A deep intelligent analysis method, device, equipment, and storage medium based on AI performance test data from a hospital-wide integrated clinical rehabilitation system.

Abstract

This invention discloses a deep intelligent analysis method, device, equipment, and storage medium for AI performance test data based on a hospital-wide integrated clinical rehabilitation system. The method includes: synchronously collecting indicator data, log data, and auxiliary correlation data; cleaning, standardizing, and fusing these data to generate a unified AI analysis dataset; using optimized isolated forest and LogBERT for threshold-free anomaly detection, and combining clinical workload to distinguish between true and false anomalies; after identifying the true bottleneck, locating the root cause based on bidirectional correlation analysis using a knowledge graph and generating a traceable evidence chain; using scenario-based LSTM for short-term peak warning and long-term degradation prediction; and feeding back the analysis results to the knowledge graph and iteratively optimizing the algorithm parameters to form a closed loop. This invention solves the problems of low efficiency, high false alarm rate, difficulty in root cause location, and inability to predict latent degradation in manual analysis, significantly improving the efficiency and accuracy of performance test data analysis.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, and in particular to a deep intelligent analysis method, device, equipment, and storage medium based on AI performance test data from a hospital-wide integrated clinical rehabilitation system. Background Technology The hospital-wide integrated clinical rehabilitation system integrates core functions across the entire process, including rehabilitation assessment, treatment plan development, rehabilitation training execution, efficacy tracking, and data archiving. Its operational stability, response speed, and data processing efficiency directly impact the continuity and accuracy of clinical rehabilitation diagnosis and treatment. Therefore, rigorous performance testing is necessary to verify the system's reliability. During performance testing, the system simultaneously generates massive amounts of test data, including millions of multi-dimensional performance metrics (such as CPU utilization, memory usage, response time, error rate, and IOPS) and gigabytes of operational log data (such as service call logs, database operation logs, and exception logs), resulting in a system characterized by "large data volume, multiple dimensions, complex relationships, and high noise levels."

[0002] Currently, the performance test data analysis of the hospital-wide integrated clinical rehabilitation system mostly adopts a combination of manual analysis and traditional monitoring tools (such as Zabbix and Nagios). This approach has many insurmountable shortcomings and is no longer suitable for the analysis of massive amounts of test data. 1. Extremely low analysis efficiency, trapped in the "log storm" dilemma: Millions of indicator data and GB of logs need to be manually screened and compared layer by layer. The analysis time often takes several hours or even days, making it impossible to quickly output analysis results, delaying the timing of system optimization, and affecting the efficiency of the clinical rehabilitation system's launch and maintenance. 2. Poor accuracy of anomaly detection, with frequent false alarms and missed alarms: Traditional analysis relies on manually setting fixed thresholds (such as "alarm when response time > 2s"), which cannot adapt to the dynamic load scenarios of clinical rehabilitation systems (such as daily rehabilitation assessment peaks and monthly efficacy statistics batch tasks). It is easy to misjudge "false anomalies" such as instantaneous network jitter and temporary data peaks as performance bottlenecks. At the same time, it may miss true anomalies such as "sudden changes in indicators but not exceeding the threshold". The false alarm rate of traditional threshold alarms is as high as 38%, which seriously interferes with manual analysis. 3. Difficulty in locating the root cause of performance bottlenecks and lack of a complete chain of evidence: Traditional analysis is mostly a one-way investigation of "logs → indicators", which cannot connect multi-source data such as system topology, historical failures, and code call chains. When performance bottlenecks occur (such as delayed query of rehabilitation data), it is difficult for humans to locate the root cause layer by layer (such as database index failure, code query logic error) and cannot generate a traceable chain of evidence, resulting in a lack of clear direction for repair work. 4. Lack of long-term trend prediction capability and the limitation of "one-time test": Traditional analysis only reviews the data from a single performance test and cannot predict long-term performance changes of the system based on historical data. It is difficult to detect "hidden performance degradation" problems such as memory leaks, which may lead to failures such as crashes and slow response in the actual operation of the clinical rehabilitation system (such as after long-term continuous operation), affecting the normal operation of diagnosis and treatment.

[0003] While some AI-assisted performance analysis solutions have been disclosed in the existing technology, they are mainly applied to fields such as e-commerce and banking. They do not take into account the special characteristics of the integrated clinical rehabilitation system of the whole hospital. The clinical rehabilitation system has special load fluctuation patterns (such as the peak of rehabilitation assessment is concentrated between 9-11 am), high data sensitivity, wide range of fault impact (directly related to patient rehabilitation diagnosis and treatment), and special core indicators (such as image transmission response time and rehabilitation data query accuracy). Existing solutions cannot adapt to these special needs and have not achieved a closed-loop analysis of the whole process of "anomaly detection-root cause location-trend prediction", which makes it difficult to meet the requirements of accuracy, efficiency and security of performance analysis of the clinical rehabilitation system. Summary of the Invention

[0004] This invention aims to overcome the shortcomings of existing technologies by providing a deep intelligent analysis method, device, equipment, and storage medium based on AI performance test data from a hospital-wide integrated clinical rehabilitation system. Through AI algorithms and multi-source data fusion technology, it achieves a closed-loop analysis of the entire process, including real-time detection and pattern recognition of performance anomalies, root cause localization of performance bottlenecks and generation of evidence chains, and prediction of long-term performance trends. This reduces analysis time from hours to minutes, improving accuracy and efficiency, adapting to the specific scenarios of integrated clinical rehabilitation systems, and ensuring stable system operation. It can at least overcome the shortcomings of existing technologies, such as low efficiency in performance test data analysis, high false alarm rate in anomaly detection, difficulty in root cause localization, and lack of long-term trend prediction capabilities.

[0005] According to a first aspect of the present invention, a deep intelligent analysis method based on AI performance test data from a hospital-wide integrated clinical rehabilitation system is provided, comprising the following steps: S1. Multi-source performance test data acquisition and preprocessing: Simultaneously collect multi-source data generated from the performance test of the hospital's integrated clinical rehabilitation system. The multi-source data includes indicator data, log data, and auxiliary correlation data. Clean, standardize, and fuse the multi-source data to establish the correlation mapping between indicator data, log data, and auxiliary correlation data, and generate a unified AI analysis dataset. S2. Real-time detection and pattern recognition of performance anomalies: Based on AI analysis datasets, an optimized isolated forest algorithm is used to detect anomalies in indicator data and identify abrupt change patterns in indicators; the LogBERT model is used to perform cluster analysis on log data to identify similar error patterns; and clinical workload characteristic data is combined to distinguish between true and false anomalies in the detection results, identify true bottlenecks, and trigger alarms. S3. Root cause location and evidence chain generation of performance bottlenecks: When an alarm is triggered, the root cause of the performance bottleneck is located through bidirectional correlation analysis based on a pre-built performance analysis knowledge graph, and a traceable evidence chain containing bottleneck phenomena, related indicators, related logs, root cause and remediation suggestions is automatically generated. S4. Performance Trend Prediction: Based on AI analysis datasets and historical performance data, a scenario-based LSTM model is used to predict short-term and long-term performance trends. When a potential bottleneck is predicted, an early warning is triggered and optimization suggestions are output. S5. Analysis Result Feedback and Model Optimization: Collect feedback data on the analysis results, supplement the analysis data into the knowledge graph and historical case library, and iteratively optimize the model parameters of each algorithm to form a closed loop of the entire process.

[0006] Therefore, by synchronously collecting multi-source data and establishing association mappings, and combining optimized isolated forest, LogBERT, bidirectional association analysis of knowledge graphs and scenario-based LSTM models for collaborative processing, a closed-loop intelligent analysis of the entire process from anomaly detection and root cause localization to trend prediction has been achieved. Compared with manual analysis, the analysis time has been reduced from several hours to minutes. Anomaly detection does not require manual threshold setting, root cause localization can automatically generate a complete chain of evidence, and long-term prediction can detect latent degradation risks in advance, significantly improving the analysis efficiency, accuracy and intelligence level of the performance test data of the hospital-wide integrated clinical rehabilitation system.

[0007] In some implementations, in step S1: The metrics data should include at least several core performance metrics across multiple dimensions, such as CPU utilization, memory usage, rehabilitation assessment report generation response time, image transmission response time, error rate, and IOPS. Log data includes the operation logs of all system modules, with clinically relevant logs retained after preprocessing; The auxiliary correlation data includes system topology, historical failure cases, code call chains, and clinical workload characteristics data.

[0008] Therefore, by specifying that the indicator data should include at least core dimensions such as CPU, memory, and response time, and that the log data retains clinically relevant logs, and that the auxiliary data includes topology, cases, call chains, and load characteristics, the comprehensiveness and relevance of multi-source data are ensured. This provides high-quality and highly relevant input for subsequent AI analysis, avoids interference from irrelevant data, and improves the sensitivity of anomaly detection and the accuracy of root cause localization.

[0009] In some implementations, the association mapping is established in step S1, specifically by establishing association mappings for four types of data: "indicators, logs, system topology, and historical faults," generating a unified AI analysis dataset, and storing the preprocessed data in a distributed database.

[0010] Therefore, by establishing a correlation mapping for four types of data and storing them in a distributed database, the organic integration of heterogeneous data is achieved, enabling mutual verification and collaborative analysis of abnormal indicators, log errors, system topology relationships, and historical failure cases. This correlation mapping provides a data foundation for subsequent bidirectional correlation root cause localization, significantly improving the efficiency and accuracy of root cause analysis. At the same time, distributed storage ensures data security and efficient querying.

[0011] In some implementations, the distinction between true and false anomalies in step S2 specifically includes: Cross-validate the test results with clinical workload characteristics data and system status data; For transient abnormalities whose duration is within a preset threshold and do not affect core clinical functions, they are judged as false abnormalities and are only recorded without alarm. If the duration exceeds the preset threshold and is associated with error logs, it is determined to be a true bottleneck and an alarm is triggered.

[0012] Therefore, by cross-validating the detection results with clinical load characteristics and system status, and distinguishing between transient false anomalies and persistent true bottlenecks based on duration thresholds, the problem of high false alarm rate of traditional fixed threshold alarms is effectively solved. This method can filter out false alarms caused by load fluctuations, network jitter, etc., and only trigger alarms for persistent bottlenecks that truly affect the core clinical functions, which greatly reduces the burden of handling invalid alarms for operation and maintenance personnel.

[0013] In some implementations, in step S3: The performance analysis knowledge graph integrates system topology, historical failure cases, code call chains, and system configuration data. The nodes in the graph include bottleneck phenomena, abnormal indicators, log events, and system modules, and the edges represent the relationships. It also has an automatic update function. Two-way correlation analysis includes: forward correlation of bottleneck phenomena with abnormal indicators and log events to identify suspicious system modules, and reverse correlation of historical failure cases and code call chains to trace the root cause, and verification through time series data matching; The confidence level of the root cause is not lower than the preset confidence threshold.

[0014] Therefore, by constructing a knowledge graph that includes system topology, historical faults, and code call chains, and by adopting a two-way analysis mechanism that uses forward association to lock suspicious modules and reverse tracing to root causes, the root causes can be located quickly and accurately. Combined with time-series data matching verification and confidence assessment (not lower than the preset confidence threshold), the output root causes have high reliability. At the same time, the automatic updating function of the knowledge graph enables the analysis capabilities to continuously improve with the accumulation of data.

[0015] In some implementations, in step S4: The scenario-based LSTM model is trained based on historical performance data and clinical workload patterns within a preset historical data period. The time window length, number of hidden layer nodes, and learning rate parameters are optimized, and the model prediction error is controlled within a preset error threshold. The time frame for short-term forecasts is a preset short-term time window, used to predict performance indicators during peak periods and guide resource expansion. The long-term prediction timeframe is a preset long-term time window, used to identify hidden performance degradation trends such as memory leaks and connection pool exhaustion.

[0016] Therefore, by using a scenario-based LSTM model trained based on historical data and clinical workload patterns, we can achieve short-term (preset short-term time window) peak-hour performance warnings and long-term (preset long-term time window) latent degradation predictions. Short-term predictions can guide maintenance personnel to expand resources in advance to avoid performance degradation during peak hours, while long-term predictions can detect latent risks such as memory leaks and connection pool exhaustion weeks in advance, preventing sudden system crashes and ensuring the continuity and stability of clinical rehabilitation services.

[0017] In some implementations, step S5 iteratively optimizes the model parameters of each algorithm, specifically including: adding newly discovered abnormal patterns, new root cause cases, and prediction error feedback data to the knowledge graph and historical case library during the analysis, and adjusting the detection sensitivity of the Isolation Forest algorithm, the clustering threshold of the LogBERT model, and the training parameters of the LSTM model accordingly to achieve continuous optimization of the model.

[0018] Therefore, by feeding newly discovered anomaly patterns, root cause cases, and prediction errors into the knowledge graph and case library, and adjusting the parameters of algorithms such as Isolation Forest, LogBERT, and LSTM accordingly, the analysis model is continuously optimized and adaptively evolved. With the accumulation of analysis data, the system's anomaly detection sensitivity, root cause localization accuracy, and trend prediction precision are continuously improved, enabling it to better adapt to system upgrades and business changes, forming a virtuous cycle of becoming smarter with use.

[0019] According to a second aspect of the present invention, a deep intelligent analysis device based on AI performance test data from a hospital-wide integrated clinical rehabilitation system is also provided, comprising at least: The multi-source data acquisition module is used to synchronously collect indicator data, log data, and auxiliary correlation data generated by the performance test of the hospital's integrated clinical rehabilitation system, supporting real-time and offline acquisition. The data preprocessing module is used to clean, standardize, and fuse multi-source data, establish correlation mappings between various types of data, and generate a unified AI analysis dataset. The core AI analysis module integrates an anomaly detection unit, a root cause localization unit, and a trend prediction unit. The anomaly detection unit employs an optimized isolated forest algorithm and a LogBERT model to achieve real-time detection and pattern recognition of performance anomalies, and combines clinical workload feature data to distinguish between true and false anomalies. The root cause localization unit uses a performance analysis knowledge graph for bidirectional correlation analysis to achieve root cause localization of performance bottlenecks and generation of evidence chains. The trend prediction unit uses a scenario-based LSTM model to predict short-term and long-term performance trends. The results output and visualization module is used to output analysis results in the form of charts or reports, and supports querying and exporting. The knowledge graph and database module is used to store the performance analysis knowledge graph, as well as various raw data, analysis results, and feedback data, and supports automatic updates. The feedback optimization module is used to collect user feedback, supplement the analysis data into the knowledge graph and case library, and iteratively optimize the parameters of each algorithm in the core AI analysis module to form a closed loop throughout the entire process.

[0020] Therefore, through the collaborative work of the multi-source data acquisition module, data preprocessing module, AI analysis core module (including three units: anomaly detection, root cause localization, and trend prediction), result output and visualization module, knowledge graph and database module, and feedback optimization module, a complete automated analysis device from data acquisition, processing, analysis to feedback optimization has been constructed. This device can automatically perform in-depth analysis of performance test data without human intervention, output intuitive analysis results and traceable evidence chains, and has closed-loop self-optimization capabilities, significantly reducing the technical threshold and workload of operation and maintenance personnel.

[0021] In some implementations, the anomaly detection unit uses an optimized isolated forest algorithm to perform unsupervised modeling of time-series index data across multiple dimensions and automatically identify index mutation patterns; the LogBERT model is used to perform cluster analysis on log data, extract semantic features, and identify similar error patterns; the anomaly detection unit also combines clinical workload feature data to distinguish between true and false anomalies in the detection results.

[0022] Therefore, by optimizing the isolated forest in the anomaly detection unit for unsupervised modeling and mutation identification of multi-dimensional indicator data, using LogBERT for log semantic clustering and pattern recognition, and combining it with clinical load characteristics to distinguish between true and false anomalies, intelligent anomaly detection without manual annotation and threshold setting is achieved. This solution effectively adapts to the special characteristics of load fluctuations in clinical rehabilitation systems, and can both capture indicator mutations and filter false anomalies caused by load changes, resulting in more accurate and reliable detection results.

[0023] In some implementations, the root cause localization unit integrates the performance analysis knowledge graph with system topology, historical failure cases, code call chains and system configuration data. Through bidirectional analysis, it outputs the root cause by forward correlation of bottleneck phenomena with indicators and logs, and backward combination of historical cases and code call chains. It also automatically generates a traceable evidence chain containing bottleneck phenomena, related indicators, related logs, root cause and remediation suggestions.

[0024] Therefore, through the forward (bottleneck → metric / log → module) and reverse (historical case / call chain → root cause) bidirectional analysis of the knowledge graph in the root cause localization unit, the root cause is automatically output and a traceable evidence chain containing bottleneck phenomena, related metrics, related logs, root cause and remediation suggestions is generated. This achieves the automation of root cause localization and the completeness of the evidence chain. Operation and maintenance personnel can directly obtain the root cause and remediation plan without checking layer by layer, which greatly shortens the fault repair time and significantly improves the efficiency of batch task processing.

[0025] According to a third aspect of the present invention, an electronic device is also provided, including one or more processors and a memory, the memory storing one or more programs, which, when executed by the processor, implement the above-described deep intelligent analysis method based on AI performance test data of a hospital-wide integrated clinical rehabilitation system.

[0026] According to a fourth aspect of the present invention, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the above-described deep intelligent analysis method based on AI performance test data of a hospital-wide integrated clinical rehabilitation system.

[0027] The beneficial effects of this invention are: 1. Significantly improved analysis efficiency: This invention reduces the time for in-depth analysis of GB-level logs and millions of indicator data from several hours of traditional manual analysis to less than 5 minutes through parallel acquisition of multi-source data and AI collaborative analysis, improving analysis efficiency by more than 30 times and effectively supporting the rapid optimization and operation and maintenance decisions of clinical rehabilitation systems.

[0028] 2. Accurate and reliable anomaly detection: The technology adopts a threshold-free anomaly detection technique that integrates an optimized isolated forest and LogBERT model. It eliminates the need for manually setting fixed thresholds and distinguishes between true and false anomalies by combining clinical workload characteristics, which significantly reduces the false alarm rate and the missed alarm rate and avoids interference from invalid alarms with operation and maintenance work.

[0029] 3. Rapid and Traceable Root Cause Analysis: Based on bidirectional association analysis technology using knowledge graphs, it can automatically associate multi-source data such as system topology, historical failure cases, and code call chains to quickly locate the root cause of performance bottlenecks and generate a complete and traceable evidence chain that includes bottleneck phenomena, related indicators, related logs, root causes, and repair suggestions. The root cause location accuracy is high, and the fault repair time is significantly shortened.

[0030] 4. Early warning of hidden performance degradation: By using a scenario-based LSTM model to learn from historical performance data and clinical workload patterns, performance warnings during short-term peak periods and predictions of long-term hidden performance degradation (such as memory leaks and connection pool exhaustion) can be achieved. Potential risks can be detected weeks in advance to avoid sudden system crashes and ensure the continuity of clinical diagnosis and treatment services.

[0031] 5. Continuous optimization through a closed-loop process: By analyzing results and updating models through iteration, newly discovered anomaly patterns, root cause cases, and prediction errors are added to the knowledge graph and case library in each analysis. The parameters of each AI algorithm are continuously optimized so that the capabilities of the analysis system can be continuously improved with the accumulation of data, adapting to system upgrades and business changes.

[0032] 6. Strong adaptability to medical scenarios: This invention is designed with the special load patterns (such as peak rehabilitation assessment and dense image transmission), core indicators (such as image transmission response time and rehabilitation data query accuracy) and data compliance requirements of the integrated clinical rehabilitation system for the whole hospital in a customized manner. It can be directly deployed in medical institutions without affecting normal diagnosis and treatment business and has good engineering implementation capabilities.

[0033] 7. Significantly reduce operation and maintenance costs: The automated and intelligent deep analysis process reduces reliance on manual analysis experience, lowers manpower input and fault repair costs, and avoids diagnostic interruptions and economic losses caused by system crashes through early warning, resulting in significant engineering and social benefits.

[0034] In summary, this invention discloses a deep intelligent analysis method, device, equipment, and storage medium for AI performance test data based on a hospital-wide integrated clinical rehabilitation system. The method includes: synchronously collecting indicator data, log data, and auxiliary correlation data; cleaning, standardizing, and fusing these data to generate a unified AI analysis dataset; employing optimized isolated forest and LogBERT for threshold-free anomaly detection, and combining clinical workload to distinguish between true and false anomalies; identifying true bottlenecks; locating root causes and generating traceable evidence chains based on bidirectional correlation analysis using a knowledge graph; employing scenario-based LSTM for short-term peak warnings and long-term degradation predictions; and feeding back the analysis results to the knowledge graph and iteratively optimizing algorithm parameters to form a closed loop. This invention solves the problems of low efficiency, high false alarm rate, difficulty in root cause location, and inability to predict latent degradation in manual analysis, significantly improving the efficiency and accuracy of performance test data analysis. Attached Figure Description

[0035] Figure 1 This is an overall flowchart of a deep intelligent analysis method based on AI performance test data from a hospital-wide integrated clinical rehabilitation system, according to one embodiment of the present invention.

[0036] Figure 2 This is a flowchart of multi-source data acquisition and preprocessing according to one embodiment of the present invention.

[0037] Figure 3 This is a flowchart illustrating performance anomaly detection and true / false differentiation according to one embodiment of the present invention.

[0038] Figure 4 This is a schematic diagram illustrating root cause localization and evidence chain generation in one embodiment of the present invention.

[0039] Figure 5 This is a schematic diagram illustrating long-term performance trend prediction and closed-loop optimization in one embodiment of the present invention.

[0040] Figure 6This is a system framework diagram of a deep intelligent analysis device based on AI performance test data from a hospital-wide integrated clinical rehabilitation system, according to one embodiment of the present invention.

[0041] Figure 7 This is a schematic diagram of an electronic device according to one embodiment of the present invention. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0043] Reference Figure 1-5 As shown, this invention provides a deep intelligent analysis method based on AI performance test data from a hospital-wide integrated clinical rehabilitation system. This method can be executed by an electronic device, which can be implemented using software and / or hardware. The method includes the following steps: Step S1: Multi-source performance test data acquisition and preprocessing: 1.1. Data Acquisition: Three types of core data are collected synchronously: indicator data, log data, and auxiliary correlation data. The multi-source data acquisition module adopts a distributed acquisition architecture, deploying lightweight acquisition agents on the hospital's core servers, databases, and application service nodes to push indicator data in real time via TCP / IP protocol; log data is centrally collected using the ELK (Elasticsearch, Logstash, Kibana) technology stack or the Fluentd tool; auxiliary correlation data is imported through the configuration management database (CMDB) and code version management system. The acquisition frequency supports dynamic configuration, with a default of 1 time / second, which can be automatically adjusted according to clinical workload (e.g., increased to 2 times / second during peak periods).

[0044] 1.2. Customized Cleaning, Standardization, and Integration: Outliers were removed from the collected indicator data, and missing values ​​were filled using time-series interpolation. Log data was structured and formatted. Indicator data was normalized to the [0,1] interval. Specific data cleaning rules included: outliers in the indicator data (such as values ​​exceeding the mean ± 3 standard deviations) were removed using the 3σ principle, and then missing values ​​were filled using linear interpolation; data segments with more than 5 consecutive missing time points were marked as unreliable intervals and excluded from subsequent analysis. For log data, regular expressions were used to extract structured fields such as timestamps, log levels, module names, error codes, and stack traces, filtering out DEBUG level logs and pure heartbeat logs. Data standardization used the Min-Max normalization method, mapping each indicator value to the [0,1] interval, with the formula: x' = (x - min) / (max - min), where min and max are the minimum and maximum values ​​of the indicator in historical data. Based on this, a correlation mapping of "metrics-logs-system topology-historical faults" is established to generate a unified AI analysis dataset, which is stored in a compliant distributed database. The specific data fusion method is as follows: using timestamps as the core index, align metric data with log data (time window error ≤ 1 second); use service call chain IDs (Trace IDs) in the system topology to associate metrics, logs, and code call chains; through keyword matching and semantic similarity calculation, associate symptom descriptions in historical fault cases with currently detected abnormal patterns; finally, a unified AI analysis dataset is generated, with each record containing: timestamp, metric name and value, log summary, module, topology node, and associated fault case ID.

[0045] Step S2: Real-time detection and pattern recognition of performance anomalies: 2.1. Optimize the Isolation Forest algorithm for anomaly detection: The optimized Isolation Forest algorithm is used to perform unsupervised modeling of time-series indicator data with more than 20 dimensions, automatically identify the mutation patterns of the indicators, and optimize the detection sensitivity for core clinical indicators (image transmission response time, rehabilitation data query accuracy).

[0046] Specific measures to optimize the isolated forest algorithm include: increasing the number of trees to 200 (default 100) to improve detection stability; setting a lower subsampling size (default 256, core indicators set to 128) for core clinical indicators (such as image transmission response time) to improve sensitivity to local mutations; and introducing an adaptive contamination coefficient to dynamically estimate the proportion of abnormalities based on historical normal data to avoid misjudgments caused by changes in data distribution.

[0047] 2.2. LogBERT Model for Log Anomaly Identification: The LogBERT model is used to perform cluster analysis on GB-level log data, extract semantic features of the logs, identify similar error patterns, and automatically filter invalid logs; The specific implementation process of the LogBERT model: First, use a log parsing tool (such as Drain) to convert the raw log into a template sequence; Then, a pre-trained BERT model (fine-tuned on the medical log corpus) is used to encode each log template into a 768-dimensional semantic vector; Semantic vectors are grouped by clustering algorithms (such as DBSCAN), and log patterns that are the same or similar are grouped into one category to identify high-frequency error patterns. Set the cluster radius ε=0.5 and the minimum number of samples minPts=5 to mark isolated vectors (noise points) as potential anomalies.

[0048] As an alternative, models such as LogAnomaly and DeepLog can also be used, simply by replacing BERT with the corresponding encoder. These are all equivalent implementations of the present invention.

[0049] 2.3. Distinguishing between true and false anomalies: The above detection results are cross-validated with clinical load characteristic data and system status data: For transient anomalies lasting less than 1-2 seconds and not affecting core clinical functions, they are judged as false anomalies, and only recorded without alarm; for anomalies lasting more than the preset threshold (default 30 seconds, which can be customized and adjusted according to clinical needs) and associated with error logs, they are judged as true bottlenecks and alarms are triggered.

[0050] The default threshold is 30 seconds, which can be configured according to the business importance of the clinical rehabilitation system. For example, for the image transmission module, it can be set to 10 seconds to ensure timely response; for the non-real-time statistics module, it can be set to 60 seconds to reduce false alarms.

[0051] The system also supports adaptive threshold learning based on historical data, dynamically adjusting the threshold by analyzing the distribution of abnormal durations over the past 30 days.

[0052] Step S3: Root cause identification and evidence chain generation of performance bottlenecks: 3.1. Construction of a dedicated knowledge graph for medical scenarios: Integrate system topology, historical failure cases, code call chains, and system configuration data to construct a dedicated knowledge graph for performance analysis. The nodes in the graph include bottleneck phenomena, abnormal indicators, log events, system modules, etc., and the edges represent the relationships. This knowledge graph has an automatic update function.

[0053] The construction of a knowledge graph involves the following steps: Entity extraction—extracting nodes such as services, databases, and middleware from the system topology; extracting bottleneck phenomena and root cause types from historical failure cases; and extracting methods and interfaces from the code call chain.

[0054] Relationship extraction—Based on system call logs and configuration relationships, establish edges such as "call", "dependency", and "deployed in"; based on historical fault data, establish edges such as "caused", "related to", and "similar to".

[0055] Knowledge integration – using graph databases (such as Neo4j) for storage, supporting graph traversal analysis using the Cypher query language.

[0056] Automatic updates – When a new failure case is confirmed, nodes and edges are automatically added through the rules engine, or the confidence weights of existing entities are updated.

[0057] 3.2. Bidirectional Correlation Root Cause Analysis: Once the true bottleneck is identified, perform bidirectional correlation analysis: Starting with bottleneck phenomena, correlate abnormal indicators and log events to identify suspicious system modules; By combining historical failure cases and code call chains in reverse, possible root causes can be traced, and time-series data can be matched to locate bottlenecks caused by human error or other reasons.

[0058] The root cause and its confidence level are output through two-way cross-validation.

[0059] The specific algorithm for bidirectional association analysis: Forward propagation—starting from the bottleneck phenomenon node that triggers the alarm, breadth-first search (BFS) is used to traverse the knowledge graph to collect abnormal indicator nodes and log event nodes that are directly connected to the bottleneck phenomenon (within 1 hop), and then it is extended to system module nodes within 2 hops.

[0060] Backpropagation—Starting from the suspected module node, it searches backwards for historical failure case nodes, filtering out cases closest to the current time window through timestamp matching and call chain path alignment. Confidence Calculation—For each candidate root cause, a weighted voting mechanism is used to calculate the confidence score, considering the number of positive hits, the reverse matching degree, and the time sequence alignment. When the confidence score is greater than or equal to the preset confidence threshold (default 0.9), it is determined to be the root cause.

[0061] 3.3. Evidence Chain and Remediation Suggestion Output: Automatically generate a complete and traceable evidence chain containing "bottleneck phenomenon → related indicators → related logs → root cause (confidence ≥ preset confidence threshold) → historical cases → remediation suggestions", and output it in visual or text form.

[0062] Step S4, Performance Trend Prediction: 4.1. Scenario-based LSTM model construction: Collect historical performance data and real-time data for more than 6 months, and combine them with the load patterns of the clinical rehabilitation system (such as daily rehabilitation assessment peaks and monthly efficacy statistics batch tasks) to optimize the training parameters of the LSTM deep learning model, such as the temporal window length, the number of hidden layer nodes, and the learning rate, and build a prediction model adapted to the clinical scenario.

[0063] Specific parameter configurations for the LSTM model: The default time window length is 24 hours (i.e., using data from the past 24 hours to predict the future), which can be adjusted to 7 days according to the periodic pattern; Number of hidden layer nodes: 64 for the first layer and 32 for the second layer, using Dropout (dropout rate 0.2) to prevent overfitting; Learning rate: Initially 0.001, using the Adam optimizer, decaying by 0.9 every 10 epochs; Loss function: Mean Squared Error (MSE), with a regularization term added; Training data volume: At least 6 months of historical data, using a rolling validation method (first 80% training, last 20% validation).

[0064] As an alternative, GRU, bidirectional LSTM, or Transformer-based time series models (such as Informer) can be used. Only the model structure needs to be adjusted, and the prediction method principle is the same. All of these are equivalent implementations of the present invention.

[0065] 4.2. Two-dimensional prediction: The trained LSTM model is used to perform two-dimensional prediction: short-term prediction (1-2 days) is used to predict performance indicators during peak periods to guide resource expansion; long-term prediction (1-3 months) is used to identify implicit performance degradation trends such as memory leaks and connection pool exhaustion, with a focus on the image transmission module and rehabilitation assessment module.

[0066] 4.3. Early Warning Output: Generates visualized forecast reports. When a potential bottleneck is predicted, an early warning is automatically triggered, and targeted optimization suggestions are output simultaneously.

[0067] Step S5: Analysis Results Feedback and Model Optimization (Closed-Loop Management): Collect user feedback on the analysis results, supplement the knowledge graph and historical fault case library with new data generated during the analysis process (including newly discovered abnormal patterns, new root cause cases, prediction error feedback, etc.), and iteratively optimize the parameters of the isolated forest, LogBERT, knowledge graph inference and LSTM models accordingly to continuously improve the accuracy of analysis and scenario adaptability, forming a closed loop of the whole process.

[0068] Specific implementation methods for feedback optimization: User feedback collection – provides “Accurate / Inaccurate” and “Helpful / Useless” buttons through the results output and visualization module, and allows users to annotate the correct root causes.

[0069] Incremental learning – For the anomaly detection unit, the isolated forest is incrementally trained using newly confirmed anomaly samples (updating the tree splitting threshold); for LogBERT, the BERT model is fine-tuned using a new log template.

[0070] Regular retraining – Once a week, retrain the LSTM model using the accumulated new data and update the parameters.

[0071] Knowledge graph update – When a new root cause is identified and its confidence level is higher than 0.9, it is automatically added as a new historical case node, and the associated edges are updated.

[0072] Reference Figure 6 The diagram shown is a functional module schematic of the deep intelligent analysis device based on AI performance test data from the hospital-wide integrated clinical rehabilitation system of the present invention.

[0073] The deep intelligent analysis device based on AI performance test data from a hospital-wide integrated clinical rehabilitation system, as described in this invention, can be installed in an electronic device. Depending on the functions implemented, the deep intelligent analysis device based on AI performance test data from a hospital-wide integrated clinical rehabilitation system may include a multi-source data acquisition module, a data preprocessing module, an AI analysis core module, a result output and visualization module, a knowledge graph and database module, and a feedback optimization module. These modules can also be referred to as units, which are a series of computer program segments that can be executed by the processor of an electronic device and perform a fixed function, stored in the memory of the electronic device.

[0074] In this embodiment, the functions of each module / unit are as follows: The multi-source data acquisition module is used to synchronously collect indicator data, log data, and auxiliary correlation data generated by the performance test of the hospital's integrated clinical rehabilitation system, supporting real-time and offline acquisition. The data preprocessing module is used to clean, standardize, and fuse multi-source data, establish correlation mappings between various types of data, and generate a unified AI analysis dataset. The core AI analysis module integrates an anomaly detection unit, a root cause localization unit, and a trend prediction unit. The anomaly detection unit employs an optimized isolated forest algorithm and a LogBERT model to achieve real-time detection and pattern recognition of performance anomalies, and combines clinical workload feature data to distinguish between true and false anomalies. The root cause localization unit uses a performance analysis knowledge graph for bidirectional correlation analysis to achieve root cause localization of performance bottlenecks and generation of evidence chains. The trend prediction unit uses a scenario-based LSTM model to predict short-term and long-term performance trends. The results output and visualization module is used to output analysis results in the form of charts or reports, and supports querying and exporting. The knowledge graph and database module is used to store the performance analysis knowledge graph, as well as various raw data, analysis results, and feedback data, and supports automatic updates. The feedback optimization module is used to collect user feedback, supplement the analysis data into the knowledge graph and case library, and iteratively optimize the parameters of each algorithm in the core AI analysis module to form a closed loop throughout the entire process.

[0075] As a preferred embodiment, in the anomaly detection unit, the optimized Isolation Forest algorithm is used to perform unsupervised modeling of time-series indicator data in multiple dimensions and automatically identify indicator mutation patterns; the LogBERT model is used to perform cluster analysis on log data, extract semantic features and identify similar error patterns; the anomaly detection unit also combines clinical workload feature data to distinguish between true and false anomalies in the detection results.

[0076] As a preferred embodiment, the root cause localization unit integrates system topology, historical failure cases, code call chains, and system configuration data. Through bidirectional analysis—forwardly linking bottleneck phenomena with indicators and logs, and backwardly combining historical cases and code call chains—it outputs the root cause and automatically generates a traceable chain of evidence containing bottleneck phenomena, related indicators, related logs, root cause, and remediation suggestions.

[0077] As a preferred approach, the collaboration and data flow between the modules are as follows: The multi-source data acquisition module synchronously collects indicator data, log data, and auxiliary correlation data, supporting both real-time and offline acquisition. The data preprocessing module cleans, standardizes, and fuses the collected data, establishing correlation mappings between various data types to generate a unified AI analysis dataset. Within the core AI analysis module: the anomaly detection unit employs an optimized Isolation Forest algorithm and LogBERT model to achieve real-time detection and pattern recognition of performance anomalies, and distinguishes between true and false anomalies by combining clinical workload characteristic data; the root cause localization unit uses a performance analysis knowledge graph for bidirectional correlation analysis to locate the root cause of performance bottlenecks and generate evidence chains; the trend prediction unit uses a scenario-based LSTM model to predict short-term and long-term performance trends; the knowledge graph and database module stores the knowledge graph, raw data, analysis results, and feedback data; the results output and visualization module outputs analysis results in the form of charts or reports, supporting querying and export; the feedback optimization module collects user feedback, supplements the analysis data to the knowledge graph and case library, and iteratively optimizes the parameters of each algorithm in the core AI analysis layer, forming a closed-loop process.

[0078] Reference Figure 7 The diagram shown is a schematic representation of a preferred embodiment of the electronic device 1 of the present invention.

[0079] The electronic device 1 includes, but is not limited to, a memory 11, a processor 12, a display 13, and a network interface 14. The electronic device 1 connects to a network via the network interface 14 to acquire raw data. The network can be an intranet, the Internet, a Global System for Mobile Communications (GSM), Wideband Code Division Multiple Access (WCDMA), a 4G network, a 5G network, Bluetooth, Wi-Fi, a voice communication network, or other wireless or wired networks.

[0080] The memory 11 includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, etc.

[0081] Preferably, the memory 11 can be an internal storage unit of the electronic device 1, such as a hard disk or RAM of the electronic device 1. In other embodiments, the memory 11 can also be an external storage device of the electronic device 1, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD) card, flash card, etc., equipped with the electronic device 1. Of course, the memory 11 can also include both internal storage units and external storage devices of the electronic device 1.

[0082] In this embodiment, the memory 11 is typically used to store the operating system and various application software installed on the electronic device 1, such as the program code of the deep intelligent analysis program 10 based on the AI ​​performance test data of the integrated clinical rehabilitation system for the whole hospital. In addition, the memory 11 can also be used to temporarily store various types of data that have been output or will be output.

[0083] The processor 12 can preferably be a central processing unit (CPU), controller, microcontroller, microprocessor, or other data processing chip. The processor 12 is typically used to control the overall operation of the electronic device 1, such as performing data interaction or communication-related control and processing. In this embodiment, the processor 12 is used to run program code stored in the memory 11 or process data, such as running the program code of a deep intelligent analysis program 10 based on AI performance test data from a hospital-wide integrated clinical rehabilitation system.

[0084] The display 13 can be referred to as a display screen or display unit. Preferably, the display 13 can be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an organic light-emitting diode (OLED) touchscreen, etc. The display 13 is used to display information processed in the electronic device 1 and to display a visual working interface, such as displaying the results of data statistics.

[0085] The network interface 14 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface), which is typically used to establish a communication connection between electronic device 1 and other electronic devices.

[0086] Figure 7 Only an electronic device 1 with components 11-14 and a deep intelligent analysis program 10 based on AI performance test data from a hospital-wide integrated clinical rehabilitation system is shown. However, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0087] Optionally, the electronic device 1 may further include a user interface, which may include a display, an input unit such as a keyboard, and optionally, a standard wired interface or a wireless interface. Preferably, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an organic light-emitting diode (OLED) touchscreen. The display may also be appropriately referred to as a screen or display unit, used to display information processed in the electronic device 1 and to display a visual user interface.

[0088] The electronic device 1 may also include radio frequency (RF) circuits, sensors, and audio circuits, etc., which will not be described in detail here.

[0089] In the above embodiment, when the processor 12 executes the deep intelligent analysis program 10 based on the AI ​​performance test data of the integrated clinical rehabilitation system of the whole hospital stored in the memory 11, it can perform the following steps: S1. Synchronously collect multi-source data generated from the performance test of the hospital's integrated clinical rehabilitation system. The multi-source data includes indicator data, log data, and auxiliary correlation data. Clean, standardize, and fuse the multi-source data to establish the correlation mapping between indicator data, log data, and auxiliary correlation data, and generate a unified AI analysis dataset. S2. Based on the AI ​​analysis dataset, the optimized isolated forest algorithm is used to detect anomalies in the indicator data and identify indicator mutation patterns; the LogBERT model is used to perform cluster analysis on the log data to identify similar error patterns; and the detection results are distinguished between true and false anomalies by combining clinical workload characteristic data to determine the true bottleneck and trigger alarms. S3. When an alarm is triggered, based on a pre-built performance analysis knowledge graph, the root cause of the performance bottleneck is located through bidirectional correlation analysis, and a traceable evidence chain containing the bottleneck phenomenon, related indicators, related logs, root cause and repair suggestions is automatically generated. S4. Based on AI analysis datasets and historical performance data, a scenario-based LSTM model is used to predict short-term and long-term performance trends. When a potential bottleneck is predicted, an early warning is triggered and optimization suggestions are output. S5. Collect feedback data on the analysis results, supplement the analysis data into the knowledge graph and historical case library, and iteratively optimize the model parameters of each algorithm to form a closed loop of the entire process.

[0090] The storage device can be the memory 11 of the electronic device 1, or it can be other storage devices that are communicatively connected to the electronic device 1.

[0091] For a detailed explanation of the above steps, please refer to the above. Figure 6 Functional module diagram of an embodiment of a deep intelligent analysis device based on AI performance test data from a hospital-wide integrated clinical rehabilitation system, and... Figure 1-5 A flowchart illustrating an implementation example of a deep intelligent analysis method based on AI performance test data from a hospital-wide integrated clinical rehabilitation system.

[0092] Furthermore, this invention also proposes a computer-readable storage medium, which can be non-volatile or volatile. This computer-readable storage medium can be any one or any combination of several of the following: hard disk, multimedia card, SD card, flash memory card, SMC, read-only memory (ROM), erasable programmable read-only memory (EPROM), portable compact disk read-only memory (CD-ROM), USB memory, etc. The computer-readable storage medium includes a data storage area and a program storage area. The data storage area stores data created based on the use of blockchain nodes, and the program storage area stores a deep intelligent analysis program 10 based on AI performance test data from a hospital-wide integrated clinical rehabilitation system. When executed by a processor, the deep intelligent analysis program 10 based on AI performance test data from a hospital-wide integrated clinical rehabilitation system performs the following operations: S1. Synchronously collect multi-source data generated from the performance test of the hospital's integrated clinical rehabilitation system. The multi-source data includes indicator data, log data, and auxiliary correlation data. Clean, standardize, and fuse the multi-source data to establish the correlation mapping between indicator data, log data, and auxiliary correlation data, and generate a unified AI analysis dataset. S2. Based on the AI ​​analysis dataset, the optimized isolated forest algorithm is used to detect anomalies in the indicator data and identify indicator mutation patterns; the LogBERT model is used to perform cluster analysis on the log data to identify similar error patterns; and the detection results are distinguished between true and false anomalies by combining clinical workload characteristic data to determine the true bottleneck and trigger alarms. S3. When an alarm is triggered, based on a pre-built performance analysis knowledge graph, the root cause of the performance bottleneck is located through bidirectional correlation analysis, and a traceable evidence chain containing the bottleneck phenomenon, related indicators, related logs, root cause and repair suggestions is automatically generated. S4. Based on AI analysis datasets and historical performance data, a scenario-based LSTM model is used to predict short-term and long-term performance trends. When a potential bottleneck is predicted, an early warning is triggered and optimization suggestions are output. S5. Collect feedback data on the analysis results, supplement the analysis data into the knowledge graph and historical case library, and iteratively optimize the model parameters of each algorithm to form a closed loop of the entire process.

[0093] The specific implementation of the computer-readable storage medium of the present invention is largely the same as the specific implementation of the deep intelligent analysis method based on the AI ​​performance test data of the hospital-wide integrated clinical rehabilitation system, and will not be repeated here.

[0094] To better explain the technical solution of the present invention, the present invention also provides the following specific embodiments: Example 1: Detection and Root Cause Localization of Performance Abnormalities in Clinical Rehabilitation Image Transmission Module Application Scenario: In a hospital-wide integrated clinical rehabilitation system at a top-tier hospital, the image transmission module is responsible for uploading, downloading, and querying DICOM rehabilitation images. After performance testing, the system generated millions of data points (including over 20 dimensions such as image transmission response time, IOPS, network bandwidth, and error rate) and gigabytes of operational log data. Traditional manual analysis would take more than two hours, and is prone to missing anomalies, failing to quickly pinpoint root causes and impacting clinical diagnosis and treatment.

[0095] Implementation process: Step S1: Multi-source performance test data acquisition and preprocessing: The multi-source data acquisition module collects 20+ dimensions of indicator data from the image transmission module at a frequency of once per second; the acquisition agent is deployed on the image transmission service node and the database node; the entire module's operation log is collected synchronously and centrally through the ELK technology stack; at the same time, system topology, historical fault cases (including 5000+ records), code call chains, and clinical workload characteristic data are loaded.

[0096] The data preprocessing module removes outliers from the indicator data using the 3σ principle and fills in missing values ​​using linear interpolation, normalizing the indicators to the [0,1] range. It then uses regular expressions to extract key fields from the log data and filters DEBUG logs. Finally, it establishes a mapping between "indicators-logs-system topology-historical faults," using timestamp alignment and Trace ID association to generate a unified AI analysis dataset. Preprocessing takes approximately 8 minutes.

[0097] Step S2: Real-time detection and pattern recognition of performance anomalies: The optimized isolated forest algorithm (200 trees, 128 subsamples, and 0.05 adaptive contamination coefficient) in the anomaly detection unit monitored real-time indicator data and found a significant abrupt change pattern: "the image transmission response time suddenly increased from 450ms to 2.1s and the error rate increased from 0.08% to 5.2%".

[0098] The LogBERT model (using BERT-base-uncased fine-tuned on medical logs, DBSCAN clustering ε=0.5, minPts=5) performed cluster analysis on the log data and identified a similar error pattern of "MySQL connection timeout" occurring simultaneously in multiple image transmission services.

[0099] Based on the clinical workload characteristics data (currently 10:00 AM, which is a peak time for image transmission), the system determined that the abnormality lasted for more than 30 seconds and was associated with the error log, confirming it as a "true bottleneck" and triggering an alarm.

[0100] Step S3: Root cause identification and evidence chain generation of performance bottlenecks: The root cause localization unit performs bidirectional correlation analysis based on a knowledge graph (Neo4j storage) containing more than 1,200 entity nodes and more than 3,000 relation edges: in the forward direction, starting from the node "image transmission response time spike", BFS traverses the correlation index anomalies and log events to lock the database module; in the reverse direction, it combines historical failure cases and code call chains to trace back and find that the image query logic was modified in the most recent code release.

[0101] The confidence level was calculated using a weighted voting mechanism, and the root cause was identified as: "An error in the new code's query logic caused the status field of the orders table to not use an index, triggering a full table scan, exhausting the database connection pool, and consequently exceeding the IOPS limit." The system output a root cause confidence level of 0.92.

[0102] The system automatically generates a chain of evidence: "Bottleneck phenomenon (image transmission response time 2.1s) → Correlation indicators (error rate 5.2%, peak IOPS) → Correlation log (MySQL connection timeout) → Root cause (query logic error, full table scan) → Historical cases (similar index failure issues) → Repair suggestions (optimize SQL statements, add composite indexes)". The average response time for bidirectional correlation analysis is 1.8 seconds.

[0103] Step S4, Long-term performance trend prediction: The trend prediction unit uses historical data from the image transmission module over the past 8 months and a contextualized LSTM model (24-hour window, 64 / 32 hidden layers, learning rate 0.001, Dropout 0.2) to make short-term and long-term predictions. Short-term predictions show that without improvement, image transmission response time during peak hours the following day will exceed 3 seconds. Long-term predictions show that if the query logic is not optimized, connection pool exhaustion will occur 3-5 times per week, and memory usage will show a linear increasing trend.

[0104] Step S5: Analysis Results Feedback and Model Optimization The maintenance personnel optimized the SQL statements and added indexes according to the repair suggestions. After the repair, the image transmission response time was restored to within 450ms. The system supplemented the analysis data (abnormal patterns, root cause cases, and repair solutions) into the knowledge graph and historical case library, fine-tuned the Isolation Forest and LogBERT models through incremental learning, and triggered weekly LSTM retraining.

[0105] Application effect: The entire analysis process takes only 5 minutes (compared to 2.5 hours for traditional manual analysis), improving efficiency by 30 times.

[0106] The false alarm rate was reduced to 6.8%, and the root cause localization accuracy reached 92%.

[0107] After the repair, the image transmission efficiency was improved by 37%, avoiding the impact of image transmission interruption on clinical rehabilitation diagnosis and treatment.

[0108] Example 2: Identification of Latent Memory Leaks in the Rehabilitation Assessment Module Application Scenario: In the integrated clinical rehabilitation system of a secondary-level hospital, the rehabilitation assessment module experienced slow response and occasional crashes after long-term operation, but routine performance testing and manual analysis failed to identify a clear cause. A hidden memory leak is suspected.

[0109] Implementation process: Steps S1-S2 (Data Acquisition and Anomaly Detection): Collect historical performance data (including memory usage, GC frequency, response time, etc.) and operation logs of the rehabilitation assessment module over the past 8 months; the anomaly detection unit did not find any obvious mutations, but the LSTM trend prediction model needs to make predictions based on long-term data.

[0110] Step S3 (Trend Prediction Model Training): A scenario-based LSTM model is trained using 8 months of historical data. The model input consists of daily memory usage time series, GC frequency, request volume, and other features. The output is the future memory usage trend. After parameter optimization (time series window set to 30 days, number of hidden layer nodes 64, learning rate 0.001), the model prediction error is controlled at 4.2%.

[0111] Step S4 (Trend Prediction and Early Warning): The LSTM model predicts that without intervention, the memory usage of the rehabilitation assessment module will reach 95% of the preset threshold after 3 weeks. Subsequently, frequent GC will cause the response time to exceed 3 seconds, posing a risk of crash. The system outputs an early warning and provides optimization suggestions: "Check for unreleased object references in the code and add a memory cleanup mechanism; expand the heap memory if necessary."

[0112] Step S5 (Implementation and Feedback Optimization): The development team checked the code based on the suggestions and found that a cached object was not being cleaned up in time. After the fix, the memory usage stabilized within a safe range. At the same time, the memory capacity was increased by 20% as redundancy. The system added the analysis data (memory leak pattern, fix solution) to the knowledge graph and updated the training set of the LSTM model, further optimizing the prediction accuracy.

[0113] Application effect: The hidden memory leak was detected three weeks in advance, preventing system crashes and ensuring the continuity of rehabilitation assessment services.

[0114] The cost of reducing medical interruptions and repairs due to system crashes is approximately 120,000 yuan.

[0115] The practicality and reliability of this invention in identifying latent performance degradation have been verified.

[0116] Alternative embodiments Those skilled in the art will understand that, without departing from the core idea of ​​this invention, the following technical features can be substituted, and all such substitutions fall within the protection scope of this invention: Anomaly detection algorithm alternatives: In addition to Isolation Forest, One-Class Support Vector Machine (One-ClassSVM) and Autoencoder can be used for unsupervised anomaly detection; in addition to LogBERT, models such as LogAnomaly, DeepLog, and PLELog can be used for log anomaly identification.

[0117] Root cause localization alternatives: In addition to knowledge graphs, decision trees, random forests, or rule-based expert systems can be used for root cause inference, or causal inference algorithms (such as PC algorithms and LiNGAM) can be used to discover causal relationships from data.

[0118] Trend prediction model alternatives: In addition to LSTM, models such as GRU, TCN (temporal convolutional network), and Transformer (such as Informer and Autoformer) can also be used for time series prediction.

[0119] Data fusion method alternatives: In addition to association based on timestamps and call chain IDs, graph neural networks (GNNs) can be used to encode multi-source data into a graph structure to achieve end-to-end learning.

[0120] Alternative feedback optimization methods: In addition to incremental learning and periodic retraining, online learning can also be used to update model parameters immediately after each analysis.

[0121] It should be noted that the sequence numbers of the above embodiments of the present invention are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0122] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, electronic device, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0123] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A deep intelligent analysis method for AI performance test data based on a hospital-wide integrated clinical rehabilitation system, characterized in that, Includes the following steps: S1. Multi-source performance test data acquisition and preprocessing: Simultaneously collect multi-source data generated by the performance test of the hospital's integrated clinical rehabilitation system. The multi-source data includes indicator data, log data, and auxiliary correlation data. Clean, standardize, and fuse the multi-source data to establish the correlation mapping between the indicator data, log data, and auxiliary correlation data, and generate a unified AI analysis dataset. S2. Real-time detection and pattern recognition of performance anomalies: Based on the AI ​​analysis dataset, an optimized isolated forest algorithm is used to detect anomalies in the indicator data and identify abrupt change patterns in the indicators. The LogBERT model is used to perform cluster analysis on log data to identify similar error patterns; combined with clinical workload characteristic data, the test results are distinguished between true and false anomalies, and true bottlenecks are identified and alarms are triggered. S3. Root cause location and evidence chain generation of performance bottlenecks: When an alarm is triggered, the root cause of the performance bottleneck is located through bidirectional correlation analysis based on a pre-built performance analysis knowledge graph, and a traceable evidence chain containing bottleneck phenomena, related indicators, related logs, root cause and remediation suggestions is automatically generated. S4. Performance Trend Prediction: Based on the AI ​​analysis dataset and historical performance data, a scenario-based LSTM model is used to predict short-term and long-term performance trends. When a potential bottleneck is predicted, an early warning is triggered and optimization suggestions are output. S5. Analysis Result Feedback and Model Optimization: Collect feedback data on the analysis results, supplement the analysis data into the knowledge graph and historical case library, and iteratively optimize the model parameters of each algorithm to form a closed loop of the entire process.

2. The deep intelligent analysis method for AI performance test data based on the integrated clinical rehabilitation system of the whole hospital, as described in claim 1, is characterized in that... In step S1: The metrics data include at least several core performance metrics from multiple dimensions, such as CPU utilization, memory usage, rehabilitation assessment report generation response time, image transmission response time, error rate, and IOPS. The log data includes the system's full module operation logs, with clinically relevant core logs retained after preprocessing. The auxiliary correlation data includes system topology, historical failure cases, code call chains, and clinical workload characteristics data.

3. The deep intelligent analysis method for AI performance test data based on the integrated clinical rehabilitation system of the whole hospital, as described in claim 1, is characterized in that... The establishment of association mapping in step S1 specifically involves: establishing association mappings for four types of data, namely "indicators, logs, system topology, and historical faults", generating a unified AI analysis dataset, and storing the preprocessed data in a distributed database.

4. The deep intelligent analysis method for AI performance test data based on the integrated clinical rehabilitation system of the whole hospital, as described in claim 1, is characterized in that... The distinction between true and false anomalies in step S2 specifically includes: Cross-validate the test results with clinical workload characteristics data and system status data; For transient abnormalities whose duration is within a preset threshold and do not affect core clinical functions, they are judged as false abnormalities and are only recorded without alarm. If the duration exceeds the preset threshold and is associated with error logs, it is determined to be a true bottleneck, and an alarm is triggered.

5. The deep intelligent analysis method based on AI performance test data of a hospital-wide integrated clinical rehabilitation system according to any one of claims 1-4, characterized in that, In step S3: The performance analysis knowledge graph integrates system topology, historical failure cases, code call chains, and system configuration data. The nodes in the graph include bottleneck phenomena, abnormal indicators, log events, and system modules, and the edges represent the relationships. It also has an automatic update function. The bidirectional correlation analysis includes: forward correlation of bottleneck phenomena with abnormal indicators and log events to identify suspicious system modules, and reverse combination of historical failure cases and code call chains to trace the root cause, and verification through time series data matching. The confidence level of the root cause is not lower than a preset confidence threshold.

6. The deep intelligent analysis method for AI performance test data based on a hospital-wide integrated clinical rehabilitation system according to any one of claims 1-4, characterized in that, In step S4: The scenario-based LSTM model is trained based on historical performance data and clinical workload patterns within a preset historical data period. The time window length, number of hidden layer nodes, and learning rate parameters are optimized, and the model prediction error is controlled within a preset error threshold. The time range of the short-term forecast is a preset short-term time window, which is used to predict performance indicators during peak periods and guide resource expansion. The time range for the long-term prediction is a preset long-term time window, used to identify implicit performance degradation trends such as memory leaks and connection pool exhaustion.

7. The deep intelligent analysis method for AI performance test data based on the integrated clinical rehabilitation system of the whole hospital, as described in claim 1, is characterized in that... In step S5, iteratively optimizing the model parameters of each algorithm specifically includes: adding newly discovered abnormal patterns, new root cause cases, and prediction error feedback data to the knowledge graph and historical case library during this analysis, and adjusting the detection sensitivity of the Isolation Forest algorithm, the clustering threshold of the LogBERT model, and the training parameters of the LSTM model accordingly to achieve continuous optimization of the model.

8. A deep intelligent analysis device based on AI performance test data from a hospital-wide integrated clinical rehabilitation system, characterized in that, include: The multi-source data acquisition module is used to synchronously collect indicator data, log data, and auxiliary correlation data generated by the performance test of the hospital's integrated clinical rehabilitation system, supporting real-time and offline acquisition. The data preprocessing module is used to clean, standardize, and fuse the multi-source data, establish correlation mappings between multiple types of data, and generate a unified AI analysis dataset. The core AI analysis module integrates an anomaly detection unit, a root cause localization unit, and a trend prediction unit. The anomaly detection unit employs an optimized isolated forest algorithm and a LogBERT model to achieve real-time detection and pattern recognition of performance anomalies, and combines clinical workload feature data to distinguish between true and false anomalies. The root cause localization unit uses a performance analysis knowledge graph for bidirectional correlation analysis to achieve root cause localization of performance bottlenecks and generation of evidence chains. The trend prediction unit uses a scenario-based LSTM model to predict short-term and long-term performance trends. The results output and visualization module is used to output analysis results in the form of charts or reports, and supports querying and exporting. The knowledge graph and database module is used to store the performance analysis knowledge graph, as well as various raw data, analysis results, and feedback data, and supports automatic updates. The feedback optimization module is used to collect user feedback, supplement the analysis data into the knowledge graph and case library, and iteratively optimize the parameters of each algorithm in the core AI analysis module to form a closed loop throughout the entire process.

9. The deep intelligent analysis device based on AI performance test data of a hospital-wide integrated clinical rehabilitation system according to claim 8, characterized in that, In the anomaly detection unit, the optimized isolated forest algorithm is used to perform unsupervised modeling of time-series indicator data in multiple dimensions and automatically identify indicator mutation patterns; the LogBERT model is used to perform cluster analysis on log data, extract semantic features and identify similar error patterns; the anomaly detection unit also combines clinical workload feature data to distinguish between true and false anomalies in the detection results.

10. The deep intelligent analysis device based on AI performance test data of a hospital-wide integrated clinical rehabilitation system according to claim 8, characterized in that, In the root cause localization unit, the performance analysis knowledge graph integrates system topology, historical failure cases, code call chains, and system configuration data. Through bidirectional analysis—forwardly linking bottleneck phenomena with indicators and logs, and backwardly combining historical cases and code call chains—it outputs the root cause and automatically generates a traceable evidence chain containing bottleneck phenomena, related indicators, related logs, root cause, and repair suggestions.

Images