Instrument log recording and intelligent analysis method in medical instrument software system

Abstract

The invention provides an instrument log recording and intelligent analysis method in a medical instrument software system, and the method comprises the steps: carrying out the preprocessing of operation log data of medical instrument equipment, carrying out the unified conversion of the log data of different sources and different formats, and obtaining the operation log data of the medical instrument equipment after the standardization processing; mining operation state change rules and abnormal modes of the medical equipment in different diagnosis and treatment stages through a time sequence mode according to the extracted log data semantic features in combination with the diagnosis and treatment workflow of the medical equipment; and analyzing the actual process of the equipment participating in the diagnosis and treatment activity from the equipment log data, and comparing and analyzing the actual process with a standard diagnosis and treatment work process to determine the information of the diagnosis and treatment work process, including a key path, a bottleneck link and an abnormal deviation.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to an instrument log recording and intelligent analysis method in a medical instrument software system. Background Art

[0002] Medical devices generate a large amount of log data during actual operation, recording important information such as the operating status, parameter changes, and abnormal conditions of the equipment. These log data not only reflect the operating rules of the equipment, but also contain potential failure modes. However, due to the large amount of data and complex format, it is very difficult to directly use these data. At the same time, the operation of medical devices is closely related to the clinical diagnosis and treatment workflow, and the operating status of the equipment will directly affect the development of diagnosis and treatment activities and the guarantee of medical quality. Therefore, there is an urgent need for an effective method that can combine the operating rules of medical devices and the diagnosis and treatment workflow, deeply analyze the intrinsic connection between equipment logs and diagnosis and treatment events, explore the omen pattern of equipment failure, characterize the dynamic behavior characteristics of equipment in the actual diagnosis and treatment process, and construct equipment operation status maps and diagnosis and treatment workflows to provide data support and decision-making basis for the operation and maintenance management of medical devices and the improvement of diagnosis and treatment quality. However, both the semantic complexity of log data and its correlation with the diagnosis and treatment process make this process full of challenges. In other words, at present, the convenience and effectiveness of data utilization in the analysis of medical device log data and status monitoring and early warning are low. Summary of the invention

[0003] The present invention provides a device log recording and intelligent analysis method in a medical device software system, which mainly includes: Preprocess the operation log data of medical devices and equipment, and uniformly convert the log data from different sources and in different formats to obtain the normalized operation log data of medical devices and equipment; perform semantic understanding and feature extraction on the normalized operation log data of medical devices and equipment, and identify the keywords, terms and time in the log data through word segmentation, part-of-speech tagging and named entity recognition, and mine the semantic features of the log data; combine the extracted semantic features of the log data with the medical device and equipment diagnosis and treatment workflow, and mine the operation status change rules and abnormal patterns of medical devices and equipment at different diagnosis and treatment stages through time series patterns; graphically represent the operation status change rules and abnormal patterns of medical devices and equipment at different diagnosis and treatment stages, and construct a medical device and equipment operation status map to characterize the dynamics of the equipment in the actual diagnosis and treatment process. state behavior characteristics; based on the medical device operation status map, by real-time monitoring of the equipment operation parameters and status changes, determine whether the equipment operation status is abnormal. If the equipment operation status is detected to be abnormal or the parameters exceed the preset threshold range, a fault warning is issued to prompt the equipment to carry out maintenance work and record the medical device operation status monitoring warning information, and start the corresponding equipment maintenance plan; analyze the actual process of the equipment participating in the diagnosis and treatment activities from the equipment log data, and determine the diagnosis and treatment workflow information by comparing and analyzing with the standard diagnosis and treatment workflow, including critical paths, bottleneck links and abnormal deviations; integrate the medical device operation status monitoring warning information, diagnosis and treatment workflow information and equipment maintenance plan to form a device failure prediction and safety risk assessment method based on log big data, and apply it to the medical device software system.

[0004] The technical solution provided by the embodiment of the present invention may have the following beneficial effects: The present invention discloses a device log recording and intelligent analysis method in a medical device software system. The method preprocesses, semantically understands and extracts features from the operation log of the medical device, mines the operation status change rules and abnormal patterns of the device at different diagnosis and treatment stages, and constructs a device operation status map. Based on this map, the present invention monitors the device operation parameters in real time, detects abnormalities and issues fault warnings. At the same time, the present invention analyzes the actual process of the device participating in the diagnosis and treatment activities, compares it with the standard process, and determines the critical path, bottleneck link and abnormal deviation. By integrating the equipment operation status monitoring and early warning information, diagnosis and treatment workflow information and equipment inspection and maintenance plan, a device failure prediction and safety risk assessment method is formed, which improves the operating efficiency and safety of medical devices and optimizes the diagnosis and treatment process. BRIEF DESCRIPTION OF THE DRAWINGS

[0005] Figure 1 The present invention is a flowchart of a device log recording and intelligent analysis method in a medical device software system.

[0006] Figure 2 A schematic diagram of a device log recording and intelligent analysis method in a medical device software system of the present invention.

[0007] Figure 3 It is another schematic diagram of the device log recording and intelligent analysis method in the medical device software system of the present invention. DETAILED DESCRIPTION

[0008] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this specification.

[0009] like Figure 1-3 In this embodiment, a device log recording and intelligent analysis method in a medical device software system may specifically include: S101. Preprocess the medical device operation log data, convert the log data from different sources and in different formats uniformly, and obtain the medical device operation log data after normalization.

[0010] Receive operation log data sent by medical equipment; perform quality assessment on the operation log data according to a preset data quality scoring standard to obtain a quality assessment score; if the quality assessment score is higher than a preset threshold, determine that the operation log data is valid data; clean the valid data to obtain standardized log data; align the standardized log data according to a unified time scale to obtain time series data; segment the time series data, calculate the statistical features within each time window, and obtain a time series feature vector; map the time series feature vector to a preset general data structure template to obtain feature data in a unified format; perform dimensionality reduction on the feature data in a unified format to obtain feature data after dimensionality reduction; use a clustering algorithm to group the feature data after dimensionality reduction to obtain normalized operation log data of medical equipment.

[0011] Specifically, the data quality of the collected medical device operation log data is evaluated, regular expression matching is used to check the consistency of data format, the Z-score method is used to identify outliers, the original log data is scored according to the preset data quality scoring standard, and valid data with a quality score higher than the threshold is screened out. The noise and outliers in the log data are removed through data cleaning, the mutation points in the data are smoothed by the moving average method, the missing data are supplemented by the multiple interpolation method based on historical data of similar dates, and the units and dimensions of data in different formats are unified by regularization to obtain the cleaned standardized log data. The cleaned standardized log data is processed in time series, and the timestamp alignment method is used to align the log data of different devices and different time periods according to a unified time scale. The sliding window method is used to segment the data, and the statistical features in each time window are calculated, including the mean, variance, skewness, and kurtosis, to construct the time series feature vector of the equipment operation status. A general data structure template is designed to map data of various formats to this template to achieve unified conversion of log data from different sources and formats. The constructed time series feature vectors were fused and integrated, the t-SNE algorithm was used to reduce the feature dimension, and the DBSCAN clustering algorithm was used to group similar operating states to obtain the normalized medical device operation log data. When evaluating the quality of medical device operation log data, the regular expression "^\d{4}-\d{2}-\d{2}\d{2}:\d{2}:\d{2},\d+,\d+(.\d+)?$" was used to check the format consistency of timestamps, device IDs, and numerical values. The standard score of each data point was calculated by the Z-score method, and it was marked as an outlier if it exceeded 3 standard deviations. The data quality scoring criteria included three dimensions: completeness, accuracy, and timeliness, with weights of 30%, 40%, and 30% respectively. Data with a total score of more than 80 points were considered valid. During the data cleaning process, the 5-point moving average method was used to smooth the mutation points. For missing data, the mean was calculated based on the historical data of the same time period in the previous 7 days to complete the data. Regularization processing converts temperature data in different units into degrees Celsius and pressure data into kilopascals. When processing time series, adjust the timestamps of all devices to a unified scale of one data point every 5 minutes. Use a 60-minute sliding window to calculate the mean, variance, skewness, and kurtosis of the data in the window to construct a 4-dimensional time series feature vector. Design a general data structure template containing the fields: timestamp, device ID, temperature, pressure, flow, and status code, and convert raw data in different formats into this template through a mapping function. In the data fusion and integration stage, the t-SNE algorithm is used to reduce the 4-dimensional feature vector to 2 dimensions, with the parameters set to perplexity 30, learning rate 200, and number of iterations 1000. The DBSCAN clustering algorithm parameters are set to eps=0.5 and MinPts=5 to divide similar operating states into different groups.The resulting normalized log data contains information such as time, device ID, original data, feature vectors, and cluster labels, which is convenient for subsequent analysis and monitoring.

[0012] S102. Perform semantic understanding and feature extraction on the normalized medical device operation log data, identify keywords, terms and time in the log data through word segmentation, part-of-speech tagging and named entity recognition, and mine the semantic features of the log data.

[0013] Obtain the normalized medical device operation log data; construct a medical device professional dictionary based on the log data, which includes terms extracted from the medical device manual and professional literature; perform word segmentation on the log data to obtain the log text sequence after word segmentation. Perform part-of-speech tagging on the log text sequence after word segmentation; use the annotated medical text corpus to train the hidden Markov model; use the trained hidden Markov model to annotate the part-of-speech of each word in the log text sequence to obtain the log text with part-of-speech information. Perform named entity recognition on the log text with part-of-speech information; use the conditional random field algorithm to identify the device name, parameter index, and time expression entity in the log text; use the annotated medical text to train the conditional random field to obtain a structured named entity sequence. Based on the structured named entity sequence, calculate the word importance score, filter out keywords based on the importance score, and identify professional terms in combination with the pre-defined term template. Use the word embedding method to convert keywords, professional terms, and time information into numerical vectors; splice the numerical vectors to obtain the semantic feature vector of the log data.

[0014] Specifically, the normalized medical device operation log data is segmented, and terms are extracted from medical device manuals and professional literature and manually reviewed to build a medical device professional dictionary. The dictionary and rule-based segmentation method is used to segment the log text. The rules include the maximum matching principle and unregistered word recognition. Professional terms are identified through professional dictionaries to obtain the log text sequence after segmentation. Part-of-speech tagging is performed on the log text sequence after segmentation. The hidden Markov model is trained using the annotated medical text corpus to annotate each word with its part of speech, including noun, verb, and adjective. The part-of-speech transfer probability matrix is ​​used to optimize the tagging results to obtain log text with part-of-speech information. Named entity recognition is performed on the log text with part-of-speech information, and the conditional random field algorithm is used to identify entities such as device names, parameter indicators, and time expressions in the log. Parts of speech, context windows, and character features are selected as features of the conditional random field, and the annotated medical text is used for training. The identified time expressions are standardized into a unified format to form a structured named entity sequence. The frequency of occurrence of each entity is counted, the co-occurrence relationship between entities is analyzed, and the entity co-occurrence matrix is ​​constructed. Extract keywords and terms from structured named entity sequences, use the TF-IDF algorithm combined with entity co-occurrence matrix to calculate word importance scores, and select words with higher scores as keywords. Combine pre-defined term templates, such as "noun + verb + noun" or "adjective + noun", to identify professional terms. Use word embedding technology to convert keywords, terms and time information into numerical vectors, and finally concatenate these vectors to construct the semantic feature vector of log data. When processing the operation log data of medical devices, first extract 10,000 terms from 5,000 pages of medical device instructions and 1,000 professional documents. After automatic screening and manual review, finally build a medical device professional dictionary containing 8,000 entries. Use a dictionary and rule-based word segmentation algorithm, set the maximum match length to 6 characters, and the unregistered word recognition threshold to 0.8 to segment the log text. For example, "Blood pressure meter measurement results show 130 / 85mmHg" is segmented into "Blood pressure meter / measurement / result / display / 130 / 85 / mmHg". Then, a hidden Markov model was trained using a corpus of 100,000 annotated medical texts, with the number of states set to 40 and the number of iterations set to 100, with the part of speech labeled for each word. The size of the part of speech transition probability matrix was 40x40, which was used to optimize the labeling results. Then named entity recognition was performed, and the part of speech, the context window of 3 words, and the character features were selected as the features of the conditional random field. 50,000 annotated medical texts were used for training, and the iterations were repeated 500 times. The recognized time expressions were standardized into the "YYYY-MM-DDHH:MM:SS" format. The frequency of occurrence of each entity was counted, and a 100x100 entity co-occurrence matrix was constructed.The TF-IDF algorithm is combined with the entity co-occurrence matrix to score the importance of words, and the top 100 words with the highest scores are selected as keywords. Ten term templates are predefined, such as "noun + verb + noun" or "adjective + noun", to identify professional terms. Finally, the Word2Vec word embedding technology is used to convert keywords, terms and time information into 300-dimensional numerical vectors, and the dimension of the semantic feature vector obtained after splicing is 900. In this way, the original unstructured log text "2024-05-01 09:30:15 Blood pressure monitor A001 measurement results show that patient Zhang San's blood pressure is 130 / 85mmHg, heart rate 75 times / min" is processed and converted into structured data containing equipment name, parameter indicators, and time information, and keywords such as "blood pressure monitor", "measurement", "blood pressure" and professional terms such as "blood pressure measurement results" are extracted, and finally a 900-dimensional semantic feature vector is generated for subsequent data analysis and decision support.

[0015] S103. Combine the extracted semantic features of the log data with the medical device equipment diagnosis and treatment workflow to mine the changing rules and abnormal patterns of the operating status of the medical device equipment at different diagnosis and treatment stages through time series patterns.

[0016] Receive semantic feature vectors carrying unique identification numbers of medical devices, where the semantic feature vectors are generated by different devices in different time periods; use a dynamic time warping algorithm to align the semantic feature vectors to a unified time axis to obtain aligned time series semantic feature data; use a hidden Markov model to divide the aligned time series semantic feature data into diagnosis and treatment stages, and calculate the optimal state sequence; use a sliding window method to extract local features from the time series data divided into diagnosis and treatment stages, calculate statistical features and frequency domain features in each window, and obtain the feature sequence of each diagnosis and treatment stage; perform cluster analysis on the feature sequence of each diagnosis and treatment stage, determine the typical operating state of the equipment at different diagnosis and treatment stages, and explore the law of state change; set an abnormal threshold, and if the Mahalanobis distance between the sample and the cluster center exceeds the abnormal threshold, it is identified as an abnormal pattern, and the cause of the abnormality is inferred based on the decision tree algorithm.

[0017] Specifically, the extracted semantic features of log data are aligned in time series, and the dynamic time warping algorithm is used to align the semantic feature vectors of different devices and different time periods to a unified time axis, and the cosine similarity is used to calculate the distance between vectors. The missing data are processed by cubic spline interpolation, and the redundant data are processed by linear resampling to obtain the aligned time series semantic feature data. According to the pre-defined medical device diagnosis and treatment workflow, the aligned time series semantic feature data are divided into diagnosis and treatment stages using the hidden Markov model. The labeled diagnosis and treatment stage data is used for supervised learning and training of the hidden Markov model. The optimal state sequence is calculated by the Viterbi algorithm, and the continuous time series data is divided into discrete diagnosis and treatment stages, and the corresponding diagnosis and treatment stage labels are annotated for each time point. For the time series data with divided diagnosis and treatment stages, the sliding window method is used to extract local features, and time windows of different sizes are set. Statistical features are calculated in each window, including mean, variance, skewness, kurtosis and frequency domain features, including the first five main frequency components after Fourier transform. The principal component analysis method is used to reduce the feature dimension and retain the principal components that explain 95% of the variance. The recursive feature elimination method is used to select the most relevant feature subset from the reduced-dimensional features to obtain the feature sequence of each diagnosis and treatment stage. The Gaussian mixture model is used to cluster the feature sequence of each diagnosis and treatment stage to determine the typical operating state of the equipment at different diagnosis and treatment stages. The state transition matrix is ​​calculated to represent the state change law, and the distance between the sample and the cluster center is measured by the Mahalanobis distance. The 3σ principle is used to set the abnormal threshold and identify the abnormal pattern. The identified abnormal patterns are classified, and the abnormal causes are inferred based on the decision tree algorithm. Finally, the operating state change law and abnormal pattern of medical equipment at each diagnosis and treatment stage are obtained. When processing the operation log data of medical equipment, the 900-dimensional semantic feature vector is first aligned in time series. The dynamic time warping algorithm is used, the window size is set to 60 seconds, the step size is 10 seconds, and the distance between vectors is calculated using cosine similarity. For missing data, cubic spline interpolation is used to fill, and the interpolation node interval is set to 5 seconds. For redundant data, the linear resampling method is used to downsample the data to one sampling point per second. Subsequently, a hidden Markov model with 5 hidden states was trained using a training set containing 1000 labeled samples, corresponding to 5 main diagnosis and treatment stages. The Viterbi algorithm was used to calculate the optimal state sequence, and the continuous time series data was divided into discrete diagnosis and treatment stages. In each diagnosis and treatment stage, three sliding windows of 30 seconds, 60 seconds and 120 seconds were set, and the step size was 1 / 3 of the window size. Statistical features such as mean, variance, skewness, and kurtosis were calculated in each window, and the first 5 main frequency components were extracted as frequency domain features by fast Fourier transform. The features were reduced to 50 dimensions by principal component analysis, retaining 95.2% of the variance information. Using recursive feature elimination, the 25 most relevant features were finally selected.The Gaussian mixture model was used for clustering, and 3-5 cluster centers were set for each diagnosis and treatment stage to represent different operating states. A 5x5 state transition matrix was calculated to represent the transition probability between states. The 3σ principle was used to set the abnormal threshold, that is, samples whose Mahalanobis distance from the cluster center exceeded 3 times the mean standard deviation were marked as abnormal. Finally, a decision tree algorithm with a depth of 4 was used to classify and infer the causes of abnormal samples based on 25 features, with an accuracy rate of 87.5%.

[0018] S104. Graphically represent the changing patterns and abnormal patterns of the operating status of medical devices at different diagnosis and treatment stages, and construct an operating status map of medical devices to characterize the dynamic behavior characteristics of the equipment during the actual diagnosis and treatment process.

[0019] Obtain the changing rules and abnormal patterns of the operating status of medical equipment at different diagnosis and treatment stages; construct the nodes and edges of the graph according to the changing rules and abnormal patterns, where the nodes include the diagnosis and treatment stage, operating status and abnormal pattern, and the node attributes include duration, parameter value range and frequency of occurrence; the edge represents the conversion relationship between states, and the edge weight represents the conversion probability. For the constructed graph, the Node2Vec algorithm is used for vectorization representation, and the attribute information of the nodes and edges is encoded into a multidimensional vector space; the semantic similarity between nodes is calculated by cosine similarity. If the semantic similarity is higher than the preset threshold, the nodes are merged to obtain a simplified graph structure. For the simplified graph structure, a graph convolutional neural network is used to aggregate and propagate node features. Each layer of the graph convolutional neural network contains multiple neurons; the neural network parameters are trained by the back-propagation algorithm to obtain an updated node feature representation.

[0020] Specifically, the nodes and edges of the graph are constructed according to the changing rules and abnormal patterns of the operating status of medical devices at different diagnosis and treatment stages. The diagnosis and treatment stage, operating status and abnormal pattern are defined as node types, and the node attributes are set including duration, parameter value range and occurrence frequency. The transition relationship between states is represented by the edge, and the weight of the edge represents the transition probability. The triggering conditions of the abnormal pattern are quantified as the attributes of the edge, including the trigger threshold and trigger time. The parameter value range is determined by the mean and standard deviation of the statistical historical data, and the occurrence frequency is calculated according to the number of occurrences of the node in the total sample. After obtaining the initial graph structure, the Node2Vec algorithm is used to vectorize the constructed initial graph, and the attribute information of the nodes and edges is encoded into a 128-dimensional vector space. The semantic similarity between nodes is calculated by cosine similarity, and nodes with similarity higher than 0.9 are merged to simplify the graph structure. The quality of the simplified graph structure is evaluated, and the connectivity, clustering coefficient and average path length of the graph are calculated to ensure that the simplification does not affect the overall structural characteristics of the graph. Based on the evaluation results, a three-layer graph convolutional neural network is used to aggregate and propagate node features. Each layer contains 64, 32, and 16 neurons, respectively, and ReLU is used as the activation function. The network parameters are trained by the back-propagation algorithm to capture the local structural information and global semantic information of the nodes, update the feature representation of the nodes, and strengthen the association between the nodes in the graph. The updated graph is stored in the graph database, and an index is established to accelerate the query. The Cypher query language is used to quickly retrieve the operating status of the equipment and locate the abnormal mode, such as MATCH (n: operating status)-[r: trigger]->(m: abnormal mode) WHEREn. parameter value>r. trigger threshold RETURNm. The force-directed layout algorithm is used to visualize the graph, and the change of the equipment status over time is represented by the time series graph, which intuitively shows the dynamic behavior characteristics of medical devices in the actual diagnosis and treatment process. When constructing the operating status graph of medical devices, 5 diagnosis and treatment stages, 7 operating states, and 3 types of abnormal modes are first defined as node types. Each node is set with attributes such as duration, parameter value range, and occurrence frequency. For example, the average duration of the "preparation phase" node is 10 minutes, the temperature parameter range is 20±2℃, and the frequency of occurrence is 100%. The state transition relationship is represented by an edge, such as the transition probability from "preparation phase" to "inspection phase" is 0.95. The triggering conditions of the abnormal mode are quantified as the attributes of the edge, such as the triggering threshold of the "overtemperature" abnormality is 25℃, and the triggering time is 30 seconds. The Node2Vec algorithm is used to vectorize the initial graph, set the window size to 10, the walk length to 80, and generate a 128-dimensional node vector. Through cosine similarity calculation, it is found that the similarity of the "mild anomaly" and "moderate anomaly" nodes is 0.92, which exceeds the 0.9 threshold, and they are merged into the "mild and moderate anomaly" node.After simplification, the quality of the graph was evaluated, with a connectivity of 1, an average clustering coefficient of 0.68, and an average path length of 2.3, proving that simplification did not significantly affect the graph structure. Subsequently, a three-layer graph convolutional neural network was used, with 64 neurons in the input layer, 32 neurons in the hidden layer, and 16 neurons in the output layer. The ReLU activation function was used, the learning rate was set to 0.01, and 1000 rounds of training were performed. The updated graph was stored in the Neo4j graph database, and a composite index of node types and attributes was established. The Cypher language was used to query abnormal patterns, such as "MATCH(n: running status {stage: 'treatment'})-[r: trigger]->(m: abnormal pattern) WHEREn. temperature>r. trigger threshold RETURNm" to quickly locate temperature anomalies that occurred during treatment. Finally, the force-directed layout algorithm of the D3.js library was used to visualize the graph. The size of the node represents the frequency of occurrence, the thickness of the edge represents the conversion probability, and the depth of the color represents the degree of abnormality. The time series diagram shows the device's state changes within 24 hours, with the x-axis representing time and the y-axis representing different states, clearly presenting the dynamic behavior characteristics of the device.

[0021] S105. Based on the medical device operation status map, the device operation parameters and status changes are monitored in real time to determine whether the device is in an abnormal operation state. If the device operation state is detected to be abnormal or the parameters exceed the preset threshold range, a fault warning is issued to prompt the device to carry out maintenance work and record the medical device operation status monitoring warning information, and start the corresponding equipment maintenance plan.

[0022] The normal operating parameter range and state transition rules of the equipment are obtained from the operation status map of the medical device. The state transition rules include the state transition sequences of different diagnosis and treatment stages. The real-time operation parameters and state information of the equipment are obtained, and the operation parameters and state information are preprocessed to obtain a standardized equipment operation data stream. The standardized equipment operation data stream is analyzed using the isolation forest algorithm, and it is determined whether the equipment is in an abnormal operation state by comparing it with the preset parameter thresholds and state transition rules. According to the abnormal operation state, the corresponding level of fault warning information is generated, and the fault warning information includes abnormal parameters, abnormal time and abnormal type.

[0023] Specifically, the normal operation parameter range and state transition rules of the equipment are extracted from the operation status map of medical equipment, and the node attributes and edge relationships are obtained using the Cypher query language to establish a real-time monitoring indicator system, including the thresholds and normal fluctuation ranges of key parameters such as temperature, pressure, and flow, as well as the state transition sequences of different diagnosis and treatment stages. The equipment operation parameters and status information are obtained in real time, and the collected data is processed through the data preprocessing module. The median filter is used for denoising, and after min-max normalization, the nearest neighbor interpolation method is used to fill in the missing values ​​to generate a standardized equipment operation data stream. The standardized equipment operation data stream is analyzed using the isolation forest algorithm. By comparing with the preset parameter thresholds and state transition rules, it is determined whether the equipment is in an abnormal operation state. If an abnormality is detected, the degree and type of the abnormality are calculated. According to the abnormality detection results, the corresponding level of fault warning information is generated, the abnormal parameters, abnormal time and abnormal type are recorded, and the warning information is distributed to relevant personnel through the message queue system. At the same time, the equipment inspection and maintenance plan generation module is triggered, and the decision tree algorithm is used to select the appropriate maintenance plan according to the type and degree of abnormality, and the corresponding inspection and maintenance plan is formulated. The maintenance plan is executed through the workflow engine, and relevant personnel and resources are arranged to ensure that the equipment is maintained and repaired in time. When implementing the monitoring and early warning of the operation status of medical equipment, the Cypher query language is first used to extract the normal operation parameter range and state transition rules of the equipment from the Neo4j graph database. For example, query "MATCH (n: operation status) RETURN n. parameter name, n. minimum value, n. maximum value" to obtain the normal range of parameters such as temperature, pressure, and flow, and set the temperature threshold to 36.5 ± 0.5 ° C and the pressure threshold to 120 ± 10 mmHg. The state transition rule is obtained through "MATCH (n: operation status) - [r: conversion] -> (m: operation status) RETURN n. name, m. name, r. probability", such as the transition probability from "ready state" to "working state" is 0.95. The equipment operation data is collected at a frequency of 10Hz, denoised by 5-point median filtering, min-max normalization is performed using the formula (x-min) / (max-min), and the missing values ​​are interpolated using the average of the nearest 3 points. The processed data is input into an isolation forest model configured with 100 decision trees and a sampling rate of 0.1 for anomaly detection. If an anomaly is detected, such as the temperature exceeding 37.5℃ for 30 seconds, an early warning message with a level of "severe" is generated. The early warning information is sent to relevant personnel through the RabbitMQ message queue in JSON format {"Device ID":"MED001","Anomaly Type":"Temperature is too high","Anomaly value":38.2,"Time":"2024-05-0110:30:15"}.At the same time, a decision tree algorithm with a depth of 5 and a minimum sample split number of 10 is used to select maintenance plans based on the type and degree of abnormality. For example, temperature abnormalities trigger the "cooling system inspection" process. Finally, the maintenance plan is executed through the Activiti workflow engine, and tasks are automatically assigned to maintenance personnel. The task priority is set to "high" and the maintenance is required to be completed within 2 hours.

[0024] Based on the abnormal operating status that occurs during the medical device log data analysis and mining and status monitoring and early warning process, including data loss and system failure, formulate corresponding equipment inspection and maintenance plans; combine the equipment inspection and maintenance plans with the medical device log data for visualization, so as to monitor the operating status of medical devices in real time and discover and deal with abnormalities.

[0025] Obtain the operation log data of medical equipment, and use the sliding window method to segment the operation log data of medical equipment to obtain the segmented operation log data; for the segmented operation log data, use the cubic spline interpolation algorithm to repair the data loss and obtain the repaired operation log data; establish data integrity evaluation indicators based on the repaired operation log data and the historical operation data of the equipment, and the data integrity evaluation indicators include data missing rate, interpolation accuracy and data consistency; apply the clustering algorithm to the repaired operation log data to obtain the system fault classification results, and build a fault type library based on the system fault classification results; use the ARIMA model to perform time series analysis on the repaired operation log data, extract the equipment operation status characteristics, and set the anomaly detection threshold according to the equipment operation status characteristics; based on the anomaly detected by the anomaly detection threshold, the severity of the anomaly is evaluated in combination with the anomaly duration, deviation degree and impact range factors; based on the anomaly severity score, the fault type in the fault type library and the importance of the equipment, use the decision tree algorithm to prioritize the maintenance tasks, and the decision tree algorithm uses the fault type, equipment importance and maintenance difficulty as the characteristics of the decision tree.

[0026] Specifically, the data collector is used to obtain the operation log data of medical equipment in real time, and the sliding window method is used to process the data in segments. The cubic spline interpolation algorithm is used to repair the data loss problem. Combined with the historical operation data of the equipment, data integrity evaluation indicators are established, including data missing rate, interpolation accuracy and data consistency. Based on the repaired log data, the K-means clustering algorithm is used to classify system faults and build a fault type library. At the same time, the ARIMA model is used for time series analysis to extract the characteristics of the equipment operation status and set the anomaly detection threshold. According to the detected anomalies, the severity of the anomaly is evaluated using a rule-based scoring system, and the anomaly score is generated by considering factors such as the duration of the anomaly, the degree of deviation and the scope of influence. According to the fault type, anomaly score and equipment importance, an equipment inspection and maintenance plan is formulated, and the decision tree algorithm is used to prioritize the maintenance tasks. The fault type, equipment importance and maintenance difficulty are used as the characteristics of the decision tree to generate an inspection and maintenance plan containing maintenance steps, required resources and estimated completion time. The inspection and maintenance plan is combined with real-time log data, and Grafana is used to create an interactive dashboard to display the equipment operation status, abnormal alarms and maintenance progress, and the equipment health status is intuitively presented through color coding and dynamic charts. When an abnormality is detected, an alarm is automatically triggered, and a notification is pushed to relevant personnel through the message queue system. The abnormal status and handling progress are updated in real time on the dashboard, and the entire abnormal handling process is recorded to achieve real-time monitoring and abnormal handling of the operating status of medical devices. When implementing the medical device equipment operating status monitoring and early warning system, the equipment operation log data is first obtained through the data collector at a frequency of 10Hz, and the data is segmented using a 60-second sliding window. For the detected data loss, the cubic spline interpolation algorithm is used for repair, and the interpolation node interval is set to 5 seconds. The data integrity is evaluated by calculating the data missing rate, the threshold is set to 5%, the interpolation accuracy, the R² value is greater than 0.95, and the data consistency, and the deviation of adjacent data points is less than 3 standard deviations. Subsequently, the K-means clustering algorithm is used with the K value set to 5 to classify the system faults, and a fault type library containing categories such as "temperature anomaly" and "pressure fluctuation" is constructed. At the same time, the ARIMA (1,1,1) model is applied to analyze the time series data of the equipment operating status, extract the characteristics such as trend, periodicity and residual, and set the abnormality detection threshold to 3 times the standard deviation. After the anomaly is detected, the rule-based scoring system considers the anomaly duration weight of 0.3, the deviation degree weight of 0.5, and the impact range weight of 0.2 for scoring, with a score range of 0 to 100. Based on the scoring results, the decision tree algorithm is used with a depth of 4 and a minimum number of sample splits of 10 to prioritize the maintenance tasks and generate a maintenance plan with specific steps, required tools, and an estimated completion time of 2 hours.Finally, we created an interactive dashboard that is updated in real time through Grafana, set 5 color codes to indicate the device status, including green for normal, yellow for slight abnormality, orange for moderate abnormality, red for severe abnormality, and gray for offline. We used a line chart to show the 24-hour trend of key parameters, and a bar chart to show the progress of maintenance tasks. When an abnormality is detected, the RabbitMQ message queue pushes the alarm information to the relevant personnel, and updates the abnormal status and handling progress on the dashboard, recording the entire process from abnormality discovery to resolution, realizing all-round monitoring and timely handling of the operating status of medical devices.

[0027] S106. Analyze the actual process of the equipment participating in the diagnosis and treatment activities from the equipment log data, and determine the diagnosis and treatment workflow information by comparing and analyzing it with the standard diagnosis and treatment workflow, including critical paths, bottleneck links and abnormal deviations.

[0028] The timestamp and event information are obtained from the device log data. The event information is obtained by denoising the original log using the median filter algorithm. Based on the denoised event information, the frequent event sequence is extracted to obtain the actual process of the device participating in the diagnosis and treatment activities. For the standard diagnosis and treatment workflow, the standard diagnosis and treatment workflow is converted into a directed acyclic graph structure. The nodes of the directed acyclic graph structure represent the diagnosis and treatment steps, and the edges represent the sequential relationship between the steps. The actual diagnosis and treatment process is sequence-aligned with the standard diagnosis and treatment workflow, and the edit distance between the actual diagnosis and treatment process and the standard diagnosis and treatment workflow is calculated. Based on the edit distance and timestamp information, the decision tree algorithm is used to classify abnormal deviations and identify the bottleneck links on the critical path. The bottleneck links are steps whose execution time exceeds the preset threshold and whose resource utilization is higher than the preset utilization threshold.

[0029] Specifically, timestamps and event information are extracted from device log data, the original log is denoised using a median filter algorithm, continuous events are segmented using a 60-second sliding window, and the GSP (Generalized Sequential Patterns) algorithm is used to extract frequent event sequences to obtain the actual process of the device participating in diagnosis and treatment activities. The standard diagnosis and treatment workflow is formally represented, and the process steps are converted into a directed acyclic graph structure. The nodes represent the diagnosis and treatment steps, and the edges represent the sequential relationship between the steps. The key path in the standard process is extracted using a topological sorting algorithm. The actual process is normalized, and the event sequence is mapped to the steps in the standard process to form an actual process representation consistent with the standard process structure. The Needleman-Wunsch algorithm is used to align the actual diagnosis and treatment process with the standard diagnosis and treatment workflow, and the edit distance of the two processes is calculated to identify abnormal deviations such as missing steps, wrong sequence, and redundant steps. At the same time, the execution time of each step is calculated through timestamp analysis. Based on the comparison results and time analysis, the decision tree algorithm is used to classify abnormal deviations, identify bottleneck links on the critical path, define the steps with execution time exceeding the preset threshold and resource utilization rate higher than 90% as bottleneck links, and calculate the time proportion and deviation frequency of each step. Generate a diagnosis and treatment workflow information report containing critical paths, bottleneck links and abnormal deviations, use BPMN (Business Process Model and Notation) to graphically represent the diagnosis and treatment process, and highlight the critical path, bottleneck links and abnormal deviations through different colors and marks. When implementing the diagnosis and treatment workflow analysis of medical devices, first extract data containing timestamps and event information from the device log, such as "2024-05-01 09:30:15 device startup". The original log is denoised by the 5-point median filter algorithm to remove outliers. A 60-second sliding window with a step size of 30 seconds is used to segment continuous events. Using the GSP algorithm, the minimum support is set to 0.1 and the maximum interval is 5, and frequent event sequences are extracted to obtain the actual diagnosis and treatment process such as "device startup→parameter setting→treatment start→treatment end→data saving". The standard diagnosis and treatment workflow was converted into a directed acyclic graph, containing 10 nodes (diagnosis and treatment steps) and 15 edges (relationships between steps). The Kahn algorithm was used for topological sorting, and the key path was extracted, which contained 6 key steps. The actual process was normalized, and the "parameter setting" was mapped to the "treatment preparation" step in the standard process. The Needleman-Wunsch algorithm was used to set the matching score to 1, the mismatch penalty to -1, and the vacancy penalty to -2, and the actual process and the standard process were sequenced. The edit distance was calculated to be 3, and 1 missing step, 1 sequence error, and 1 redundant step were identified. The average execution time of each step was calculated through timestamp analysis, such as the average execution time of "treatment preparation" was 5 minutes.The CART decision tree algorithm was used to classify abnormal deviations, with the maximum depth set to 5 and the minimum number of sample splits set to 10. Steps with execution time exceeding the preset threshold, such as 10 minutes and resource utilization higher than 90%, were defined as bottlenecks, and two bottleneck steps were identified. Finally, a flowchart was drawn using the BPMN2.0 standard, with critical paths marked in green, bottlenecks marked in red, and abnormal deviations marked in yellow, to generate a diagnosis and treatment workflow analysis report containing detailed statistical information.

[0030] Analyze the actual diagnosis and treatment workflow, determine the critical path of the diagnosis and treatment process, and identify the key activities and nodes that play a decisive role in the entire process. By analyzing the bottlenecks of the actual diagnosis and treatment workflow, determine the bottleneck links that restrict process efficiency, and locate the diagnosis and treatment activities and resource allocation that need to be optimized; compare and analyze the deviations between the actual diagnosis and treatment workflow and the standard process, and identify abnormal deviations that violate the norms or are inefficient.

[0031] Receive workflow information carrying the diagnosis and treatment activity identifier, where the workflow information is generated during the actual diagnosis and treatment process; construct a directed acyclic graph model based on the workflow information, in which nodes represent diagnosis and treatment activities and edges represent dependencies between activities; use the critical path method algorithm to calculate the longest path of the directed acyclic graph model to obtain the key activities and nodes that play a decisive role in the entire process; for each activity node in the directed acyclic graph model, calculate the resource utilization and waiting time, and if the resource utilization exceeds a preset threshold or the waiting time exceeds a preset multiple of the average execution time of the activity, determine the node as a bottleneck link; obtain a standard diagnosis and treatment flowchart, perform structural matching between the directed acyclic graph model and the standard diagnosis and treatment flowchart, determine the similarity of the two flowcharts, and identify structural deviations such as missing nodes, redundant nodes, and incorrect sequences.

[0032] Specifically, the actual diagnosis and treatment workflow is modeled, and the process is represented as a directed acyclic graph. The nodes represent the diagnosis and treatment activities, and the edges represent the dependencies between the activities. The longest path is calculated by the critical path method (CPM) algorithm to determine the key activities and nodes that play a decisive role in the entire process. Based on the actual diagnosis and treatment workflow model, the theory of constraints (TOC) method is used to calculate the resource utilization and waiting time of each activity node. Through statistical analysis of historical data, the resource utilization threshold is set to 85%, and the waiting time threshold is set to 1.5 times the average execution time of the activity. Nodes exceeding the threshold are judged as bottleneck links, and the diagnosis and treatment activities and resource configurations that need to be optimized are located. The actual diagnosis and treatment workflow is matched with the standard process by graph structure, and the similarity of the two flowcharts is calculated by the tree edit distance algorithm. Structural deviations such as missing nodes, redundant nodes, and sequence errors are identified. At the same time, the time and resource attributes of the corresponding nodes are compared to find deviations with abnormal execution efficiency. The severity of the identified deviations is evaluated, and the deviation type weight coefficient is set, such as the weight of missing nodes is 0.8, the weight of sequence errors is 0.6, and the weight of abnormal execution efficiency is 0.4. According to the results of critical path, bottleneck link and abnormal deviation analysis, a diagnosis and treatment process optimization scoring model is constructed, and each diagnosis and treatment activity node is quantitatively scored using the weighted scoring method. The calculation formula is node score = critical path score * 0.4 + bottleneck score * 0.3 + deviation score * 0.3. The standardization score and efficiency score of the overall process are calculated to determine the diagnosis and treatment activities with greater optimization potential. When implementing the optimization analysis of the diagnosis and treatment workflow of medical devices and equipment, the actual diagnosis and treatment process is first modeled as a directed acyclic graph containing 20 nodes, each node representing a diagnosis and treatment activity, such as "patient registration" and "preliminary diagnosis". The critical path method (CPM) algorithm is used to calculate the longest path, and 8 key nodes are identified, with a total critical path duration of 120 minutes. Subsequently, the theoretical constraint method (TOC) is used to analyze the resource utilization and waiting time of each node. By statistically analyzing the historical data of the past 6 months, the resource utilization threshold is set to 85%, and the waiting time threshold is set to 1.5 times the average execution time. The analysis results show that 3 nodes exceed the threshold, among which the resource utilization rate of the "imaging examination" node reaches 92%, and the waiting time is 2.3 times the average execution time, which is determined to be the main bottleneck link. The actual process is compared with the standard process, and the similarity is calculated using the tree edit distance algorithm, with a score of 0.75 and a full score of 1. One missing node "preoperative evaluation", two sequence errors, and three nodes with abnormal execution efficiency are identified. The deviation is evaluated, and the node missing weight is set to 0.8, the sequence error is 0.6, and the execution efficiency is abnormal 0.4. An optimization scoring model is constructed, and the weighted scoring method is used to calculate the score of each node. For example, the "imaging examination" node: the critical path score is 0.9, the bottleneck score is 0.8, the deviation score is 0.6, and the total score is 0.90.4+0.80.3+0.6*0.3=0.78.The standardization score of the overall process was 0.82 and the efficiency score was 0.75, identifying "imaging examination" and "doctor consultation" as the two links with the greatest potential for optimization.

[0033] S107. Integrate the medical device equipment operation status monitoring and warning information, diagnosis and treatment workflow information, and equipment inspection and maintenance plan to form a device failure prediction and safety risk assessment method based on log big data, and apply it to the medical device software system.

[0034] The monitoring and early warning information of the operation status of medical equipment, the diagnosis and treatment workflow information, and the equipment inspection and maintenance plan are integrated to obtain a multidimensional feature vector. Based on the multidimensional feature vector, the ARIMA model is used to perform trend analysis on the historical operation data of the equipment to obtain the equipment failure prediction results. The hierarchical analysis method is used to construct a safety risk assessment index system, in which the equipment failure prediction results, diagnosis and treatment process abnormalities, and maintenance plan execution are used as evaluation factors to construct a judgment matrix and perform a consistency test. The fault prediction model and safety risk assessment method are integrated into the medical device software system to achieve real-time data update and analysis result output.

[0035] Specifically, the Kalman filter algorithm is used to integrate the medical equipment operation status monitoring and early warning information, diagnosis and treatment workflow information, and equipment inspection and maintenance plan. Through data cleaning, standardization, and feature extraction, a multidimensional feature vector containing equipment status, diagnosis and treatment process, and maintenance information is constructed. Based on the fused multidimensional feature vector, the ARIMA model is used to perform trend analysis and periodic detection on the historical operation data of the equipment. Combined with the failure mode and effect analysis (FMEA) method, an equipment failure prediction model is established. The accuracy and reliability of the prediction model are evaluated by ten-fold cross validation and ROC curve analysis, and the AUC value and F1 score of the model are calculated. The hierarchical analysis method (AHP) is used to construct a safety risk assessment index system. The equipment failure prediction results, diagnosis and treatment process abnormalities, and maintenance plan execution are used as evaluation factors. The judgment matrix is ​​constructed and consistency is tested. The safety risk level of the equipment is obtained through indicator weight calculation and comprehensive scoring. The Hadoop distributed file system is used to store log big data, and MapReduce is used for parallel processing and analysis. The fault prediction model and safety risk assessment method are integrated into the medical device software system. Real-time data update and analysis result output are achieved through RESTful API, real-time warning information is pushed through the WebSocket protocol, and the D3.js library is used to design an interactive visualization interface to display the prediction results and risk assessment report. When implementing the fault prediction and safety risk assessment of medical equipment, the Kalman filter algorithm is first used to fuse data from different sources, and the observation noise covariance R is set to 0.1 and the process noise covariance Q is set to 0.01 to achieve the optimal estimation of the data. Through this method, the equipment operation status, diagnosis and treatment process, and maintenance plan information are integrated into a vector containing 50 features. Subsequently, the ARIMA (2,1,2) model is applied to analyze the historical operation data of the equipment and predict the equipment status trend in the next 7 days. Combined with the FMEA method, 5 key failure modes are identified and a fault prediction model is established. The model performance is evaluated using ten-fold cross validation, and the average AUC value is 0.92 and the F1 score is 0.88. The AHP method was used to construct a safety risk assessment index system, which included 3 first-level indicators and 9 second-level indicators. The judgment matrix was constructed using a 9-point scoring method. The consistency ratio CR was 0.04, which was less than 0.1 and passed the consistency test. The weight of the equipment failure prediction result was calculated to be 0.5, the weight of the diagnosis and treatment process abnormality was 0.3, and the weight of the maintenance plan execution was 0.2. The Hadoop cluster was used to store and process about 100GB of log data generated daily, and the mean time between failures (MTBF) and mean time to repair (MTTR) of the equipment were calculated through MapReduce jobs. A RESTful API was designed to update the equipment status data every 5 minutes, and the early warning information was pushed in real time through the WebSocket protocol.Use the D3.js library to create interactive dashboards, including equipment status trend graphs, fault prediction probability distribution graphs, and risk level heat maps, to achieve intuitive presentation and real-time updates of data.

[0036] The description of the above embodiments is only used to help understand the technical solutions and core ideas of the present application. Ordinary technical personnel in this field should understand that they can still modify the technical solutions recorded in the aforementioned embodiments, or replace some of the technical features therein with equivalents. However, these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A device log recording and intelligent analysis method in a medical device software system, characterized in that: The method comprises: Preprocess the operation log data of medical devices and equipment, and uniformly convert the log data from different sources and in different formats to obtain the normalized operation log data of medical devices and equipment; perform semantic understanding and feature extraction on the normalized operation log data of medical devices and equipment, and identify the keywords, terms and time in the log data through word segmentation, part-of-speech tagging and named entity recognition, and mine the semantic features of the log data; combine the extracted semantic features of the log data with the medical device and equipment diagnosis and treatment workflow, and mine the operation status change rules and abnormal patterns of medical devices and equipment at different diagnosis and treatment stages through time series patterns; graphically represent the operation status change rules and abnormal patterns of medical devices and equipment at different diagnosis and treatment stages, and construct a medical device and equipment operation status map to characterize the dynamics of the equipment in the actual diagnosis and treatment process. state behavior characteristics; based on the medical device operation status map, by real-time monitoring of the equipment operation parameters and status changes, determine whether the equipment operation status is abnormal. If the equipment operation status is detected to be abnormal or the parameters exceed the preset threshold range, a fault warning is issued to prompt the equipment to carry out maintenance work and record the medical device operation status monitoring warning information, and start the corresponding equipment maintenance plan; analyze the actual process of the equipment participating in the diagnosis and treatment activities from the equipment log data, and determine the diagnosis and treatment workflow information by comparing and analyzing with the standard diagnosis and treatment workflow, including critical paths, bottleneck links and abnormal deviations; integrate the medical device operation status monitoring warning information, diagnosis and treatment workflow information and equipment maintenance plan to form a device failure prediction and safety risk assessment method based on log big data, and apply it to the medical device software system.

2. The method according to claim 1, characterized in that The medical device operation log data is preprocessed to uniformly convert the log data from different sources and in different formats to obtain the standardized medical device operation log data, including: Receive operation log data sent by medical devices; Perform a quality assessment on the operation log data according to the preset data quality scoring standard to obtain a quality assessment score; If the quality assessment score is higher than the preset threshold, the operation log data is determined to be valid data; Clean the valid data to obtain standardized log data; Align the standardized log data according to a unified time scale to obtain time series data; The time series data is segmented and the statistical features in each time window are calculated to obtain the time series feature vector; Mapping the time series feature vector to a preset general data structure template to obtain feature data in a unified format; Performing dimensionality reduction processing on the feature data in a unified format to obtain feature data after dimensionality reduction; A clustering algorithm is used to group the feature data after dimensionality reduction to obtain the normalized medical device operation log data.

3. The method according to claim 1, characterized in that The method of performing semantic understanding and feature extraction on the normalized medical device operation log data, identifying keywords, terms and time in the log data through word segmentation, part-of-speech tagging and named entity recognition, and mining the semantic features of the log data includes: Obtain standardized medical equipment operation log data; Construct a medical device professional dictionary based on log data. The professional dictionary includes terms extracted from medical device instructions and professional literature. Perform word segmentation on the log data to obtain the log text sequence after word segmentation; Perform part-of-speech tagging on the log text sequence after word segmentation; Hidden Markov models are trained using annotated medical text corpora; The trained hidden Markov model is used to mark the part of speech of each word in the log text sequence, and the log text with part of speech information is obtained; Perform named entity recognition on log text with part-of-speech information; The device name, parameter index and time expression entity in the log text are identified through the conditional random field algorithm; Use annotated medical texts to train the conditional random field to obtain a structured named entity sequence; Based on the structured named entity sequence, the word importance score is calculated, keywords are screened out according to the importance score, and professional terms are identified by combining with the pre-defined term template; Use word embedding methods to convert keywords, professional terms, and time information into numerical vectors; The numerical vectors are concatenated to obtain the semantic feature vector of the log data.

4. The method according to claim 1, characterized in that The extracted semantic features of the log data are combined with the medical device diagnosis and treatment workflow to mine the operation status change rules and abnormal patterns of the medical device at different diagnosis and treatment stages through time series patterns, including: receiving a semantic feature vector carrying a unique identification number of a medical device, the semantic feature vector being generated by different devices at different time periods; According to the semantic feature vector, a dynamic time warping algorithm is used to align it to a unified time axis to obtain aligned time series semantic feature data; The Hidden Markov Model is used to divide the diagnosis and treatment stages of the aligned time series semantic feature data and calculate the optimal state sequence; For the time series data divided into diagnosis and treatment stages, the sliding window method is used to extract local features, and the statistical features and frequency domain features are calculated in each window to obtain the feature sequence of each diagnosis and treatment stage; Perform cluster analysis on the characteristic sequences of each diagnosis and treatment stage, determine the typical operating status of the equipment at different diagnosis and treatment stages, and explore the law of status change; An abnormal threshold is set. If the Mahalanobis distance between the sample and the cluster center exceeds the abnormal threshold, it is identified as an abnormal pattern, and the cause of the abnormality is inferred based on the decision tree algorithm.

5. The method according to claim 1, characterized in that The above-mentioned method of graphically representing the operation status change rules and abnormal patterns of medical devices in different diagnosis and treatment stages and constructing an operation status map of medical devices to characterize the dynamic behavior characteristics of the devices in the actual diagnosis and treatment process includes: Obtain the changing rules and abnormal patterns of the operating status of medical devices at different stages of diagnosis and treatment; The nodes and edges of the graph are constructed according to the change rules and abnormal patterns, where the nodes include the diagnosis and treatment stage, operation status and abnormal pattern, and the node attributes include duration, parameter value range and occurrence frequency; The edge represents the transition relationship between states, and the edge weight represents the transition probability; For the constructed graph, the Node2Vec algorithm is used for vector representation, encoding the attribute information of nodes and edges into a multi-dimensional vector space; The semantic similarity between nodes is calculated by cosine similarity. If the semantic similarity is higher than the preset threshold, the nodes are merged to obtain a simplified graph structure. For the simplified graph structure, a graph convolutional neural network is used to aggregate and propagate node features. Each layer of the graph convolutional neural network contains multiple neurons. The neural network parameters are trained through the back-propagation algorithm to obtain the updated node feature representation.

6. The method according to claim 1, characterized in that The medical device operation status map is based on real-time monitoring of the device operation parameters and status changes to determine whether the device is in an abnormal operation state. If the device operation state is abnormal or the parameters exceed the preset threshold range, a fault warning is issued to prompt the device to carry out maintenance work and record the medical device operation status monitoring warning information, and start the corresponding equipment maintenance plan, including: Obtain the normal operating parameter range and state transition rules of the equipment from the operation state map of the medical device. The state transition rules include the state transition sequence of different diagnosis and treatment stages; Obtain the real-time operating parameters and status information of the equipment, perform data preprocessing on the operating parameters and status information, and obtain standardized equipment operation data stream; The isolation forest algorithm is used to analyze the standardized equipment operation data stream, and by comparing it with the preset parameter thresholds and state transition rules, it is determined whether the equipment is in an abnormal operating state; Generate fault warning information of corresponding level according to abnormal operation status, including abnormal parameters, abnormal time and abnormal type; It also includes: formulating corresponding equipment inspection and maintenance plans based on the abnormal operating status that occurs during the medical device log data analysis and mining and status monitoring and early warning, including data loss and system failure; combining the equipment inspection and maintenance plan with the medical device log data for visual representation, so as to monitor the operating status of the medical device in real time and discover and deal with abnormalities.

7. The method according to claim 6, characterized in that The abnormal operation status that occurs during the analysis and mining of medical device log data and status monitoring and early warning, including data loss and system failure, formulates a corresponding equipment inspection and maintenance plan; combines the equipment inspection and maintenance plan with the medical device log data for visualization, so as to monitor the operation status of the medical device in real time, discover and deal with abnormalities, including: Obtaining the operation log data of medical equipment, and using the sliding window method to segment the operation log data of medical equipment to obtain the segmented operation log data; using the cubic spline interpolation algorithm to repair the data loss of the segmented operation log data to obtain the repaired operation log data; establishing data integrity evaluation indicators based on the repaired operation log data and the historical operation data of the equipment, the data integrity evaluation indicators include data missing rate, interpolation accuracy and data consistency; Apply clustering algorithm to the repaired operation log data to obtain the system fault classification results, and build a fault type library based on the system fault classification results; The ARIMA model is used to perform time series analysis on the operation log data after repair, extract the equipment operation status characteristics, and set the anomaly detection threshold according to the equipment operation status characteristics; based on the anomaly detected by the anomaly detection threshold, the severity of the anomaly is evaluated in combination with the anomaly duration, deviation degree and impact range factors; According to the anomaly severity score, the fault type in the fault type library and the equipment importance, the decision tree algorithm is used to prioritize the maintenance tasks. The decision tree algorithm uses the fault type, equipment importance and maintenance difficulty as the features of the decision tree.

8. The method according to claim 1, characterized in that The actual process of the equipment participating in the diagnosis and treatment activities is analyzed from the equipment log data, and the diagnosis and treatment workflow information is determined by comparing and analyzing with the standard diagnosis and treatment workflow, including key paths, bottleneck links and abnormal deviations, including: Obtain timestamps and event information from device log data. Event information is obtained by denoising the original log using a median filter algorithm. Based on the event information after denoising, the frequent event sequence is extracted to obtain the actual process of the equipment participating in the diagnosis and treatment activities; For the standard diagnosis and treatment workflow, the standard diagnosis and treatment workflow is converted into a directed acyclic graph structure, where the nodes of the directed acyclic graph represent the diagnosis and treatment steps, and the edges represent the sequential relationship between the steps; Perform sequence alignment on the actual diagnosis and treatment process and the standard diagnosis and treatment workflow, and calculate the edit distance between the actual diagnosis and treatment process and the standard diagnosis and treatment workflow; According to the edit distance and timestamp information, a decision tree algorithm is used to classify abnormal deviations and identify bottleneck links on the critical path. The bottleneck link is a step whose execution time exceeds a preset threshold and whose resource utilization rate is higher than a preset utilization rate threshold. It also includes: analyzing the actual diagnosis and treatment workflow, determining the critical path of the diagnosis and treatment process, identifying key activities and nodes that play a decisive role in the entire process, and determining the bottleneck links that restrict the efficiency of the process by analyzing the bottleneck of the actual diagnosis and treatment workflow, and locating the diagnosis and treatment activities and resource allocation that need to be optimized; comparing and analyzing the deviations between the actual diagnosis and treatment workflow and the standard process, and identifying abnormal deviations that violate specifications or are inefficient.

9. The method according to claim 8, characterized in that The analysis of the actual diagnosis and treatment workflow determines the critical path of the diagnosis and treatment workflow, identifies the key activities and nodes that play a decisive role in the entire process, analyzes the bottlenecks of the actual diagnosis and treatment workflow, determines the bottleneck links that restrict the efficiency of the process, and locates the diagnosis and treatment activities and resource allocation that need to be optimized; compares and analyzes the deviations between the actual diagnosis and treatment workflow and the standard process, identifies abnormal deviations that violate the norms or are inefficient, including: Receive workflow information carrying a diagnosis and treatment activity identifier, where the workflow information is generated during the actual diagnosis and treatment process; construct a directed acyclic graph model based on the workflow information, where nodes represent diagnosis and treatment activities and edges represent dependencies between activities; The critical path method algorithm is used to calculate the longest path of the directed acyclic graph model to obtain the key activities and nodes that play a decisive role in the entire process; For each active node in the directed acyclic graph model, the resource utilization and waiting time are calculated. If the resource utilization exceeds the preset threshold or the waiting time exceeds the preset multiple of the average execution time of the activity, the node is determined to be a bottleneck link. The standard diagnosis and treatment flowchart is obtained, and the directed acyclic graph model is structurally matched with the standard diagnosis and treatment flowchart to determine the similarity of the two flowcharts, and identify structural deviations such as missing nodes, redundant nodes, and incorrect sequences.

10. The method according to claim 1, characterized in that The medical device operation status monitoring and early warning information, diagnosis and treatment workflow information, and equipment inspection and maintenance plan are integrated to form a device failure prediction and safety risk assessment method based on log big data, and applied to the medical device software system, including: Integrate the medical equipment operation status monitoring and early warning information, diagnosis and treatment workflow information, and equipment inspection and maintenance plan to obtain multi-dimensional feature vectors; According to the multi-dimensional feature vector, the ARIMA model is used to analyze the trend of the equipment historical operation data to obtain the equipment failure prediction results; The analytic hierarchy process was used to construct a safety risk assessment index system, in which equipment failure prediction results, abnormal diagnosis and treatment process, and maintenance plan execution were used as evaluation factors to construct a judgment matrix and conduct consistency tests. Integrate fault prediction models and safety risk assessment methods into medical device software systems to achieve real-time data updates and analysis result outputs.

Images