Knowledge model data management system based on artificial intelligence

Abstract

The invention discloses a knowledge model data management system based on artificial intelligence, and relates to the field of data management, and the system comprises the following components: a data collection and storage module, an intelligent prediction and analysis module, an abnormal comprehensive processing module, a knowledge graph and traceability module and a man-machine interaction management module. According to the invention, the intelligent prediction analysis module is combined with a time sequence prediction model and a causal reasoning algorithm, abnormal fluctuation and causal association thereof in data can be accurately judged, the accuracy of anomaly detection is effectively improved, and meanwhile, the knowledge graph and traceability module uses the constructed knowledge graph and reinforcement learning algorithm to improve the accuracy of anomaly detection. According to the method, the relation chain can be quickly traced from the abnormal data, the problem source can be positioned, the abnormal traceability efficiency is remarkably improved, and the functions act together, so that the system can more quickly and accurately discover and process problems when facing a complex data environment, and the stability and reliability of data management are guaranteed.

Description

TECHNICAL FIELD

[0001] The application relates to the field of data management, and particularly relates to a knowledge model data management system based on artificial intelligence. BACKGROUND

[0002] In the current digital era, with the continuous improvement of enterprise informatization, the amount of data presents an explosive growth, and the data sources are increasingly diversified, covering databases, sensors, log files and various forms, and the data contains rich value, which plays a crucial role in decision support and business optimization of enterprises.

[0003] The traditional data management method is often not up to the task when facing large-scale and high-complexity data. On the one hand, the traditional method mainly relies on manual rule setting or simple statistical methods in the data cleaning and preprocessing stage, and it is difficult to effectively remove noise and outliers in the data, which affects the accuracy of subsequent data analysis. On the other hand, in the aspect of anomaly detection and tracing, the traditional method usually relies on fixed thresholds or simple correlation analysis, and cannot accurately capture the time sequence dependence and causal relationship in the data, resulting in high false positive rate and false negative rate of anomaly detection and low tracing efficiency.

[0004] In summary, the traditional data management method has problems such as incomplete data cleaning, inaccurate anomaly detection, low tracing efficiency and inflexible model updating when dealing with large-scale and high-complexity data. These problems seriously restrict the efficiency and accuracy of data management and affect the full play of data value. Therefore, it is particularly important to develop a knowledge model data management system based on artificial intelligence. SUMMARY

[0005] The purpose of the application is to make up for the shortcomings of the prior art, and provide a knowledge model data management system based on artificial intelligence, which can realize automatic data cleaning, intelligent prediction analysis, anomaly comprehensive processing, knowledge graph and tracing and other functions by integrating advanced machine learning algorithms and knowledge graph technology, and has important significance for improving the accuracy and efficiency of data management.

[0006] To solve the above technical problems, the application provides the following technical scheme: a knowledge model data management system based on artificial intelligence, which comprises the following components: a data acquisition and storage module, an intelligent prediction analysis module, an anomaly comprehensive processing module, a knowledge graph and tracing module and a man-machine interaction management module.

[0007] The data collection storage module: collects data from various data sources such as databases, sensors, and log files, and performs preliminary cleaning and format conversion on the collected data, and transmits it to the storage unit in a unified format, uses distributed storage technology to support fast read and write of data, and stores classifiedly according to data timestamp and type, laying a solid foundation for subsequent data processing and analysis;

[0008] The intelligent prediction analysis module: builds a time series prediction model, learns and trains from historical data obtained from the data storage module, establishes a data time series change model to predict future data, compares the prediction results with actual data in real time, calculates the prediction error, and when the error exceeds the preset threshold, preliminarily judges that there is abnormal fluctuation, at the same time, uses Bayesian network and structural causal model causal reasoning algorithm to mine the causal relationship between data variables, combines with the time series prediction result to further accurately judge whether the abnormal fluctuation has causal association, improves the accuracy of abnormal detection, and outputs the abnormal information;

[0009] The abnormal comprehensive processing module: receives the abnormal information sent by the intelligent prediction analysis module, uses support vector machine and random forest machine learning classification algorithm to classify and identify abnormal data, determines the type and severity of the abnormality, displays the abnormal detection result in a visual form, and sends warning information to the management personnel, so that the management personnel can timely grasp the abnormal situation;

[0010] The knowledge graph and traceability module: according to the data in the data storage module, extracts entities, attributes and relationships through semantic analysis and entity recognition, and constructs a knowledge graph covering data sources, processing flow, related systems and personnel information, when the abnormal comprehensive processing module detects an abnormality, based on the knowledge graph, uses path search algorithm and graph neural network correlation analysis algorithm to trace the relationship chain from the abnormal data, quickly locates the problem source, and visualizes the traceability result to the management personnel to assist their decision-making and processing;

[0011] The human-computer interaction management module: provides a friendly operation interface for the management personnel, the management personnel can set system parameters through the interface, at the same time, the module displays abnormal detection results, traceability information and related data statistical analysis charts, which facilitates the management personnel to intuitively understand the data management situation, and according to the traceability result, timely takes processing measures.

[0012] Further, the data collection and storage module adopts a combination of rules and machine learning in the data cleaning process. First, the system sets a series of basic cleaning rules according to data types and business rules, and preliminarily filters the original data through these rules to eliminate data that obviously does not meet the requirements. Then, the isolated forest algorithm in machine learning is used to further detect outliers in the data after preliminary cleaning. The isolated forest algorithm constructs a binary tree structure to isolate data points in the tree, and determines whether a data point is an outlier according to its path length and other characteristics in the tree. When a data point is determined to be an outlier, the system uses a clustering-based repair algorithm to process it. This data cleaning method can effectively remove noise and abnormal data in the data, greatly improving the data quality and providing a reliable data basis for subsequent data processing and analysis.

[0013] Further, the time series prediction model constructed by the intelligent prediction and analysis module adopts an improved hybrid attention time series prediction algorithm that combines self-attention mechanism and time window attention mechanism, and the formula is:

[0014]

[0015] wherein is the predicted value at time t, σ is the activation function, w is the time window size, and α t,i is the attention weight between data at time t and time i, f θ (x i ) is the output of the input data x i through the feature extraction function with parameters θ, and the calculation method of the attention weight α t,i is:

[0016]

[0017] wherein score(a,b) is a similarity calculation function, score(a,b) = a T W s b, W s is a trainable weight matrix, and the time window size w is determined by analyzing the fluctuation period of historical data, specifically: the average interval between adjacent peaks of historical data is taken as the initial w value, and then the w value is adjusted within a range of 10% above and below the initial value through cross-validation, and the w value with the smallest prediction error is selected. The weight matrix W s is randomly initialized, and then trained and optimized through the back propagation algorithm with the mean square error between the predicted value and the actual value as the loss function, wherein n is the number of training data samples, and y tThe algorithm can more accurately capture the time sequence dependence and local characteristics of the data, and compared with the traditional time sequence prediction algorithm, the accuracy of the prediction and the adaptability to complex data patterns are improved.

[0018] Further, the causal reasoning of the intelligent prediction analysis module adopts a causal analysis algorithm based on a dynamic Bayesian network, which constructs a dynamic Bayesian network model, nodes represent data variables, edges represent causal relationships between variables, and the probability transfer formula is:

[0019]

[0020] where X t is a data variable set at time t, X t,i is the i-th variable in X t , Parents(X t,i ) is the parent node set of X t,i , the network structure is determined by a constraint-based structure learning method, the network structure is first initialized according to the correlation of the data and the domain knowledge, then the independence relationship between variables is judged by conditional independence test such as chi-square test, the network structure is gradually added or deleted, and the network structure is adjusted, the network parameters are calculated by maximum likelihood estimation method, for each node X t,i and its parent node set Parents(X t,i ), the maximum likelihood estimation of the parameter θ is:

[0021]

[0022] The algorithm can dynamically learn the causal relationship between data variables and adapt to data changes, and when judging the causal association of abnormal fluctuations, compared with the traditional Bayesian network method, it has higher accuracy and timeliness.

[0023] Further, the abnormal comprehensive processing module adopts an abnormal classification algorithm that fuses multiple features, which extracts the time sequence features, statistical features and context features of the data, constructs a feature vector F=[F ts ,F stat ,F ctx ], wherein F ts is a time sequence feature vector, which is composed of frequency features extracted by discrete Fourier transform and wavelet transform, F stat is a statistical feature vector, which contains mean, variance, skewness, kurtosis and other statistical quantities, and F ctx is a context feature vector, which is constructed according to the collection time and source system information of the data, and an improved support vector machine algorithm is used for classification, and the classification function is:

[0024]

[0025] where sgn is a sign function, a i is a Lagrange multiplier, y i is the class label of the sample F i , K(F,F i ) is a kernel function using an improved polynomial kernel function K(F,F i ) = (γF T′ F i +r) d , γ, r and d are trainable parameters, and b is a bias term, the Lagrange multiplier a i , the parameters γ, r, d and the bias term b are trained and optimized by a sequential minimal optimization algorithm, with the weighted sum of classification accuracy and misclassification loss as the objective function, where the weight of the classification accuracy is adjusted and determined on different data sets by cross-validation, and the misclassification loss uses a hinge loss function, the algorithm improves the accuracy and robustness of the classification of abnormal data by fusing multiple features and improving the kernel function, and can more effectively identify different types of abnormalities.

[0026] Further, when constructing the knowledge graph, the knowledge graph and traceability module adopts a knowledge graph construction algorithm based on semantic enhancement. In the entity recognition process, the algorithm combines a named entity recognition model and semantic similarity calculation. The named entity recognition model is fine-tuned using a bidirectional encoder representation model based on Transformer, and outputs an entity candidate set. For each entity candidate, the semantic similarity between the entity candidate and the words in the predefined entity type vocabulary is calculated, and the cosine similarity formula is used:

[0027]

[0028] where e is the vector representation of the entity candidate, which is obtained by pooling the word vectors output by the BERT model, v is the vector representation of the words in the vocabulary, which is obtained using a pre-trained word vector model, and when the similarity exceeds a threshold, the type of the entity candidate is determined. In relation extraction, a relation extraction model based on an attention mechanism is used to capture the semantic relationship between entity pairs in a sentence through a multi-head attention mechanism, output a relation probability matrix, and select the most probable relationship as the relationship between the entity pairs. In the knowledge graph construction process, a knowledge graph completion algorithm is introduced to learn the structure and semantic information of the knowledge graph through a graph convolution network to predict missing entities and relationships. The formula for completion is:

[0029]

[0030] where is the predicted relationship probability between entity i and entity j, σ is an activation function, W kis a trainable weight matrix, h i,k and h j,k are the feature vectors of entity i and entity j after k layers of GCN respectively, and l is the number of layers of GCN. The knowledge graph constructed by the algorithm has richer semantic information and higher integrity, providing a more reliable basis for anomaly tracing.

[0031] Further, when the knowledge graph and the tracing module perform anomaly tracing, a multi-path tracing algorithm based on reinforcement learning is adopted. The algorithm models the anomaly tracing process as a Markov decision process, where state S represents the current node position in the knowledge graph, action A represents the operation of selecting the next associated node from the current node, and reward R is used to evaluate the advantages and disadvantages of the tracing path.

[0032] The reward function is specifically designed as follows: when the algorithm reaches the source node of the problem, a reward value of 10 is given, if there is path repetition or the path has no practical significance, a reward of -1 is given, and in other cases, the reward value is the product of β and the relevance score. The weight coefficient β is adjusted through a large number of experiments on historical anomaly tracing cases to achieve an optimal balance between tracing accuracy and efficiency. The relevance score is calculated based on the relationship weight and semantic similarity between nodes. The relationship weight is initialized according to the importance and frequency of the relationship in the knowledge graph construction stage, and is dynamically adjusted in combination with user feedback and data statistics during system operation. The semantic similarity is calculated based on cosine similarity.

[0033] In terms of policy learning, the deep Q-network algorithm is adopted, and the Q-value function is expressed as Q(s,a; θ) = r + γmax a′ Q(s′,a′; θ), where θ is the network parameter, r is the reward obtained at the current step, γ is the discount factor, which is determined by cross-validation on different data sets to balance the importance of current reward and future reward, s′ is the next state reached after performing action a, and the network parameter θ is optimized by continuous training. The algorithm can quickly find the optimal anomaly tracing path in complex knowledge graph structure, significantly improving the efficiency and accuracy of anomaly tracing.

[0034] Further, the human-computer interaction management module provides a visual system monitoring interface, which adopts a layout optimization strategy based on an analytic hierarchy process. In the interface layout optimization process, a plurality of evaluation indexes need to be determined first, including whether the data display is complete, whether the user can obtain information conveniently, and whether the overall layout of the interface is beautiful. Then, a hierarchical structure model is constructed based on the indexes. Next, the relative importance between each index is determined through expert evaluation or user research, and a judgment matrix is constructed. According to the judgment matrix, the weight of each index is calculated. According to the calculated weight, the system will layout the data display area, the operation button area, and the state monitoring area interface elements, and place the information with a higher importance in a more eye-catching and more convenient position for the user to operate. In addition, the interface also supports user-defined layout. The user can freely adjust the position and size of the interface elements through drag and drop and zoom operations. The system will record the user's operation habits and preferences, and use machine learning algorithms to analyze these information, so as to automatically optimize the subsequent interface layout, thereby improving the convenience of user operation, enabling the user to more efficiently obtain system information, and significantly improving the user experience.

[0035] Further, the system further comprises a model updating module for updating and maintaining the time series prediction model of the intelligent prediction analysis module, the causal reasoning model of the causal analysis module, the abnormal classification model of the abnormal comprehensive processing module, and the related models of the knowledge graph and traceability module. The model updating module will periodically obtain newly generated historical data from the data acquisition and storage module. When the accumulation amount of new data reaches the pre-set update threshold, the model updating process is automatically triggered. When updating the time series prediction model, an incremental learning method is adopted, that is, based on the original model parameters, the model is fine-tuned according to the new data, so that the model can better adapt to the new changes and new trends of the data. For the causal reasoning model, an online learning method based on Bayes is used to update the structure and parameters of the dynamic Bayesian network according to the new data, so as to ensure that the model can accurately reflect the causal relationship between the data variables. When updating the abnormal classification model, a transfer learning algorithm is used, the model parameters trained on the original data set are used as the initial value, and then the new data set is fine-tuned to improve the model's ability to identify abnormal conditions in new data. For the related models of the knowledge graph and traceability module, the knowledge graph is updated and improved according to the entity and relationship information contained in the new data, and the same optimization algorithm as when constructing the model is used to ensure that the model performance is always in good condition. Through the work of the model updating module, each model in the system can adapt to the dynamic changes of the data in a timely manner, continuously maintain high accuracy and reliability, and ensure stable and efficient operation of the entire system.

[0036] Compared with the prior art, the knowledge model data management system based on artificial intelligence has the following beneficial effects:

[0037] I. The system can accurately judge the abnormal fluctuations and their causal relationships in the data through the intelligent prediction analysis module combined with the time series prediction model and the causal reasoning algorithm, effectively improving the accuracy of anomaly detection. At the same time, the knowledge graph and traceability module can quickly trace the relationship chain from abnormal data, locate the problem source, and significantly improve the efficiency of anomaly traceability by using the constructed knowledge graph and reinforcement learning algorithm. These functions work together to enable the system to quickly and accurately find and handle problems in complex data environments, ensuring the stability and reliability of data management.

[0038] II. The system uses a data cleaning method based on rules and machine learning through the data acquisition and storage module in the system, effectively removing noise and abnormal data in the data, greatly improving data quality. In addition, the system regularly updates each model automatically through the model update module to ensure that the model can adapt to dynamic changes in data in a timely manner, maintaining high accuracy and reliability. These measures optimize the data processing flow, reduce the need for manual intervention, and improve the automation level and overall efficiency of data management.

[0039] Other advantages, objects, and features of the present application will be set forth in part in the following specification, and in part will become apparent to those skilled in the art upon examination of the following specification, or can be learned by practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creating any creative labor.

[0041] Figure 1 It is a function implementation flowchart of a knowledge model data management system based on artificial intelligence;

[0042] Figure 2 It is a structure schematic diagram of a knowledge model data management system based on artificial intelligence. DETAILED DESCRIPTION

[0043] In order to further illustrate the technical means and effects adopted by the present application to achieve the predetermined invention purpose, the following will combine the drawings and preferred embodiments to specifically describe the specific implementation, structure, features and effects of the present application.

[0044] Embodiment one

[0045] In an automotive parts manufacturing enterprise, hundreds of precision machining equipment run continuously for 24 hours on the automated production line, sensors generate thousands of real-time data such as vibration, temperature, current every second, and the whole production line may stop due to equipment failure. The enterprise urgently needs an intelligent system to realize real-time monitoring of equipment running status, potential fault warning and rapid tracing to reduce maintenance cost and improve production efficiency.

[0046] The data acquisition and storage module connects with device sensors, PLC control system database and log files through industrial Ethernet, and real-time captures key parameters such as lathe spindle vibration frequency, motor temperature and servo system current. During the acquisition process, the system first performs preliminary cleaning according to the business rules preset by the equipment manufacturer, and eliminates obvious outliers. Then it further detects hidden abnormal data points through the Isolation Forest algorithm in machine learning to ensure data quality. The cleaned data is classified and archived in a distributed storage manner, such as stored in the cluster server in the hierarchy of "equipment number-parameter type-time stamp" for subsequent millisecond-level retrieval.

[0047] The intelligent prediction and analysis module calls the improved hybrid attention time series prediction algorithm, whose formula is:

[0048]

[0049] Where is the predicted value at time t, σ is the activation function, w is the time window size, and α t,i is the attention weight between time t and time i data, f θ (x i ) is the output of the input data x i through the feature extraction function with parameters θ. The model is trained based on the past 3 months of equipment running data to learn the time series law of parameter changes (such as the motor temperature usually rises 5-8℃ after continuous operation for 4 hours). The model predicts the parameter trend in the next 30 minutes in real time and compares it with the actual collected data. When the prediction error exceeds the preset threshold (such as vibration frequency deviation exceeds 15%), an abnormal warning is triggered. At the same time, the system uses the causal analysis algorithm based on dynamic Bayesian network, whose formula is:

[0050]

[0051] Where X t is the data variable set at time t, X t,i is the i-th variable in X t , and Parents(X t,i ) is the causal relationship between variables. Its probability transfer formula is:t,i the parent set of each node X t,i and its parent set Parents(X t,i ), the maximum likelihood estimation of the parameters θ is:

[0052]

[0053] The algorithm can dynamically learn the causal relationship between data variables, adapt to changes in data, analyze the causal relationship between parameters, and find that there is a strong causal relationship between "motor temperature rise" and "current increase", for example, which can determine that the abnormality may be caused by motor load abnormality.

[0054] After the abnormality comprehensive processing module receives the early warning information, the time series features, statistical features and context features of the data are extracted, and a feature vector is constructed by fusing the multi-feature abnormality classification algorithm. The formula is: The algorithm extracts the time series features, statistical features and context features of the data, constructs a feature vector F = [F ts , F stat , F ctx ], wherein F ts is the time series feature vector, F stat is the statistical feature vector, and F ctx is the context feature vector. The improved support vector machine algorithm is used for classification, and the classification function is:

[0055]

[0056] Wherein sgn is the sign function, α i is the Lagrange multiplier, y i is the class label of the sample F i , and K(F, F i ) is the kernel function. The improved polynomial kernel function K(F, F i ) = (γF T′ F i +r) d , γ, r and d are trainable parameters, and b is the bias term. The Lagrange multiplier α i , the parameters γ, r, d and the bias term b are trained and optimized by the sequential minimal optimization algorithm, with the weighted sum of classification accuracy and misclassification loss as the objective function. The misclassification loss uses the hinge loss function. With the help of the improved support vector machine algorithm, the system classifies the abnormality into specific types such as "motor overheating" and "bearing wear", and evaluates the severity. The abnormality result is displayed in the form of a dynamic heat map, the real-time curve of the abnormal parameter is labeled and compared with the historical threshold, and a short message warning containing the abnormal type and location is sent to the equipment management team.

[0057] The knowledge graph and traceability module integrates equipment table, maintenance records and sensor data, and constructs a knowledge graph containing "equipment components (such as main shaft, bearing) - sensor - operator - maintenance process". When a lathe vibrates abnormally, the system starts a multi-path traceability algorithm based on reinforcement learning, starting from the vibration sensor node and searching for possible fault sources along the connection relationship in the knowledge graph. The algorithm evaluates the effectiveness of the path through a reward function, for example, positioning to "insufficient bearing lubrication" obtains a high reward value, and finally presents the traceability result in a visual graph, labeling the fault chain (such as "lubricating oil pump blockage → bearing dry friction → vibration anomaly").

[0058] The human-computer interaction management module provides a three-dimensional visual monitoring interface, optimizes the layout using the analytic hierarchy process, and places key equipment status, abnormal warning list and other information in the center of the screen. The management personnel can customize the interface through drag-and-drop operations, such as placing the monitoring panel of high-frequency fault equipment on top. The system background records operation habits in real time, automatically optimizes subsequent layouts through machine learning, and the model update module periodically scans new data. When the accumulated equipment operation data reaches 20% of the historical data, it automatically triggers incremental learning, fine-tunes the parameters of the time series prediction model, and ensures that the model adapts to changes caused by equipment aging or process adjustment.

[0059] Example Two

[0060] A certain three-A hospital's imaging department processes more than 2000 CT and MRI examinations per day. Under the traditional manual film reading mode, doctors need to spend a lot of time comparing patients' historical images, and it is easy to miss subtle changes in the patient's condition. The hospital introduced an intelligent system to realize intelligent management of image data, assist doctors in quickly finding lesions and abnormalities, and trace factors that may affect the patient's condition, and improve the diagnosis efficiency of tumors, cardiovascular and cerebrovascular diseases, etc.

[0061] The data acquisition and storage module synchronizes CT and MRI image data in PACS, as well as related information such as patient medical records and blood test reports through the hospital information system interface in real time. During acquisition, the system automatically filters blurred images caused by patient movement. Through a format conversion tool, DICOM format images generated by different devices are unified into a standard format. In the storage stage, distributed clusters are used to store data by "patient ID-examination date-image type", such as archiving a patient's chest CT image and the same part of the image three months ago to the same folder, which is convenient for subsequent time series analysis.

[0062] The intelligent prediction analysis module uses an improved hybrid attention time series prediction algorithm for image features such as tumor volume and density of the patient, trains the model based on the patient's past 2 years of 4 image examination data, learns the lesion development law (such as the average increase of 2mm per month of a certain type of lung cancer lesion), and the model predicts the lesion parameters at the next examination in real time. When the error between the predicted value and the current image data exceeds the threshold value, it is prompted that the disease may be abnormal. At the same time, the system uses a causal analysis algorithm based on dynamic Bayesian networks to analyze the correlation between image features and clinical indicators. For example, when it is found that "increased density of lung nodules" and "increased CEA tumor marker" have a causal relationship, it is prompted that the tumor has a risk of deterioration.

[0063] The abnormality comprehensive processing module extracts time series features, statistical features and context features of image data, constructs a feature vector through a multi-feature abnormality classification algorithm, and uses an improved support vector machine algorithm. The system classifies the abnormality into types such as "tumor progression" and "inflammatory response", and evaluates the severity. The abnormality result is displayed on the image in the form of superimposed annotations, such as red boxes marking abnormally enlarged nodules, and a diagnostic auxiliary report containing a feature comparison table is generated and pushed to the clinician.

[0064] The knowledge graph and traceability module integrates patient medical history, medication records, image equipment parameters and other information, and constructs a knowledge graph containing "patient-lesion-treatment plan-image equipment". When a new lesion is found in the brain MRI of a patient, the system starts a multi-path traceability algorithm based on reinforcement learning, starting from the lesion node and searching for potential causes along the correlation in the knowledge graph (such as "patient medication history-drug side effects" and "image equipment parameters-imaging quality"). The algorithm preferentially selects high correlation paths according to the reward function, such as positioning to "recent chemotherapy drug use → blood-brain barrier permeability change → new lesion" to obtain positive rewards. Finally, the traceability results are displayed in the form of a time axis, and the possible influencing factor chain is labeled.

[0065] The human-computer interaction management module provides an image reading auxiliary interface, arranges the current image, historical comparison chart and abnormality warning information according to the doctor's operation habits using the analytic hierarchy process, supports related cases to be called through gesture zooming or voice instructions, and records the doctor's correction operations on the abnormality annotations in the background. The system optimizes the subsequent abnormality detection model through machine learning. The model update module automatically scans more than 500 new image data every quarter, uses transfer learning for the time series prediction model, fine-tunes the original model parameters combined with new data, and ensures that the model adapts to changes caused by new lesion characteristics or image equipment upgrades.

[0066] The above merely describes the preferred embodiments of the present application, and is not intended to limit the present application in any form. Although the present application has been disclosed with the preferred embodiments as above, it is not intended to limit the present application. Any person skilled in the art can make some changes or modifications to the above disclosed technical content to obtain equivalent embodiments with equivalent changes, as long as the changes or modifications do not deviate from the technical solution of the present application. Any modification, change, equivalent change and modification of the above embodiments made according to the technical essence of the present application still belong to the scope of the technical solution of the present application.

Claims

1. A knowledge model data management system based on artificial intelligence, characterized in that: The system includes the following components: data acquisition and storage module, intelligent prediction and analysis module, abnormality comprehensive processing module, knowledge graph and traceability module, and human-computer interaction management module; The data acquisition and storage module collects data from various data sources such as databases, sensors, and log files, performs preliminary cleaning and format conversion on the collected data, and transmits it to the storage unit in a unified format. It uses distributed storage technology to support fast reading and writing of data and classifies and stores data according to timestamps and types. The intelligent prediction and analysis module: constructs a time series prediction model, obtains historical data from the data storage module for learning and training, establishes a data time series change model to predict future data, compares the prediction results with the actual data in real time, calculates the prediction error, and preliminarily determines the existence of abnormal fluctuations when the error exceeds a preset threshold. At the same time, it uses Bayesian networks and structural causal model causal inference algorithms to explore the causal relationship between data variables, and combines the time series prediction results to further accurately determine whether the abnormal fluctuations have a causal relationship; The anomaly comprehensive processing module receives anomaly information sent by the intelligent prediction and analysis module, uses support vector machines and random forest machine learning classification algorithms to classify and identify anomaly data, determine the anomaly type and severity, display anomaly detection results in a visual form, and send warning information to management personnel; The knowledge graph and traceability module: Based on the data in the data storage module, through semantic analysis and entity recognition, it extracts entities, attributes and relationships, and constructs a knowledge graph covering data sources, processing procedures, related systems and personnel information. When the anomaly comprehensive processing module detects an anomaly, it uses the knowledge graph, path search algorithm and graph neural network association analysis algorithm to trace the relationship chain from the abnormal data, quickly locate the source of the problem, and visualize the traceability results to managers to assist them in decision-making; The human-computer interaction management module provides a user-friendly operation interface for management personnel, through which management personnel can set system parameters. At the same time, the module displays abnormality detection results, traceability information and related data statistical analysis charts, and takes timely processing measures based on the traceability results.

2. The knowledge model data management system based on artificial intelligence according to claim 1, characterized in that: During the data cleaning process, the data acquisition and storage module adopts a rule-based and machine learning approach. First, the system sets a series of basic cleaning rules based on the data type and business rules. These rules are used to preliminarily screen the original data and eliminate data that obviously does not meet the requirements. Then, the isolation forest algorithm in machine learning is used to further detect outliers in the data after preliminary cleaning. The isolation forest algorithm constructs a binary tree structure to isolate data points in the tree, and determines whether it is an outlier based on features such as the path length of the data point in the tree. When a data point is determined to be an outlier, the system uses a clustering-based repair algorithm to process it.

3. The knowledge model data management system based on artificial intelligence according to claim 1, characterized in that: The time series prediction model constructed by the intelligent prediction analysis module adopts an improved hybrid attention time series prediction algorithm, which combines the self-attention mechanism and the time window attention mechanism. The formula is: in is the predicted value at time t, σ is the activation function, w is the time window size, α t,i is the attention weight of the data at time t and time i, f θ (x i ) is the input data x i After the output of the feature extraction function with parameter θ, the attention weight α t,i The calculation method is: Where score(a,b) is the similarity calculation function, score(a,b)=a T W s b, W s is a trainable weight matrix. The time window size w is determined by analyzing the fluctuation cycle of historical data. The weight matrix W s After random initialization, the back propagation algorithm is used to calculate the mean square error between the predicted value and the actual value. Optimize the training loss function, where n is the number of training data samples, y t is the actual value at time t.

4. The knowledge model data management system based on artificial intelligence according to claim 3, characterized in that: The causal reasoning of the intelligent prediction and analysis module adopts a causal analysis algorithm based on a dynamic Bayesian network. The algorithm constructs a dynamic Bayesian network model, where nodes represent data variables and edges represent causal relationships between variables. The probability transfer formula is: where X t is the data variable set at time t, X t,i For X t The i-th variable in Parents(X t,i ) is X t,i The parent node set of each node X t,i and its parent node set Parents(X t,i ), given the sample data D, the maximum likelihood estimate of the parameter θ is: The algorithm can dynamically learn the causal relationship between data variables and adapt to data changes.

5. The knowledge model data management system based on artificial intelligence according to claim 1, characterized in that: The anomaly comprehensive processing module adopts an anomaly classification algorithm that integrates multiple features. The algorithm extracts the time series features, statistical features and context features of the data and constructs a feature vector F=[F ts ,F stat ,F ctx ], where F ts is the time series feature vector, F stat is the statistical eigenvector, F ctx is the context feature vector, and the improved support vector machine algorithm is used for classification. The classification function is: Where sgn is the sign function, α i is the Lagrange multiplier, y i For sample F i The category label, K(F,F i ) is the kernel function and the improved polynomial kernel function K(F,F i )=(γF T′ F i +r) d , γ, r and d are trainable parameters, b is the bias term, and the Lagrange multiplier α i , parameters γ, r, d and bias term b are trained and optimized through the sequential minimum optimization algorithm with the weighted sum of classification accuracy and misclassification loss as the objective function, and the misclassification loss adopts the hinge loss function.

6. The knowledge model data management system based on artificial intelligence according to claim 1, characterized in that: When constructing the knowledge graph, the knowledge graph and traceability module adopts a knowledge graph construction algorithm based on semantic enhancement. In the entity recognition process, the algorithm combines the named entity recognition model and semantic similarity calculation. The named entity recognition model is fine-tuned using a Transformer-based bidirectional encoder representation model to output a set of entity candidates. For each entity candidate, the semantic similarity between it and the vocabulary in the predefined entity type vocabulary is calculated using the cosine similarity formula: Where e is the vector representation of the entity candidate, and v is the vector representation of the word in the vocabulary. When the similarity exceeds the threshold, the type of the entity candidate is determined. In terms of relationship extraction, a relationship extraction model based on the attention mechanism is adopted. The semantic relationship between entity pairs in the sentence is captured through the multi-head attention mechanism, and the relationship probability matrix is ​​output. The relationship with the largest probability is selected as the relationship between the entity pairs. In the process of knowledge graph construction, the knowledge graph completion algorithm is introduced. The structure and semantic information of the knowledge graph are learned through the graph convolutional network to predict missing entities and relationships. The completion formula is: in is the predicted probability of the relationship between entity i and entity j, σ is the activation function, W k is the trainable weight matrix, h i,k and h j,k are the feature vectors of entity i and entity j after passing through k layers of GCN, and l is the number of GCN layers.

7. The knowledge model data management system based on artificial intelligence according to claim 6, characterized in that: When the knowledge graph and the traceability module perform anomaly tracing, a multi-path tracing algorithm based on reinforcement learning is used. This algorithm models the anomaly tracing process as a Markov decision process, where the state S represents the current node position of the traceability in the knowledge graph, the action A represents the operation of selecting the next associated node from the current node, and the reward R is used to evaluate the quality of the traceability path; The reward function is specifically designed as follows: when the algorithm reaches the node where the problem originates, a reward of 10 is given. If a path is repeated or has no practical significance, a reward of -1 is given. In other cases, the reward value is the product of β and the relevance score. The weight coefficient β is used here, and the relevance score is calculated based on the relationship weight and semantic similarity between nodes. The relationship weight is initialized according to the importance and frequency of the relationship during the knowledge graph construction phase, and is dynamically adjusted during system operation based on user feedback and data statistics. Semantic similarity is calculated based on cosine similarity. In terms of strategy learning, a deep Q-network algorithm is used, and the Q value function is expressed as Q(s,a;θ)=r+γmax a′ Q(s′,a′;θ), where θ is the network parameter, r is the reward obtained in the current step, γ is the discount factor, and s′ is the next state reached after performing action a. The network parameter θ is optimized through continuous training.

8. The knowledge model data management system based on artificial intelligence according to claim 1, characterized in that: The visual system monitoring interface provided by the human-computer interaction management module adopts a layout optimization strategy based on the hierarchical analysis method. During the interface layout optimization process, it is first necessary to determine multiple evaluation indicators, including whether the data display is complete, whether the user can obtain information conveniently, and whether the overall layout of the interface is beautiful. Then, a hierarchical structure model is constructed based on the indicators. Next, the relative importance of each indicator is determined through expert evaluation or user survey, and a judgment matrix is ​​constructed. According to the judgment matrix, the weight of each indicator is calculated. Based on the calculated weight, the system will layout the interface elements of the data display area, operation button area, and status monitoring area, and place the more important information in the interface. In addition, the interface also supports user-defined layout. Users can freely adjust the position and size of interface elements by dragging and zooming. The system will record the user's operating habits and preferences, and use machine learning algorithms to learn and analyze this information, thereby automatically optimizing the subsequent interface layout.

9. The knowledge model data management system based on artificial intelligence according to claim 1, characterized in that: The system also includes a model update module, which is used to update and maintain the time series prediction model of the intelligent prediction and analysis module, the causal reasoning model of the causal analysis module, the anomaly classification model of the anomaly comprehensive processing module, and the related models of the knowledge graph and traceability module. The model update module regularly obtains newly generated historical data from the data acquisition and storage module. When the accumulated amount of new data reaches a preset update threshold, the model update process is automatically triggered. When updating the time series prediction model, an incremental learning method is adopted, that is, based on the original model parameters, the model is fine-tuned according to the new data. For the causal reasoning model, a Bayesian online learning method is used to update the structure and parameters of the dynamic Bayesian network according to the new data. When updating the anomaly classification model, a transfer learning algorithm is used to use the model parameters trained on the original dataset as the initial values, and then targeted fine-tuning is performed in combination with the new dataset. For the related models of the knowledge graph and traceability module, the knowledge graph is updated and improved based on the entity and relationship information contained in the new data, and the same optimization algorithm as when the model was built is used to ensure that the model performance is always in good condition.

Images