, the server calculates the semantic relevance between the target breeding feedback text and other texts in the collaborative breeding feedback text sequence, which is usually achieved by calculating the cosine similarity between text vectors or using a more complex semantic matching algorithm. According to the descending order of semantic relevance, the server selects several texts with the highest relevance to the target breeding feedback text from the collaborative breeding feedback text sequence as derived breeding feedback texts. Then, these derived breeding feedback texts are sorted according to their original time order in the collaborative breeding feedback text sequence to generate a derived breeding feedback text sequence of the target breeding feedback text.

[0020] Step S130, generating a global breeding feedback text sequence according to the derived breeding feedback text sequence and the target breeding feedback text, and loading the global breeding feedback text sequence into the target breeding feedback mining network. The target breeding feedback mining network includes a set breeding feedback semantic space sequence, the target breeding feedback mining network is generated by knowledge learning based on the collaborative breeding feedback text sequence of the sample abnormal breeding trigger event, the set breeding feedback semantic space in the set breeding feedback semantic space sequence is the target breeding feedback semantic space of the breeding feedback text, and the target breeding feedback semantic space of the breeding feedback text is generated by strengthening the semantic association link of the basic breeding feedback semantic space of the breeding feedback text.

[0021] In detail, the global breeding feedback text sequence refers to a complete text sequence formed by merging the target breeding feedback text with its derived breeding feedback text sequence. This sequence provides comprehensive information about the target abnormal breeding trigger event, which is used for subsequent high-level semantic analysis and anomaly detection. For example, still taking the above explanation as an example, in the "large-scale death of pigs" event, the server merges the target breeding feedback text "dead pigs found" and its derived breeding feedback text sequence (such as "decline in pig herd activity" and "abnormal water source detection") into a global breeding feedback text sequence. This sequence contains complete information from the occurrence of abnormal phenomena to the final result.

[0022] The target breeding feedback mining network refers to a neural network model designed specifically for target breeding feedback text analysis. It learns the complex associations and temporal features between texts by training a large number of sample abnormal breeding trigger events with collaborative breeding feedback text sequences, and generates high-level feature representations that can reflect the global breeding feedback situation. For example, the server uses a large amount of historical data to train a target breeding feedback mining network. This network can receive global breeding feedback text sequences as input, and extract and collaboratively process text features through a multi-layer neural network structure, and finally generate a high-level feature vector or matrix that can fully reflect the breeding feedback situation.

[0023] The set breeding feedback semantic space sequence refers to a series of breeding feedback semantic spaces defined inside the target breeding feedback mining network. Each set breeding feedback semantic space corresponds to a specific breeding feedback context or semantic concept, which is used to represent the key semantic information in the text. In the target breeding feedback mining network, the set breeding feedback semantic space sequence may include "pig health status space", "water source status space", "feed effect space", etc. Each set breeding feedback semantic space has obtained a corresponding vector representation or matrix form through pre-training or network learning, which is used to capture the semantic information related to the set breeding feedback semantic space in the text.

[0024] The basic breeding feedback semantic space refers to the original breeding feedback semantic space that has not been strengthened by semantic association links, and may only contain basic semantic information in the text. The target breeding feedback semantic space is the result of strengthening the basic space with semantic association links, which can more accurately reflect the intrinsic connection and complex semantic relationship between texts. For example, for a text describing "pigs have difficulty breathing", its basic breeding feedback semantic space may only contain vector representations of the two words "pigs" and "breathing difficulties". The target breeding feedback semantic space after strengthening the semantic association links may also contain complex semantic information such as causal relationships and time sequence relationships between these words.

[0025] That is, in this embodiment, after obtaining the target farming feedback text and its derivative text sequence, the server merges them into a global farming feedback text sequence for more in-depth analysis. Then, the server loads this global farming feedback text sequence into a neural network specially designed for mining target farming feedback.

[0026] In detail, the server places the target breeding feedback text at the beginning of the global breeding feedback text sequence, followed by the derived breeding feedback text sequence. If the derived text sequence is long, the first N texts can be intercepted as needed to maintain the rationality of the sequence. The constructed global breeding feedback text sequence is loaded into the target breeding feedback mining network, which is trained on the collaborative breeding feedback text sequence of a large number of sample abnormal breeding trigger events, and can capture the complex associations and temporal characteristics between breeding feedback texts. The target breeding feedback mining network contains a set breeding feedback semantic space sequence, and each set breeding feedback semantic space corresponds to a specific breeding feedback scenario.

[0027] In the target farming feedback mining network, each text is first mapped to its corresponding set farming feedback semantic space. Then, the target farming feedback mining network extracts features and performs collaborative processing on these spaces through a multi-layer neural network structure to generate a high-level feature representation that can reflect the global farming feedback situation.

[0028] Step S140, using the target breeding feedback mining network, retrieves the set breeding feedback semantic space corresponding to each breeding feedback text in the global breeding feedback text sequence in the set breeding feedback semantic space sequence, performs semantic space feature collaboration on each retrieved set breeding feedback semantic space, and generates a target breeding feedback semantic collaboration space of the global breeding feedback text sequence.

[0029] In this embodiment, in the target breeding feedback mining network, the server performs deep processing on each breeding feedback text in the global breeding feedback text sequence, extracts the corresponding set breeding feedback semantic space, and generates a collaborative space that can fully reflect the semantic content of the entire sequence through feature collaboration. For each text in the global breeding feedback text sequence, the server uses the retrieval unit in the target breeding feedback mining network to find the semantic space that best matches it in the set breeding feedback semantic space sequence. This process may involve vector similarity calculation, hash table search and other technologies. The retrieved set breeding feedback semantic spaces are sent to the collaborative unit for processing. The collaborative unit performs feature fusion and collaboration on these set breeding feedback semantic spaces through advanced algorithms such as attention mechanisms and graph neural networks to capture the interactions and mutual influences between the set breeding feedback semantic spaces. After feature collaborative processing, the server generates a target breeding feedback semantic collaborative space, which is a high-dimensional vector or matrix that contains the comprehensive semantic information of all texts in the global breeding feedback text sequence, and can be used for subsequent tasks such as correlation analysis and anomaly detection.

[0030] Step S150, obtaining the breeding feedback semantic space of the target breeding feedback text and the abnormal breeding feature space of the target abnormal breeding trigger event, and loading the target breeding feedback semantic collaborative space, the breeding feedback semantic space of the target breeding feedback text, and the abnormal breeding feature space of the target abnormal breeding trigger event into the target pairing neural network to generate the correlation between the target breeding feedback text and the target abnormal breeding trigger event.

[0031] Finally, the server needs to evaluate the degree of association between the target breeding feedback text and the target abnormal breeding trigger event. To this end, the semantic space of the target breeding feedback text, the feature space of the target abnormal breeding trigger event, and the semantic coordination space of the global breeding feedback text sequence are obtained and input into the target pairing neural network for calculation.

[0032] In detail, the server first obtains the farming feedback semantic space of the target farming feedback text and the abnormal farming feature space of the target abnormal farming trigger event through the previous processing steps. The abnormal farming feature space may contain key information such as abnormal values ​​of environmental parameters and the time point when the abnormality occurs.

[0033] Next, the server loads the farming feedback semantic space of the target farming feedback text, the abnormal farming feature space of the target abnormal farming trigger event, and the target farming feedback semantic collaborative space into the target pairing neural network, which is specially designed to calculate the correlation between the text and the target abnormal farming trigger event, and may contain multiple hidden layers, attention mechanisms, and fusion layers. Components. In the target pairing neural network, spatial information from different sources is fused together, and a correlation score is generated through complex nonlinear transformations. The correlation score reflects the closeness or causal relationship between the target farming feedback text and the target abnormal farming trigger event. Finally, the server outputs the correlation score, which can be used in subsequent decision support, abnormal warning or report generation scenarios. For example, if the correlation score is high, it may mean that the feedback text is of great value for understanding the cause or impact of the target abnormal farming trigger event, and it needs to attract the special attention of the farmers.

[0034] Based on the above steps, the embodiment of the present application significantly improves the processing efficiency and analysis accuracy of aquaculture feedback data. By accurately identifying the target aquaculture feedback text and its derivative sequence from the massive aquaculture feedback big data, and constructing a global aquaculture feedback text sequence, the key information in the aquaculture process is effectively integrated, providing a comprehensive and in-depth data basis for subsequent analysis. The introduction of the target aquaculture feedback mining network, combined with the setting of the aquaculture feedback semantic space sequence, realizes the deep mining and strengthening of the aquaculture feedback semantics, and significantly enhances the model's ability to understand complex aquaculture situations. Further, the semantic space feature collaborative technology is used to generate the target aquaculture feedback semantic collaborative space, which not only strengthens the intrinsic connection between the texts, but also improves the sensitivity and accuracy of abnormal event identification. Finally, the correlation between the target aquaculture feedback text and the abnormal aquaculture triggering event is calculated through the target pairing neural network, providing timely and accurate abnormal warning and cause analysis for the aquaculture personnel, which helps to respond quickly and take effective measures to reduce aquaculture losses and improve aquaculture benefits. As a result, the aquaculture management level is improved and aquaculture safety is guaranteed.

[0035] In a possible implementation, step S120 includes: Step S121, for the target breeding feedback text and the collaborative breeding feedback text sequence of the target abnormal breeding trigger event, extract the target breeding feedback semantic space of the breeding feedback text. The target breeding feedback semantic space is generated by reinforcing the semantic association links of the basic breeding feedback semantic space of the breeding feedback text.

[0036] For example, suppose a large pig farming area has multiple intelligently monitored farms, and the server is responsible for collecting farming environment data and feedback text from farmers from these farms. Recently, the server identified a target abnormal farming trigger event - "large-scale death of pigs", and hopes to conduct an in-depth analysis of the cause and process of the event.

[0037] Specifically, the server has retrieved all the breeding feedback texts within one week before and after the target abnormal breeding trigger event "mass death of pigs" according to the target abnormal breeding trigger event, forming a collaborative breeding feedback text sequence. Now, the server needs to extract the target breeding feedback semantic space for these collaborative breeding feedback texts.

[0038] For example, the server first preprocesses each text in the collaborative farming feedback text sequence, including removing stop words and normalizing punctuation. Then, the server uses a pre-trained semantic model (such as BERT) to convert each text into a high-dimensional vector representation, which is the basic farming feedback semantic space. In order to more accurately reflect the intrinsic connection between texts, the server inputs the basic farming feedback semantic space into the target semantic association link reinforcement network. The target semantic association link reinforcement network strengthens the semantic space through specific algorithms (such as graph attention network, Transformer, etc.) to generate the target farming feedback semantic space. This process may involve identifying key entities in the text (such as farming varieties, disease names), constructing a relationship graph between entities, and strengthening these relationships through technologies such as graph neural networks.

[0039] Step S122, according to the semantic association between the target breeding feedback semantic spaces of the target breeding feedback text and the collaborative breeding feedback text, and according to the descending order of the semantic association, determine multiple derivative breeding feedback texts of the target breeding feedback text from the collaborative breeding feedback text sequence of the target abnormal breeding trigger event.

[0040] The server has extracted the target farming feedback semantic space for each collaborative farming feedback text in the collaborative farming feedback text sequence. Next, the server needs to calculate the semantic relevance between the target farming feedback text (such as "dead pigs found") and other texts in the sequence to determine the derived farming feedback text.

[0041] In detail, the server calculates the cosine similarity between the target breeding feedback text and the target breeding feedback semantic space of each text in the collaborative breeding feedback text sequence, or uses a more complex semantic matching algorithm (such as similarity calculation based on the attention mechanism) to obtain the semantic relevance. According to the descending order of semantic relevance, the server selects several texts with the highest relevance to the target breeding feedback text from the collaborative breeding feedback text sequence as derived breeding feedback texts. For example, texts such as "decline in pig herd activity" and "abnormal water source detection" may be selected.

[0042] Step S123, sorting each derived breeding feedback text according to the coordinated temporal position with the target abnormal breeding trigger event, to generate a derived breeding feedback text sequence of the target breeding feedback text.

[0043] In this embodiment, the server has determined multiple derivative breeding feedback texts of the target breeding feedback text. Now, the server needs to sort these derivative breeding feedback texts according to the original time order in the collaborative breeding feedback text sequence to generate a derivative breeding feedback text sequence.

[0044] Specifically, the server queries the timestamp of each derived breeding feedback text in the big data warehouse and sorts the derived breeding feedback texts in the order of the timestamps. After the sorting is completed, the server combines the sorted derived breeding feedback texts into a sequence as the derived breeding feedback text sequence of the target breeding feedback text. The derived breeding feedback text sequence reflects the possible changes and event development trajectories in the breeding process from before the target abnormal breeding trigger event occurs to the appearance of the target breeding feedback text.

[0045] Through the above steps, the server determines the derived breeding feedback text sequence of the target breeding feedback text "discovery of dead pigs" from the collaborative breeding feedback text sequence of the target abnormal breeding trigger event "large-scale death of pigs". The derived breeding feedback text sequence can include texts such as "decline in pig herd activity" and "abnormal water source detection". They are arranged in chronological order, providing important clues for understanding the development process of the target abnormal breeding trigger event. These derived breeding feedback texts and the associations between them will be further used for the construction of the global breeding feedback text sequence and the generation of the target breeding feedback semantic collaborative space, ultimately helping the server evaluate the correlation between the target breeding feedback text and the target abnormal breeding trigger event.

[0046] In a possible implementation, step S121 includes: Step S1211, generating a current breeding feedback label attribute set according to the breeding feedback label attribute of the current breeding feedback text, and loading the current breeding feedback label attribute set into the target semantic association link reinforcement network. The current breeding feedback text is the target breeding feedback text or the collaborative breeding feedback text in the collaborative breeding feedback text sequence of the target abnormal breeding trigger event, and the target semantic association link reinforcement network includes a basic breeding feedback semantic space sequence.

[0047] For example, a server in a large pig farming area is responsible for processing and analyzing real-time data from multiple farms, including farming environment data and feedback text from farmers. The server has identified a target abnormal farming trigger event (such as "large-scale death of pigs") and retrieved the relevant collaborative farming feedback text sequence. Now, the server needs to extract the target farming feedback semantic space for these collaborative farming feedback texts (including target farming feedback texts).

[0048] In detail, for each current breeding feedback text in the collaborative breeding feedback text sequence, the server first identifies and extracts its breeding feedback label attributes, which may include breeding varieties, disease names, symptom descriptions, feed types, etc.

[0049] For example, the server uses natural language processing technology (such as named entity recognition, keyword extraction, etc.) to analyze the text and identify key entities and descriptive words as breeding feedback label attributes. Then, all identified label attributes are organized into a set, namely the current breeding feedback label attribute set. For example, for a text describing "decreased activity of pigs", its label attribute set may include {"pigs", "decreased activity"}.

[0050] Step S1212, using the target semantic association link reinforcement network, retrieve the basic breeding feedback semantic space corresponding to the breeding feedback tag attribute of the current breeding feedback text in the basic breeding feedback semantic space sequence, perform semantic association link reinforcement on the retrieved basic breeding feedback semantic space, and generate the target breeding feedback semantic space of the current breeding feedback text.

[0051] The server has built a target semantic association link reinforcement network, which is specifically used to strengthen the semantic association links of the farming feedback text to generate a more accurate target farming feedback semantic space. Now, the server needs to load the current farming feedback label attribute set into this target semantic association link reinforcement network.

[0052] Specifically, ensure that the target semantic association link reinforcement network has been trained and is in a usable state. The target semantic association link reinforcement network contains a sequence of basic farming feedback semantic spaces, which are pre-defined and used to represent different types of farming feedback situations. Then, the current farming feedback label attribute set is loaded into the target semantic association link reinforcement network as input data, which usually involves converting the label attributes into a format that the network can understand (such as a vector representation) and passing it to the input layer of the network.

[0053] In the target semantic association link reinforcement network, the server needs to use the label attribute set of the current breeding feedback text to retrieve the corresponding basic breeding feedback semantic space, and perform semantic association link reinforcement processing on these spaces. For example, the retrieval unit in the target semantic association link reinforcement network searches for the most matching semantic space in the basic breeding feedback semantic space sequence according to the current breeding feedback label attribute set. This process may involve multiple technologies such as vector similarity calculation and hash table search. Based on the retrieved basic breeding feedback semantic space, the semantic association link reinforcement unit in the target semantic association link reinforcement network will perform reinforcement processing on it, which usually involves analyzing the association relationship between label attributes (such as causal relationship, parallel relationship, etc.), and enhancing the representation of these relationships in the semantic space through graph neural network, attention mechanism and other technologies. After reinforcement processing, the target semantic association link reinforcement network generates the target breeding feedback semantic space of the current breeding feedback text. The target breeding feedback semantic space is a high-dimensional vector or matrix that can more accurately reflect the key semantic information in the text and the intrinsic connection between label attributes.

[0054] Through the above steps, the server extracts the target farming feedback semantic space for each current farming feedback text in the collaborative farming feedback text sequence. This process involves the identification and set construction of label attributes, the loading of data into the target semantic association link reinforcement network, and the retrieval and reinforcement processing of the semantic space. The target farming feedback semantic space finally generated provides important support for subsequent tasks such as semantic association calculation and anomaly detection.

[0055] In a possible implementation, the method further includes: Step A110, obtaining a first breeding feedback label attribute set, and inputting the first breeding feedback label attribute set into a basic semantic association link reinforcement network. The first breeding feedback label attribute set is generated based on the collaborative breeding feedback text sequence of the first sample abnormal breeding trigger event, and the basic semantic association link reinforcement network includes the basic breeding feedback semantic space sequence.

[0056] In this embodiment, the server can select a representative first-example abnormal breeding trigger event (such as "large-scale death of pigs"). Based on the first-example abnormal breeding trigger event, the server retrieves the collaborative breeding feedback text sequence within a period of time before and after the event. Next, using natural language processing technology and domain knowledge base, the server extracts key breeding feedback label attributes from the collaborative breeding feedback text sequence, such as breeding varieties, symptom descriptions, environmental parameter changes, etc., and constructs a first breeding feedback label attribute set.

[0057] Thus, the server inputs the first farming feedback tag attribute set into the pre-built basic semantic association link reinforcement network. The basic semantic association link reinforcement network contains a basic farming feedback semantic space sequence for representing different types of farming feedback situations.

[0058] For example, first ensure that the basic semantic association link reinforcement network has been initialized and is in a trainable state. Then, load the first breeding feedback label attribute set into the input layer of the network in a certain format (such as a vector sequence).

[0059] Step A120, using the basic semantic association link reinforcement network, in the basic breeding feedback semantic space sequence, retrieve the basic breeding feedback semantic space corresponding to each breeding feedback label attribute before the first label node in the first breeding feedback label attribute set, perform semantic association link reinforcement on each retrieved basic breeding feedback semantic space, generate an estimated breeding feedback semantic space corresponding to the first label node, and generate a first estimated confidence corresponding to the breeding feedback label attribute on the first label node in the first breeding feedback label attribute set based on the estimated breeding feedback semantic space corresponding to the first label node. The first label node is determined from each breeding feedback label node in the first breeding feedback label attribute set, and the first estimated confidence is used to reflect the estimated derivative metric between the breeding feedback text corresponding to the first label node and the forward breeding feedback text sequence corresponding to the first label node.

[0060] Step A130: Generate a first training error according to the first estimated confidences corresponding to the breeding feedback label attributes on each first label node in the first breeding feedback label attribute set.

[0061] Step A140, updating the weight and bias information of the basic semantic association link enhancement network according to the first training error until the first training termination requirement is met, thereby generating the target semantic association link enhancement network.

[0062] In this embodiment, in the basic semantic association link reinforcement network, the server traverses each label node in the first breeding feedback label attribute set. For each first label node, the server retrieves the basic breeding feedback semantic space corresponding to all previous label nodes, and performs semantic association link reinforcement processing on these basic breeding feedback semantic spaces.

[0063] For the first label node, the retrieval unit in the network searches for the corresponding basic farming feedback semantic space in the previous label node. The semantic association link reinforcement unit is used to reinforce all the retrieved basic farming feedback semantic spaces to capture the intrinsic connection and temporal characteristics between the label attributes. After reinforcement, the estimated farming feedback semantic space corresponding to the first label node is generated.

[0064] Based on the estimated farming feedback semantic space corresponding to the first label node, the server generates a first estimated confidence corresponding to the farming feedback label attribute on the first label node, and the first estimated confidence reflects the estimated derivative metric between the label attribute and its forward farming feedback text sequence. For example, the prediction unit in the network uses the estimated farming feedback semantic space to calculate the first estimated confidence through a specific algorithm (such as logistic regression, neural network output layer, etc.).

[0065] The server generates a first training error based on the first estimated confidence of all first label nodes in the first breeding feedback label attribute set and a certain metric (such as cross entropy loss) between the actual label attributes. The server then uses this error to update the weight and bias information of the basic semantic association link reinforcement network.

[0066] For example, the first estimated confidence level can be compared with the actual label attribute to calculate a first training error. Using optimization techniques such as the back propagation algorithm, the weight and bias information of the network are updated according to the first training error. This process may involve multiple iterations until the training error is reduced to below a certain threshold or a preset training round is reached.

[0067] After enough training iterations, the performance of the basic semantic association link reinforcement network reaches a stable state. At this time, the server regards it as the target semantic association link reinforcement network and uses it for subsequent real-time data processing and analysis tasks.

[0068] During the training process, the network's performance is regularly evaluated to ensure that its performance on the validation set meets expectations. Once the network meets the training termination requirements (such as the error is below the threshold, the training round is completed, etc.), the server saves it as the target semantic association link reinforcement network for subsequent use.

[0069] Through the above steps, the server trained the basic semantic association link reinforcement network and generated the target semantic association link reinforcement network. This process involves the selection of sample events, retrieval of text sequences, extraction of label attributes, input and output processing of the network, calculation of training errors and network updates. The target network finally generated can more accurately capture the semantic associations and temporal features between the aquaculture feedback texts, providing strong support for subsequent tasks such as aquaculture anomaly detection and cause analysis.

[0070] In a possible implementation, step A120 includes: Step A121, performing semantic feature domain conversion on the estimated farming feedback semantic space corresponding to the first label node to generate a basic semantic feature domain vector distribution corresponding to the first label node. The basic semantic feature domain vector distribution includes semantic feature domain vectors corresponding to each reference farming feedback text in the reference farming feedback text sequence.

[0071] In this embodiment, the server has generated a corresponding estimated farming feedback semantic space for the first tag node through the basic semantic association link reinforcement network. Now, the server needs to convert this estimated farming feedback semantic space into a semantic feature domain for further analysis and processing.

[0072] Specifically, the server first parses and estimates the farming feedback semantic space, extracts the key semantic information and feature vectors contained therein, and then, using a predefined mapping rule or model, maps these feature vectors into a semantic feature domain, generating a basic semantic feature domain vector distribution, wherein each vector in the basic semantic feature domain vector distribution represents a semantic feature of a reference farming feedback text in a reference farming feedback text sequence.

[0073] Step A122, semantic dependency enhancement is performed on the basic semantic feature domain vector distribution to generate a target semantic feature domain vector distribution corresponding to the first label node. The target semantic feature domain vector distribution includes the first estimated confidence corresponding to each reference breeding feedback text in the reference breeding feedback text sequence, and the reference breeding feedback text sequence includes the breeding feedback text corresponding to each breeding feedback label attribute in the first breeding feedback label attribute set.

[0074] After obtaining the basic semantic feature domain vector distribution, the server needs to further strengthen the semantic dependencies between these vectors to more accurately reflect the intrinsic connections and temporal characteristics between the reference breeding feedback texts.

[0075] Specifically, the server analyzes the potential dependencies between the vectors in the basic semantic feature domain vector distribution, such as causal relationships, parallel relationships, etc. Then, using advanced algorithms such as graph neural networks and attention mechanisms, the server performs enhanced processing on these vectors to enhance the semantic dependencies between them. The processed vector distribution is the target semantic feature domain vector distribution.

[0076] Step A123: determining, from the target semantic feature domain vector distribution, a first estimated confidence corresponding to the breeding feedback label attribute on the first label node in the first breeding feedback label attribute set.

[0077] After semantic dependency enhancement, the server obtains the target semantic feature domain vector distribution. Now, the server needs to determine the first estimated confidence corresponding to the breeding feedback label attribute on the first label node in the first breeding feedback label attribute set from this distribution.

[0078] For example, the server first finds the corresponding vector of the reference farming feedback text in the target semantic feature domain vector distribution that matches the farming feedback label attribute on the first label node. Since each vector in the target semantic feature domain vector distribution contains an estimated confidence (which can be directly generated by an output layer of the network or indirectly inferred by the relationship between the vector and other vectors), the server directly extracts the corresponding first estimated confidence from this matching vector.

[0079] Finally, the extracted first estimated confidence is recorded for subsequent training error calculation and network weight update.

[0080] Through the above steps, the server generates the first estimated confidence for the farming feedback label attribute on the first label node. This process involves multiple links such as the conversion of semantic feature domains, the strengthening of semantic dependencies, and the extraction of confidence. The first estimated confidence reflects the accuracy of the network's estimate of the relationship between the label attribute and its forward farming feedback text sequence, and is an important basis for evaluating network performance and guiding network training. In subsequent iterative training, the server will continue to adjust the network's weights and bias information based on these confidences to gradually improve the network's accuracy and generalization ability.

[0081] In a possible implementation, the basic semantic association link reinforcement network includes a retrieval unit, a mapping unit, a semantic association link reinforcement unit and a prediction unit, wherein the retrieval unit is used to retrieve the basic breeding feedback semantic space, the mapping unit is used for semantic association link mapping, the semantic association link reinforcement unit is used for semantic association link reinforcement, and the prediction unit is used to generate a first estimated confidence.

[0082] Step A140 includes: updating the weights and bias information of the mapping unit, the semantic association link reinforcement unit and the prediction unit in the basic semantic association link reinforcement network according to the first training error until the first training termination requirement is met, thereby generating the target semantic association link reinforcement network.

[0083] In this embodiment, in the intelligent monitoring system of a large-scale pig breeding area, the server is responsible for training and optimizing the basic semantic association link reinforcement network to build a target network that can accurately understand and process the complex semantic associations between breeding feedback texts. This process involves the collaborative work of multiple network components, including retrieval units, mapping units, semantic association link reinforcement units, and prediction units.

[0084] In detail, the basic semantic association link reinforcement network consists of multiple specially designed units, each of which undertakes a specific task.

[0085] Retrieval unit: responsible for retrieving the semantic space related to the currently processed label node from the basic breeding feedback semantic space sequence.

[0086] Mapping unit: maps the retrieved semantic space into a feature domain suitable for further processing.

[0087] Semantic association link reinforcement unit: enhances the semantic dependency between vectors in the feature domain to reflect the intrinsic connection between texts.

[0088] Prediction unit: Generates the estimated confidence of the currently processed label node based on the enhanced feature domain vector distribution.

[0089] During the training process, the server calculates the first training error based on the difference between the estimated confidence generated by the network and the actual label attribute. For example, the server compares the first estimated confidence generated by the prediction unit with the actual label attribute obtained by some means (such as manual annotation, historical data verification, etc.). Using a predefined loss function (such as cross entropy loss, mean square error, etc.), the server calculates the difference between the first estimated confidence and the actual label attribute to obtain the first training error.

[0090] After obtaining the first training error, the server will use this first training error to guide the update of network weights and bias information to optimize network performance. The server uses the back propagation algorithm to propagate the first training error backward layer by layer from the prediction unit to the mapping unit and retrieval unit. During the back propagation process, the server calculates the adjustment amount of weights and biases in each network unit based on the error gradient, and applies these adjustments to update the weights and bias information of the network.

[0091] Iterative optimization: This process is repeated for multiple iterations, and each iteration uses new training data to generate new estimated confidence and training errors, and further adjusts the network weights and biases accordingly. As the iterations proceed, the performance of the network will gradually improve.

[0092] After a certain number of iterative training, if the network performance reaches the preset standard or the training error is reduced to an acceptable range, the server will stop the training process and save the current network state as the target semantic association link reinforcement network. After each iteration, the server will evaluate the performance of the network (such as accuracy, recall rate, F1 score, etc.) and compare it with the preset training termination requirements. If the network performance reaches the training termination requirements or the training rounds reach the preset upper limit, the server will stop the training process. Finally, the server saves the current optimized basic semantic association link reinforcement network as the target semantic association link reinforcement network for subsequent use in real-time data processing and analysis tasks.

[0093] Through the above steps, the server trained and optimized the basic semantic association link reinforcement network and generated a target network that can accurately capture the complex semantic associations between farming feedback texts. This process involves the collaborative work of network components, the calculation of training errors, the updating of network weights and bias information, and the judgment of training termination conditions. The generation of the target network not only improves the accuracy and efficiency of the intelligent monitoring system, but also provides more reliable and useful data analysis support for farmers.

[0094] In a possible implementation, the method further includes: Step B110, obtaining the text knowledge point distribution corresponding to each reference breeding feedback text in the reference breeding feedback text sequence.

[0095] Step B120, through the graph self-attention processing network, the graph self-attention semantic mining is performed on the distribution of the text knowledge points to generate the basic breeding feedback semantic space corresponding to each reference breeding feedback text.

[0096] Step B130, generating the basic breeding feedback semantic space sequence according to the basic breeding feedback semantic spaces corresponding to the reference breeding feedback texts.

[0097] In this embodiment, the server retrieves a series of reference breeding feedback texts from the stored historical data, and these reference breeding feedback texts cover breeding conditions at different times and under different breeding conditions. The server then performs natural language processing on each reference breeding feedback text to extract the knowledge point distribution in the reference breeding feedback text.

[0098] In detail, the server retrieves a reference breeding feedback text sequence from the database according to preset query conditions (such as time range, breeding species, breeding problems, etc.). For each reference breeding feedback text in the reference breeding feedback text sequence, the server uses named entity recognition, keyword extraction and other technologies to extract key entities and descriptive words in the text as knowledge points. These knowledge points may include breeding species, disease names, symptom descriptions, environmental parameters, etc. The server organizes the extracted knowledge points into a distribution form for subsequent semantic mining. This distribution can be a knowledge point set, a knowledge graph or any other data structure suitable for representing text content.

[0099] Next, the server uses a pre-trained graph self-attention processing network to perform semantic mining on the knowledge point distribution of the reference farming feedback text. This process aims to capture the intrinsic connections and contextual information between knowledge points to generate more accurate and rich semantic representations.

[0100] In detail, ensure that the graph self-attention processing network has been trained and is in an available state. The network may contain multiple graph convolutional layers, self-attention layers, and fully connected layers. The server converts the knowledge point distribution of each reference farming feedback text into an input format acceptable to the network (such as graph structure data, node feature vectors, etc.) and inputs it into the network. In the graph self-attention processing network, node features are propagated and aggregated through the graph convolution layer, and the self-attention layer is used to capture the complex dependencies between nodes. After multiple layers of processing, the network finally outputs the basic farming feedback semantic space corresponding to each text. This space is a high-dimensional vector or matrix that contains the semantic information of the key knowledge points in the text and the associations between them.

[0101] Finally, the server organizes the basic aquaculture feedback semantic space corresponding to each reference aquaculture feedback text according to their order in the original text sequence to form an ordered basic aquaculture feedback semantic space sequence, which provides important data support for subsequent advanced semantic analysis and anomaly detection tasks. The server sorts the generated basic aquaculture feedback semantic space according to the timestamp or other sorting criteria (such as breeding cycle, processing priority, etc.) of the reference aquaculture feedback text sequence. The sorted basic aquaculture feedback semantic space is added to the sequence in sequence to form a complete basic aquaculture feedback semantic space sequence, which can be saved in the database for subsequent use or directly passed to other processing modules for further analysis.

[0102] Through the above steps, the server extracts the distribution of text knowledge points from the reference farming feedback text, and conducts in-depth semantic mining through the graph self-attention processing network. The final generated basic farming feedback semantic space sequence not only contains rich semantic information, but also maintains the temporal relationship and internal connection between texts. This sequence provides a solid foundation for subsequent complex farming data analysis, anomaly detection and decision support.

[0103] In a possible implementation, the steps of obtaining the text knowledge point distribution corresponding to each reference breeding feedback text in the reference breeding feedback text sequence, and performing graph self-attention semantic mining on the text knowledge point distribution through a graph self-attention processing network to generate a basic breeding feedback semantic space corresponding to each reference breeding feedback text, include: Step C110, performing word segmentation processing on each reference breeding feedback text in the reference breeding feedback text sequence, dividing the continuous reference breeding feedback text into a word segmentation sequence, and after part-of-speech tagging of the word segmentation sequence, using named entity recognition technology to extract a key entity sequence in the word segmentation sequence, wherein the key entities in the key entity sequence include breeding varieties, disease names, and symptom descriptions.

[0104] In this embodiment, the server first performs word segmentation processing on each reference breeding feedback text in the reference breeding feedback text sequence, and divides the continuous reference breeding feedback text into independent word segmentation sequences. Subsequently, the server performs part-of-speech tagging on these word segmentation sequences to distinguish words with different parts of speech, such as nouns, verbs, and adjectives.

[0105] Specifically, the server uses a pre-trained word segmentation model (such as Jieba word segmentation) to segment the text and obtain a word segmentation sequence. A part-of-speech tagging tool (such as HanLP) is used to tag each word in the word segmentation sequence, providing a basis for subsequent named entity recognition.

[0106] After obtaining the word segmentation sequence and part-of-speech tagging, the server uses named entity recognition technology to extract key entity sequences from the word segmentation sequence. These key entities include breeding species, disease names, symptom descriptions, and other entities that are crucial to understanding the text content. For example, the server uses a named entity recognition model (such as BiLSTM-CRF) to scan the word segmentation sequence and identify the key entities in the text. The identified key entities are organized into a key entity sequence in the order in which they appear in the text.

[0107] Step C120, extracting the core words in the key entity sequence as knowledge points through a keyword extraction algorithm, and using the extracted knowledge points as nodes, establishing edges according to the co-occurrence relationship of the nodes in the corresponding reference breeding feedback text, constructing a knowledge point graph, and assigning an initial vector representation to each node in the knowledge point graph.

[0108] In this embodiment, the server extracts core vocabulary as knowledge points from the key entity sequence through a keyword extraction algorithm, and constructs a knowledge point graph with these knowledge points as nodes. The edges in the graph represent the co-occurrence relationship between nodes (knowledge points). For example, keyword extraction algorithms such as TF-IDF and TextRank can be used to extract core vocabulary from key entity sequences. With the extracted knowledge points as nodes, edges are established based on their co-occurrence relationship in the reference breeding feedback text. The co-occurrence relationship can be determined by counting the number of times the nodes appear simultaneously in the text. Then, an initial vector representation is assigned to each node in the knowledge point graph. This vector can be random or obtained through a pre-trained word embedding model (such as Word2Vec, GloVe).

[0109] Step C130, input the initial vector representation of each node into an embedding layer to obtain the target feature embedding data, and for each node in the knowledge point graph, calculate the attention score of the node and all other nodes, the attention score reflects the degree of dependence of the current node on other nodes, the attention score is calculated through dot product, concatenation and then realized by neural network, and the attention score is normalized using softmax function to obtain the attention weight of each node to other nodes.

[0110] Step C140, according to the attention weight, weighted aggregation is performed on the features of the neighboring nodes of each of the nodes to obtain an updated feature representation of the current node, and the updated feature representation after aggregation is linearly transformed and nonlinearly activated to generate a target feature representation corresponding to each node. For each reference breeding feedback text, the target feature representation of the corresponding node is integrated, and the integrated target feature representation is mapped to the basic breeding feedback semantic space through a fully connected layer.

[0111] In this embodiment, the server uses the graph self-attention processing network to perform semantic mining on the knowledge point graph, updates the feature representation of the nodes by calculating the attention scores between nodes and performing feature aggregation, and finally generates a basic breeding feedback semantic space corresponding to each reference breeding feedback text.

[0112] In detail, the initial vector representation of each node can be input into an embedding layer (such as a multi-layer perceptron, graph convolutional network, etc.) to obtain the target feature embedding data. For each node in the knowledge point graph, the attention score of the node and all other nodes is calculated, which reflects the degree of dependence of the current node on other nodes. The calculation of the attention score can be realized by dot product, concatenation and then through a neural network, and normalized using the softmax function to obtain the attention weight. According to the attention weight, the neighbor node features of each node are weighted aggregated to obtain the updated feature representation of the current node. This process takes into account the complex dependencies between nodes. The aggregated updated feature representation is linearly transformed and nonlinearly activated (such as ReLU activation function) to generate the target feature representation corresponding to each node. For each reference farming feedback text, the target feature representations of all corresponding nodes are integrated (such as summation, average, etc.), and the integrated feature representation is mapped to the basic farming feedback semantic space through a fully connected layer. The basic farming feedback semantic space is a high-dimensional vector or matrix that contains the semantic information of key knowledge points in the text and the association between them.

[0113] Through the above steps, the server extracted the distribution of text knowledge points from the reference farming feedback text, and conducted in-depth semantic mining through the graph self-attention processing network. The basic farming feedback semantic space finally generated not only captures the key knowledge point information in the text, but also considers the complex dependencies and contextual information between knowledge points. These basic farming feedback semantic spaces provide strong data support for subsequent advanced semantic analysis, anomaly detection and decision support.

[0114] Furthermore, in a practical example, suppose a large pig farming company has deployed an intelligent monitoring system in multiple farms to monitor water source parameters (such as temperature, pH value, dissolved oxygen content) in real time and record the feedback text of farmers. Recently, the company has noticed that a pig disease (such as "bacterial sepsis") occurs frequently in multiple farms, so it decided to use the server to conduct in-depth analysis of historical farming feedback texts to extract key knowledge points and build a basic farming feedback semantic space to provide support for subsequent disease warning and prevention and control.

[0115] The server obtained a reference breeding feedback text: "Recently, it was found that the activity of black pigs in the pig pen has decreased, and some pigs have red spots on their bodies, which are suspected to be early symptoms of bacterial sepsis." The server uses the pre-trained word segmentation model to segment the text and obtain the word segmentation sequence: "It has been found recently that the activity of black pigs in the pig pen has decreased, and some pigs have red spots on their bodies, which are suspected to be early symptoms of bacterial sepsis." Using a part-of-speech tagging tool (such as HanLP), the server performs part-of-speech tagging on the word sequence, and the results may include nouns (such as "black pig", "red spot"), verbs (such as "decline", "appear"), adjectives (such as "initial"), etc.

[0116] The server uses a named entity recognition model (such as BiLSTM-CRF) to scan the word segmentation sequence and identify key entities: "black pig" (farming variety), "erythema" (symptom description), "bacterial sepsis" (disease name). These entities are organized into a key entity sequence according to the order in which they appear in the text.

[0117] The server extracts core words from the key entity sequence as knowledge points: "black pig", "erythema", "bacterial sepsis". Then, a knowledge point graph is constructed with these knowledge points as nodes.

[0118] The server counts the number of co-occurrences of these knowledge points in the reference farming feedback text sequence. For example, "black pig" and "red spot" may appear simultaneously in multiple texts, indicating that they are related. Based on these co-occurrence relationships, the server creates edges in the knowledge point graph.

[0119] Assign an initial vector representation to each node. These vectors can be random or obtained through a pre-trained word embedding model (such as Word2Vec) to contain the initial semantic information of the vocabulary.

[0120] The server inputs the initial vector of each node into the embedding layer to obtain the target feature embedding data. Then, for each node in the knowledge point graph, it calculates the attention score between it and all other nodes.

[0121] Attention scores are calculated by dot product, concatenation and neural network, which reflect the degree of dependency between nodes. For example, the attention score of the "black pig" node to the "red spot" node may be high because they often appear together in the text describing the disease.

[0122] The attention scores are normalized using the softmax function to obtain the attention weight of each node to other nodes.

[0123] The server performs weighted aggregation on the neighbor node features of each node according to the attention weight and updates the feature representation of the node.

[0124] For the “black pig” node, the server will perform weighted aggregation on the features of neighboring nodes such as “erythema” and “bacterial sepsis” according to its attention weights with other nodes to obtain an updated feature representation of the “black pig” node.

[0125] Perform linear transformation and nonlinear activation (such as ReLU) on the aggregated feature representation to generate the target feature representation corresponding to each node.

[0126] For the entire reference breeding feedback text, the server integrates (such as average pooling) the target feature representations of all corresponding nodes (such as "black pig", "red spot", and "bacterial sepsis") and maps them to the basic breeding feedback semantic space through a fully connected layer. The basic breeding feedback semantic space is a high-dimensional vector that contains the semantic information of key knowledge points in the text and the associations between them.

[0127] In a possible implementation, the method further includes: Step D110, obtaining a second breeding feedback label attribute set, and inputting the second breeding feedback label attribute set into a basic breeding feedback mining network. The second breeding feedback label attribute set is generated based on the collaborative breeding feedback text sequence of the second sample abnormal breeding trigger event, and the basic breeding feedback mining network includes the set breeding feedback semantic space sequence.

[0128] In this embodiment, the server first selects a second sample abnormal breeding trigger event (such as "slow growth of pigs"), and generates a second breeding feedback label attribute set from the collaborative breeding feedback text sequence of the event, and the second breeding feedback label attribute set includes key breeding feedback label attributes related to the second sample abnormal breeding trigger event.

[0129] In detail, the server selects a representative second sample abnormal breeding trigger event from the historical data. Based on the event, the server retrieves the collaborative breeding feedback text sequence within a period of time before and after the second sample abnormal breeding trigger event occurs. Then, using natural language processing technology and domain knowledge base, the server extracts key breeding feedback label attributes from the collaborative breeding feedback text sequence and constructs a second breeding feedback label attribute set.

[0130] The server inputs the second breeding feedback tag attribute set into a pre-built basic breeding feedback mining network. The basic breeding feedback mining network contains a set breeding feedback semantic space sequence for representing different types of breeding feedback situations.

[0131] For example, first, ensure that the basic farming feedback mining network has been initialized and is in a trainable state. Then, load the second farming feedback label attribute set into the input layer of the basic farming feedback mining network in a certain format (such as a serialized label attribute vector).

[0132] Step D120, using the basic breeding feedback mining network, in the set breeding feedback semantic space sequence, retrieve the set breeding feedback semantic space corresponding to each breeding feedback label attribute before the second label node in the second breeding feedback label attribute set, perform semantic space feature coordination on each retrieved set breeding feedback semantic space, generate an estimated breeding feedback semantic space corresponding to the second label node, and generate a second estimated confidence corresponding to the breeding feedback label attribute on the second label node in the second breeding feedback label attribute set based on the estimated breeding feedback semantic space corresponding to the second label node. The second label node is determined from each breeding feedback label node in the second breeding feedback label attribute set, and the second estimated confidence is used to reflect the estimated derivative metric between the breeding feedback text corresponding to the second label node and the forward breeding feedback text sequence corresponding to the second label node.

[0133] In this embodiment, in the basic breeding feedback mining network, the server traverses each second label node in the second breeding feedback label attribute set. For each second label node, the server retrieves the set breeding feedback semantic space corresponding to all previous label nodes, and performs semantic space feature collaborative processing on these set breeding feedback semantic spaces to generate an estimated breeding feedback semantic space corresponding to the second label node.

[0134] Specifically, for each second label node, the retrieval unit in the network searches for the corresponding set aquaculture feedback semantic space in the previous label node. The collaborative unit performs feature collaborative processing on the retrieved multiple set aquaculture feedback semantic spaces to capture the intrinsic connections and temporal features between them. After collaborative processing, the estimated aquaculture feedback semantic space corresponding to the second label node is generated.

[0135] Based on the estimated farming feedback semantic space corresponding to the second label node, the server generates a second estimated confidence corresponding to the farming feedback label attribute on the node, and the confidence reflects the estimated derivative metric between the label attribute and its forward farming feedback text sequence. For example, the prediction unit in the network uses the estimated farming feedback semantic space to calculate the second estimated confidence through a specific algorithm (such as logistic regression, neural network output layer, etc.).

[0136] Step D130: Generate a second training error according to the second estimated confidences corresponding to the breeding feedback label attributes on each second label node in the second breeding feedback label attribute set.

[0137] Step D140, updating the weight and bias information of the basic aquaculture feedback mining network according to the second training error until the second training termination requirement is met, thereby generating the target aquaculture feedback mining network.

[0138] In this embodiment, the server generates a second training error based on the second estimated confidence on all second label nodes in the second breeding feedback label attribute set and a certain metric (such as cross entropy loss) between the actual label attributes. Then, the server uses this second training error to update the weight and bias information of the basic breeding feedback mining network. In detail, the second estimated confidence can be compared with the actual label attribute to calculate the second training error. Then, using optimization techniques such as the back propagation algorithm, the weight and bias information of the network are updated according to the second training error. This process may involve multiple iterations until the training error is reduced below a certain threshold or a preset training round is reached.

[0139] After enough training iterations, the performance of the basic farming feedback mining network reaches a stable state. At this time, the server regards it as the target farming feedback mining network and uses it for subsequent real-time data processing and analysis tasks. During the training process, the performance of the network is regularly evaluated to ensure that its performance on the validation set meets expectations. Once the network meets the training termination requirements (such as the error is lower than the threshold, the training round is completed, etc.), the server saves it as the target farming feedback mining network for subsequent use.

[0140] Through the above steps, the server trained the basic aquaculture feedback mining network and generated the target aquaculture feedback mining network. This process involves extracting the label attribute set from the collaborative aquaculture feedback text sequence of the second sample abnormal aquaculture trigger event, inputting the network for training, generating estimated semantic space and confidence, calculating training errors, and updating network weights and biases. The target network finally generated can more accurately understand and process the complex semantic associations and temporal features in the aquaculture feedback data, providing strong support for subsequent tasks such as aquaculture anomaly detection and cause analysis.

[0141] In a possible implementation, the basic breeding feedback mining network includes a retrieval unit, a coordination unit and a prediction unit. The retrieval unit is used to retrieve and set the breeding feedback semantic space, the coordination unit is used for semantic space feature coordination, and the prediction unit is used to output a second estimated confidence.

[0142] Step D140 includes: updating the weights and bias information of the collaborative units and prediction units in the basic aquaculture feedback mining network according to the second training error until the second training termination requirement is met, thereby generating the target aquaculture feedback mining network.

[0143] In this embodiment, the basic farming feedback mining network is a complex deep learning model, which consists of three main parts: retrieval unit, collaboration unit and prediction unit. Each unit undertakes a specific task and jointly supports the overall function of the network.

[0144] Retrieval unit: responsible for retrieving the semantic space related to the currently processed label node from the set breeding feedback semantic space sequence.

[0145] Collaborative Unit: Co-process features of multiple retrieved semantic spaces to capture their intrinsic connections and temporal features.

[0146] Prediction unit: Based on the feature representation after collaborative processing, output the second estimated confidence of the currently processed label node.

[0147] During the training process, the server calculates the second training error based on the difference between the second estimated confidence generated by the network and the actual label attribute. This error is an important indicator for measuring network performance and is used to guide subsequent weight and bias updates.

[0148] In detail, the server compares the second estimated confidence output by the prediction unit with the actual label attribute obtained by some means (such as manual annotation, historical data verification, etc.). Then, using a predefined loss function (such as cross entropy loss, mean square error, etc.), the server calculates the difference between the second estimated confidence and the actual label attribute to obtain a second training error.

[0149] After obtaining the second training error, the server will use this second training error to update the weight and bias information of the collaborative unit and the prediction unit in the basic farming feedback mining network. Since the retrieval unit is mainly responsible for data retrieval and does not involve complex calculation processes, its weight and bias usually remain unchanged during the training process.

[0150] In detail, the server uses the back-propagation algorithm to propagate the second training error backwards from the prediction unit to the collaborative unit layer by layer. During the propagation process, the error will be distributed to each weight and bias in each unit. According to the error gradient, the server calculates the adjustment amount for each weight and bias in the collaborative unit and the prediction unit. The calculation of the adjustment amount may involve the setting of hyperparameters such as learning rate and momentum. The server applies the calculated adjustment amount to the weights and biases of the collaborative unit and the prediction unit to update the values ​​of these parameters. The updated parameters will be used for the next round of forward propagation and error calculation.

[0151] The above error calculation, weight and bias update process will be repeated for multiple iterations until the network performance reaches the preset standard or the training error is reduced to an acceptable range. Each iteration uses new training data to generate a new second estimate confidence and training error. After each iteration, the server will evaluate the performance of the network (such as accuracy, recall, F1 score, etc.) and compare it with the preset training termination requirements. If the network performance reaches the training termination requirements or the training rounds reach the preset upper limit, the server will stop the training process. Finally, the server saves the currently optimized basic farming feedback mining network as the target farming feedback mining network for subsequent use in real-time data processing and analysis tasks.

[0152] Through the above steps, the server optimized the basic farming feedback mining network and generated the target farming feedback mining network. This process involves iteratively updating the weights and bias information of the collaborative unit and the prediction unit to minimize the second training error. As the iteration proceeds, the performance of the network gradually improves and eventually reaches the preset training termination requirements. The generation of the target farming feedback mining network not only improves the accuracy and efficiency of the intelligent monitoring system, but also provides farmers with more reliable and useful data analysis support.

[0153] In a possible implementation, step S140 includes: Step S141, according to the text arrangement order of the global farming feedback text sequence, the retrieved set farming feedback semantic spaces are aggregated to generate a set farming feedback semantic space sequence.

[0154] Step S142, performing semantic space feature coordination on the set farming feedback semantic space sequence to generate a farming feedback semantic coordination space set. The farming feedback semantic coordination space set includes farming feedback semantic coordination spaces that match the number of texts in the global farming feedback text sequence.

[0155] Step S143, obtaining the breeding feedback semantic collaborative space on the breeding feedback text label node of the target breeding feedback text from the breeding feedback semantic collaborative space set as the target breeding feedback semantic collaborative space of the global breeding feedback text sequence.

[0156] In this embodiment, the server first gathers the previously retrieved set breeding feedback semantic spaces in an orderly manner according to the text arrangement order of the global breeding feedback text sequence to generate a set breeding feedback semantic space sequence. In detail, the server reviews the previous processing steps to ensure that the corresponding set breeding feedback semantic space has been retrieved for each global breeding feedback text. According to the order of the texts in the global breeding feedback text sequence, the server arranges these set breeding feedback semantic spaces in sequence to generate a set breeding feedback semantic space sequence.

[0157] After obtaining the set breeding feedback semantic space sequence, the server needs to further process these spaces using feature collaboration technology to capture the intrinsic connections and temporal features between them and generate a breeding feedback semantic collaboration space set. In detail, the server selects a suitable feature collaboration algorithm (such as attention network, long short-term memory network LSTM, etc.) to process the set breeding feedback semantic space sequence. The selected feature collaboration algorithm is applied to iteratively process the set breeding feedback semantic space sequence. During the processing, the feature representations in each set breeding feedback semantic space and the dependencies between them can be considered to generate new feature representations. These new feature representations not only contain the information in the original set breeding feedback semantic space, but also incorporate relevant information in other set breeding feedback semantic spaces. After multiple iterative processing, the server generates a breeding feedback semantic collaboration space set, each breeding feedback semantic collaboration space in the breeding feedback semantic collaboration space set corresponds to a breeding feedback text in the global breeding feedback text sequence, and contains the comprehensive semantic information between the breeding feedback text and the breeding feedback texts before and after it.

[0158] Finally, the server needs to extract the farming feedback semantic collaborative space on the farming feedback text label node of the target farming feedback text from the farming feedback semantic collaborative space set as the target farming feedback semantic collaborative space of the global farming feedback text sequence. In detail, the server first determines the position of the target farming feedback text in the global farming feedback text sequence, that is, its corresponding farming feedback text label node. According to the position of the label node, the server finds the corresponding farming feedback semantic collaborative space from the farming feedback semantic collaborative space set. This space not only contains the semantic information of the target farming feedback text itself, but also integrates the comprehensive semantic information between its previous and next texts, so it can more comprehensively reflect the context of the target farming feedback text. The server saves the extracted target farming feedback semantic collaborative space for subsequent tasks such as correlation analysis and anomaly detection.

[0159] Through the above steps, the server performs feature collaborative processing on each set farming feedback semantic space in the global farming feedback text sequence and generates a target farming feedback semantic collaborative space. This process not only considers the temporal relationship between texts, but also captures the intrinsic connection and dependency between them through feature collaborative technology. The target farming feedback semantic collaborative space finally generated provides more comprehensive and accurate data support for subsequent semantic analysis, anomaly detection and other tasks.

[0160] Figure 2 The hardware structure of the aquaculture feedback mining system 100 provided in the embodiment of the present application for implementing the above-mentioned artificial intelligence-based aquaculture feedback big data mining method is shown, as shown in FIG. Figure 2 As shown, the farming feedback mining system 100 may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .

[0161] In a possible design, the farming feedback mining system 100 can be a single server or a server group. The server group can be centralized or distributed (for example, the farming feedback mining system 100 can be a distributed system). In some embodiments, the farming feedback mining system 100 can be local or remote. For example, the farming feedback mining system 100 can access information and / or data stored in the machine-readable storage medium 120 via a network. For another example, the farming feedback mining system 100 can be directly connected to the machine-readable storage medium 120 to access the stored information and / or data. In some embodiments, the farming feedback mining system 100 can be implemented on a farming feedback mining system. As an example only, the farming feedback mining system can include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, etc. or any aggregation thereof.

[0162] The machine-readable storage medium 120 can store data and / or instructions. In some embodiments, the machine-readable storage medium 120 can store data obtained from an external terminal. In some embodiments, the machine-readable storage medium 120 can store data and / or instructions that the farming feedback mining system 100 uses to execute or use to complete the exemplary methods described in this application.

[0163] During the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the artificial intelligence-based breeding feedback big data mining method as described in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected via the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0164] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned breeding feedback mining system 100. The implementation principles and technical effects are similar, and this embodiment will not be repeated here.

[0165] In addition, an embodiment of the present application also provides a readable storage medium, in which computer executable instructions are preset. When the processor executes the computer executable instructions, the above-mentioned artificial intelligence-based breeding feedback big data mining method is implemented.

[0166] It should be noted that in order to simplify the description disclosed in this application and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, multiple features are sometimes combined into one embodiment, drawings, or descriptions thereof. Similarly, it should be noted that in order to simplify the description disclosed in this application and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, multiple features are sometimes combined into one embodiment, drawings, or descriptions thereof.

Claims

1. A method for mining big data of breeding feedback based on artificial intelligence, It is characterized in that The method comprises: Obtain the target breeding feedback text and the collaborative breeding feedback text sequence of the target abnormal breeding trigger event from the breeding feedback big data mining; Determine a derived breeding feedback text sequence of the target breeding feedback text from the collaborative breeding feedback text sequence of the target abnormal breeding trigger event; A global breeding feedback text sequence is generated according to the derived breeding feedback text sequence and the target breeding feedback text, and the global breeding feedback text sequence is loaded into a target breeding feedback mining network; the target breeding feedback mining network includes a set breeding feedback semantic space sequence, the target breeding feedback mining network is generated by knowledge learning based on the collaborative breeding feedback text sequence of the sample abnormal breeding trigger event, the set breeding feedback semantic space in the set breeding feedback semantic space sequence is the target breeding feedback semantic space of the breeding feedback text, and the target breeding feedback semantic space of the breeding feedback text is generated by strengthening the semantic association link of the basic breeding feedback semantic space of the breeding feedback text; Utilizing the target breeding feedback mining network, in the set breeding feedback semantic space sequence, the set breeding feedback semantic space corresponding to each breeding feedback text in the global breeding feedback text sequence is retrieved, and semantic space feature coordination is performed on each retrieved set breeding feedback semantic space to generate a target breeding feedback semantic coordination space of the global breeding feedback text sequence; The breeding feedback semantic space of the target breeding feedback text and the abnormal breeding feature space of the target abnormal breeding trigger event are obtained, and the target breeding feedback semantic collaborative space, the breeding feedback semantic space of the target breeding feedback text, and the abnormal breeding feature space of the target abnormal breeding trigger event are loaded into the target pairing neural network to generate the correlation between the target breeding feedback text and the target abnormal breeding trigger event.

2. The method for mining big data of aquaculture feedback based on artificial intelligence according to claim 1, It is characterized in that The step of determining the derived breeding feedback text sequence of the target breeding feedback text from the coordinated breeding feedback text sequence of the target abnormal breeding trigger event includes: For the target breeding feedback text and the collaborative breeding feedback text sequence of the target abnormal breeding trigger event, extract the target breeding feedback semantic space of the breeding feedback text; the target breeding feedback semantic space is generated by strengthening the semantic association link of the basic breeding feedback semantic space of the breeding feedback text; According to the semantic association between the target breeding feedback semantic spaces of the target breeding feedback text and the collaborative breeding feedback text, a plurality of derivative breeding feedback texts of the target breeding feedback text are determined from the collaborative breeding feedback text sequence of the target abnormal breeding trigger event according to the descending order of the semantic association; The derived breeding feedback texts are sorted according to the coordinated temporal position with the target abnormal breeding trigger event to generate a derived breeding feedback text sequence of the target breeding feedback text.

3. The method for mining big data of aquaculture feedback based on artificial intelligence according to claim 2, It is characterized in that The step of extracting the target breeding feedback semantic space of the breeding feedback text from the target breeding feedback text and the breeding feedback text sequence of the target abnormal breeding trigger event comprises: Generate a current breeding feedback label attribute set according to the breeding feedback label attribute of the current breeding feedback text, and load the current breeding feedback label attribute set into the target semantic association link reinforcement network; the current breeding feedback text is the target breeding feedback text or the collaborative breeding feedback text in the collaborative breeding feedback text sequence of the target abnormal breeding trigger event, and the target semantic association link reinforcement network includes a basic breeding feedback semantic space sequence; The target semantic association link reinforcement network is used to retrieve the basic breeding feedback semantic space corresponding to the breeding feedback tag attribute of the current breeding feedback text in the basic breeding feedback semantic space sequence, and the semantic association link reinforcement is performed on the retrieved basic breeding feedback semantic space to generate the target breeding feedback semantic space of the current breeding feedback text.

4. The method for mining big data of aquaculture feedback based on artificial intelligence according to claim 3, It is characterized in that The method further comprises: Acquire a first breeding feedback label attribute set, and input the first breeding feedback label attribute set into a basic semantic association link reinforcement network; the first breeding feedback label attribute set is generated according to the collaborative breeding feedback text sequence of the first sample abnormal breeding trigger event, and the basic semantic association link reinforcement network includes the basic breeding feedback semantic space sequence; Using the basic semantic association link reinforcement network, in the basic breeding feedback semantic space sequence, the basic breeding feedback semantic spaces corresponding to the breeding feedback label attributes before the first label node in the first breeding feedback label attribute set are retrieved, and the semantic association link reinforcement is performed on the retrieved basic breeding feedback semantic spaces to generate an estimated breeding feedback semantic space corresponding to the first label node. According to the estimated breeding feedback semantic space corresponding to the first label node, a first estimated confidence corresponding to the breeding feedback label attribute on the first label node in the first breeding feedback label attribute set is generated; the first label node is determined from the breeding feedback label nodes in the first breeding feedback label attribute set, and the first estimated confidence is used to reflect the estimated derivative metric value between the breeding feedback text corresponding to the first label node and the forward breeding feedback text sequence corresponding to the first label node; Generate a first training error according to first estimated confidences respectively corresponding to the breeding feedback label attributes on each first label node in the first breeding feedback label attribute set; The weight and bias information of the basic semantic association link reinforcement network are updated according to the first training error until the first training termination requirement is met, thereby generating the target semantic association link reinforcement network.

5. The method for mining big data of aquaculture feedback based on artificial intelligence according to claim 4, It is characterized in that The step of generating, based on the estimated farming feedback semantic space corresponding to the first label node, a first estimated confidence corresponding to the farming feedback label attribute on the first label node in the first farming feedback label attribute set comprises: The estimated breeding feedback semantic space corresponding to the first label node is converted into a semantic feature domain to generate a basic semantic feature domain vector distribution corresponding to the first label node; the basic semantic feature domain vector distribution includes semantic feature domain vectors corresponding to each reference breeding feedback text in the reference breeding feedback text sequence; Performing semantic dependency enhancement on the basic semantic feature domain vector distribution to generate a target semantic feature domain vector distribution corresponding to the first label node; the target semantic feature domain vector distribution includes first estimated confidences corresponding to each reference breeding feedback text in the reference breeding feedback text sequence, and the reference breeding feedback text sequence includes breeding feedback texts corresponding to each breeding feedback label attribute in the first breeding feedback label attribute set; A first estimated confidence corresponding to the breeding feedback label attribute on the first label node in the first breeding feedback label attribute set is determined from the target semantic feature domain vector distribution.

6. The method for mining big data of aquaculture feedback based on artificial intelligence according to claim 4, It is characterized in that The basic semantic association link reinforcement network includes a retrieval unit, a mapping unit, a semantic association link reinforcement unit and a prediction unit, wherein the retrieval unit is used to retrieve the basic farming feedback semantic space, the mapping unit is used for semantic association link mapping, the semantic association link reinforcement unit is used for semantic association link reinforcement, and the prediction unit is used to generate a first estimated confidence; The updating of the weight and bias information of the basic semantic association link reinforcement network according to the first training error until the first training termination requirement is met to generate the target semantic association link reinforcement network includes: The weights and bias information of the mapping unit, the semantic association link reinforcement unit and the prediction unit in the basic semantic association link reinforcement network are updated according to the first training error until the first training termination requirement is met, thereby generating the target semantic association link reinforcement network.

7. The method for mining big data of aquaculture feedback based on artificial intelligence according to claim 3, It is characterized in that The method further comprises: Obtain the text knowledge point distribution corresponding to each reference breeding feedback text in the reference breeding feedback text sequence; Through the graph self-attention processing network, the graph self-attention semantic mining is performed on the distribution of the text knowledge points to generate the basic breeding feedback semantic space corresponding to each reference breeding feedback text; Generate the basic breeding feedback semantic space sequence according to the basic breeding feedback semantic spaces corresponding to the reference breeding feedback texts; The steps of obtaining the text knowledge point distribution corresponding to each reference breeding feedback text in the reference breeding feedback text sequence, and performing graph self-attention semantic mining on the text knowledge point distribution through a graph self-attention processing network to generate a basic breeding feedback semantic space corresponding to each reference breeding feedback text, include: Perform word segmentation processing on each reference breeding feedback text in the reference breeding feedback text sequence, divide the continuous reference breeding feedback text into word segmentation sequences, and perform part-of-speech tagging on the word segmentation sequences, and then use named entity recognition technology to extract key entity sequences in the word segmentation sequences, wherein the key entities in the key entity sequences include breeding varieties, disease names, and symptom descriptions; Extract the core words in the key entity sequence as knowledge points by using a keyword extraction algorithm, and use the extracted knowledge points as nodes, establish edges according to the co-occurrence relationship of the nodes in the corresponding reference breeding feedback text, construct a knowledge point graph, and assign an initial vector representation to each node in the knowledge point graph; Input the initial vector representation of each node into an embedding layer to obtain target feature embedding data, and for each node in the knowledge point graph, calculate the attention score of the node and all other nodes, the attention score reflects the degree of dependence of the current node on other nodes, the calculation of the attention score is realized by dot product, concatenation and neural network, and the softmax function is used to normalize the attention score to obtain the attention weight of each node to other nodes; According to the attention weight, the features of the neighboring nodes of each node are weighted aggregated to obtain an updated feature representation of the current node. After linear transformation and nonlinear activation of the aggregated updated feature representation, the target feature representation corresponding to each node is generated. For each reference breeding feedback text, the target feature representation of the corresponding node is integrated, and the integrated target feature representation is mapped to the basic breeding feedback semantic space through a fully connected layer.

8. The method for mining big data of aquaculture feedback based on artificial intelligence according to claim 1, It is characterized in that The method further comprises: Acquire a second breeding feedback label attribute set, and input the second breeding feedback label attribute set into a basic breeding feedback mining network; the second breeding feedback label attribute set is generated according to the collaborative breeding feedback text sequence of the second sample abnormal breeding trigger event, and the basic breeding feedback mining network includes the set breeding feedback semantic space sequence; Using the basic breeding feedback mining network, in the set breeding feedback semantic space sequence, the set breeding feedback semantic space corresponding to each breeding feedback label attribute before the second label node in the second breeding feedback label attribute set is retrieved, and the semantic space feature coordination is performed on each retrieved set breeding feedback semantic space to generate an estimated breeding feedback semantic space corresponding to the second label node, and according to the estimated breeding feedback semantic space corresponding to the second label node, a second estimated confidence corresponding to the breeding feedback label attribute on the second label node in the second breeding feedback label attribute set is generated; the second label node is determined from each breeding feedback label node in the second breeding feedback label attribute set, and the second estimated confidence is used to reflect the estimated derivative metric value between the breeding feedback text corresponding to the second label node and the forward breeding feedback text sequence corresponding to the second label node; Generate a second training error according to the second estimated confidences respectively corresponding to the breeding feedback label attributes on each second label node in the second breeding feedback label attribute set; updating the weight and bias information of the basic aquaculture feedback mining network according to the second training error until the second training termination requirement is met, thereby generating the target aquaculture feedback mining network; The basic aquaculture feedback mining network includes a retrieval unit, a coordination unit and a prediction unit, wherein the retrieval unit is used to retrieve and set the aquaculture feedback semantic space, the coordination unit is used for semantic space feature coordination, and the prediction unit is used to output a second estimated confidence; The step of updating the weight and bias information of the basic aquaculture feedback mining network according to the second training error until the second training termination requirement is met to generate the target aquaculture feedback mining network includes: The weights and bias information of the collaborative units and the prediction units in the basic aquaculture feedback mining network are updated according to the second training error until the second training termination requirement is met, thereby generating the target aquaculture feedback mining network.

9. The method for mining big data of aquaculture feedback based on artificial intelligence according to claim 1, It is characterized in that The step of performing semantic space feature coordination on each of the retrieved set aquaculture feedback semantic spaces to generate a target aquaculture feedback semantic coordination space of the global aquaculture feedback text sequence includes: According to the text arrangement order of the global farming feedback text sequence, the retrieved set farming feedback semantic spaces are aggregated to generate a set farming feedback semantic space sequence; Performing semantic space feature coordination on the set breeding feedback semantic space sequence to generate a breeding feedback semantic coordination space set; the breeding feedback semantic coordination space set includes breeding feedback semantic coordination spaces that match the number of texts in the global breeding feedback text sequence; From the breeding feedback semantic collaborative space set, the breeding feedback semantic collaborative space on the breeding feedback text label node of the target breeding feedback text is obtained as the target breeding feedback semantic collaborative space of the global breeding feedback text sequence.

10. A breeding feedback mining system, It is characterized in that The breeding feedback mining system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the breeding feedback big data mining method based on artificial intelligence as described in any one of claims 1 to 9.

Images