Medical equipment state monitoring method based on artificial intelligence

Abstract

The invention provides a medical equipment state monitoring method based on artificial intelligence. The method is applied to the technical field of medical equipment state monitoring, and comprises the following steps: performing synchronous acquisition through a multi-source sensor network to form an original data set; performing data cleaning and standardization processing on the original data set to generate a preprocessed structured data matrix; inputting the structured data matrix into the hybrid model, extracting a multi-dimensional time sequence feature vector, and outputting a feature set containing equipment health degree scores; predicting the state category of the medical equipment according to the feature set, generating a preliminary judgment result, and calculating a corresponding confidence probability; and performing judgment according to the preliminary judgment result and the corresponding confidence coefficient probability, if the confidence coefficient is lower than a preset threshold value, starting an artificial rechecking process, otherwise, directly generating an equipment maintenance instruction and transmitting the equipment maintenance instruction to a terminal execution unit. In this way, the monitoring accuracy can be improved.

Description

Technical Field

[0001] The present disclosure relates to the technical field of medical device status monitoring, and in particular to a medical device status monitoring method based on artificial intelligence. Background Art

[0002] With the continuous advancement of medical technology, various high-end medical devices play a vital role in clinical diagnosis, treatment, and surgery. However, medical device failures, anomalies, or improper operation can directly impact patient health and even lead to immeasurable consequences. Real-time monitoring of equipment using artificial intelligence (AI) can effectively improve device reliability and safety. By analyzing device operating status data, such as temperature, pressure, and current, potential failures can be predicted, enabling proactive maintenance or component replacement to avoid medical accidents caused by equipment failure. Through machine learning, data analysis, and predictive algorithms, abnormal device behavior and potential failures can be identified, triggering early warnings and prompting personnel to conduct inspections and repairs, significantly reducing the occurrence of medical accidents and improving patient safety. Real-time monitoring of device status can prevent equipment failures, reduce unnecessary repairs and downtime, and avoid the high costs associated with equipment problems. AI can also analyze device usage, optimize usage frequency and maintenance cycles, and further reduce operating costs for healthcare institutions. Therefore, AI-based monitoring of medical device status is of great significance.

[0003] Currently, existing technologies for medical device status monitoring primarily rely on single-sensor data collection and threshold alarm mechanisms. These technical limitations have become increasingly apparent. Traditional monitoring systems often employ single-dimensional signal analysis, making it difficult to reflect the comprehensive status of the device under complex operating conditions. Medical device operation involves the coupling of multiple physical fields, such as electromechanical coordination and thermodynamic changes. Relying solely on single sensor data can easily lead to the loss of key features, making it difficult to accurately identify early wear or intermittent failures. At the state recognition algorithm level, traditional methods commonly suffer from the fusion of spatiotemporal features. When using models such as convolutional neural networks (CNNs) or support vector machines (SVMs), CNNs can extract local spatial features but struggle to capture the long-term temporal patterns of device degradation. While LSTM networks excel at temporal modeling, they lack the ability to represent spatial features such as mechanical vibration waveforms. Furthermore, classification decision mechanisms based on empirical thresholds lack probabilistic assessments. When the device is in a critical state between normal and faulty, existing systems are prone to misjudging the device as "false normal" or "over-warning," resulting in low accuracy in monitoring the medical device status. Summary of the Invention

[0004] The present disclosure provides a medical device status monitoring method based on artificial intelligence. The method includes:

[0005] S1, obtains real-time operating data of medical equipment, including current signals and temperature parameters, through synchronous collection of multi-source sensor networks to form raw data sets;

[0006] S2, performing data cleaning and standardization processing on the original data set, including noise filtering, missing value interpolation and data normalization, to generate a preprocessed structured data matrix;

[0007] S3 inputs the structured data matrix into a hybrid model that integrates a convolutional neural network and a long short-term memory network, extracts multidimensional time series feature vectors, and outputs a feature set containing the device health score;

[0008] S4, predicting the medical device status category using a random forest classifier based on the feature set, generating a preliminary determination result including three levels of labels: normal, warning, and fault, and calculating the corresponding confidence probabilities;

[0009] S5, making a determination based on the preliminary determination result and the corresponding confidence probability. If the confidence is lower than a preset threshold, a manual review process is initiated; otherwise, a device maintenance instruction is directly generated and transmitted to the terminal execution unit.

[0010] Furthermore, S3 includes:

[0011] S31, dividing the structured data matrix into a plurality of subsequence segments through a sliding time window, each segment covering a continuous sampling period of 5 minutes;

[0012] S32, extracting spatial local features using parallel convolution layers according to the subsequence segments, and generating a convolution feature map including frequency domain energy distribution;

[0013] S33, inputting a bidirectional LSTM network into the convolutional feature map to capture temporal dependencies and outputting a hidden state vector with context association;

[0014] S34, performing weighted fusion of key features through a self-attention mechanism according to the hidden state vector to form an attention-enhanced feature representation;

[0015] S35 , mapping the attention enhancement feature to a health score space through a fully connected layer, and outputting a device health index in the range of 0 to 1.

[0016] Further:

[0017] If the health score is lower than J1 and the current harmonic distortion rate exceeds b, it is determined to be a level 3 fault state;

[0018] If the health score is between J1 and J2 and the temperature rise rate is higher than w, it is determined to be a level 2 warning state;

[0019] If the health score is higher than J2, it is determined to be in a level 1 normal state; where J1 is the first health score threshold, J2 is the second health score threshold, J1<J2, b is the current harmonic distortion rate threshold, and w is the temperature rise rate threshold.

[0020] Further:

[0021] If a single current surge occurs in the first-level normal state and exceeds the rated value k1, a reassessment is performed, where k1 is the first percentage threshold;

[0022] If the temperature rise rate exceeds the threshold value k2 in three consecutive detection cycles in the second-level warning state, it is determined to be a third-level fault state; k2 is the second percentage threshold.

[0023] Furthermore, S4 includes:

[0024] S41, inputting the health score and the 12-dimensional time series features into a random forest model according to the feature set;

[0025] S42, based on the model output of S41, calculate the probability distribution of each state category and the confidence of the decision path;

[0026] S43, dynamically adjusting the classification threshold boundary using a Bayesian update rule according to the confidence level;

[0027] S44, starting a secondary verification process for data points in the classification boundary area according to the adjusted threshold value of S43;

[0028] S45 , generating a final status report including a timestamp and a fault code based on the verification result of S44 .

[0029] Further:

[0030] If the matching degree between the feature vector and the historical fault library in the secondary verification process exceeds k3, the fault classification is confirmed; k3 is the third percentage threshold;

[0031] If the LSTM hidden state entropy value of the boundary area data point is higher than Sz, it is judged as an unknown anomaly type; Sz is the entropy value threshold;

[0032] If the verification results are inconsistent for two consecutive times, S1 to S3 are executed again.

[0033] Further:

[0034] For unknown abnormal type data, if its Mahalanobis distance in the feature space exceeds the 3σ range, it is marked as a high-risk event;

[0035] For high-risk events, if they occur more than g times within time t, a new fault category label is automatically created; t is the time threshold, g is the number threshold, and σ is the standard deviation threshold;

[0036] For data that still has classification ambiguity after sensor calibration, it is forced to enter the manual diagnosis queue and the automatic decision-making is suspended.

[0037] Furthermore, S5 includes:

[0038] S51, based on the status report of S4, generates a JSON instruction set containing the device ID, fault code and recommended measures;

[0039] S52, calling a three-dimensional visualization engine to generate a device health status heat map according to the instruction set;

[0040] S53, based on the heat map, push key alarm information to the mobile terminal and write it into the blockchain for evidence storage;

[0041] S54, based on the evidence records in S53, automatically generates a preventive maintenance work order and assigns it to the nearest service station;

[0042] S55: Update the fault resolution case library in the equipment knowledge graph based on the maintenance feedback data of S54.

[0043] Further:

[0044] If the same medical device triggers the same fault code more than three times within 24 hours, the spare parts replacement priority escalation procedure will be initiated;

[0045] If the maintenance work order response time exceeds 4 hours and is not confirmed, it will be automatically transferred to the higher-level management center;

[0046] If similar failure modes occur in related devices after the knowledge graph is updated, group warning notifications will be sent in batches.

[0047] The present invention uses a multi-source sensor network to synchronously collect multi-dimensional raw data such as current signals, temperature parameters, mechanical vibration waveforms and equipment log information to ensure that the operating status of medical equipment is fully captured. This multi-dimensional data collection method improves the accuracy and reliability of monitoring; a random forest classifier is used to classify the equipment status, which can generate three-level labels such as normal, warning and fault, and provide confidence probability based on the classification results. This multi-category judgment method can not only perform a detailed analysis of the health status of the equipment, but also provide a reliable basis for maintenance decisions, thereby improving the accuracy of monitoring.

[0048] It should be understood that the contents described in the Summary of the Invention section are not intended to limit the key or important features of the embodiments of the present disclosure, nor are they intended to limit the scope of the present disclosure. Other features of the present disclosure will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The above and other features, advantages, and aspects of the various embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. The accompanying drawings are provided for a better understanding of the present disclosure and do not constitute a limitation of the present disclosure. In the accompanying drawings, the same or similar reference numerals represent the same or similar elements, wherein:

[0050] Figure 1 A flowchart of a medical device status monitoring method based on artificial intelligence according to an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0051] To make the purpose, technical solutions, and advantages of the embodiments of the present disclosure more clear, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present disclosure, not all of the embodiments. Based on the embodiments of the present disclosure, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present disclosure.

[0052] In this document, the term "and / or" simply describes a relationship between related objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Furthermore, the character " / " in this document generally indicates that the related objects are in an "or" relationship.

[0053] Figure 1 A flowchart of a medical device status monitoring method based on artificial intelligence according to an embodiment of the present disclosure is shown, the method comprising:

[0054] S1, obtains real-time operating data of medical equipment, including current signals and temperature parameters, through synchronous collection of multi-source sensor networks to form raw data sets;

[0055] S2, performing data cleaning and standardization processing on the original data set, including noise filtering, missing value interpolation and data normalization, to generate a preprocessed structured data matrix;

[0056] S3 inputs the structured data matrix into a hybrid model that integrates a convolutional neural network and a long short-term memory network, extracts multidimensional time series feature vectors, and outputs a feature set containing the device health score;

[0057] S4, predicting the medical device status category using a random forest classifier based on the feature set, generating a preliminary determination result including three levels of labels: normal, warning, and fault, and calculating the corresponding confidence probabilities;

[0058] S5, making a determination based on the preliminary determination result and the corresponding confidence probability. If the confidence is lower than a preset threshold, a manual review process is initiated; otherwise, a device maintenance instruction is directly generated and transmitted to the terminal execution unit.

[0059] According to an embodiment of the present invention, a multi-source sensor network is used to synchronously collect multi-dimensional raw data of current signals and temperature parameters to ensure that the operating status of medical equipment is fully captured. This multi-dimensional data collection method improves the accuracy and reliability of monitoring. A random forest classifier is used to classify the equipment status, which can generate three-level labels such as normal, warning and fault, and provide confidence probabilities based on the classification results. This multi-category judgment method can not only perform a detailed analysis of the health status of the equipment, but also provide a reliable basis for maintenance decisions, thereby improving the accuracy of monitoring.

[0060] In some embodiments, S3 includes: S31, according to the structured data matrix, dividing it into multiple subsequence segments through a sliding time window, each segment covering a continuous sampling period of 5 minutes; S32, according to the subsequence segments, using a parallel convolution layer to extract spatial local features, and generating a convolution feature map containing frequency domain energy distribution; S33, according to the convolution feature map, inputting a bidirectional LSTM network to capture temporal dependencies, and outputting a hidden state vector with context association; S34, according to the hidden state vector, performing weighted fusion of key features through a self-attention mechanism to form an attention-enhanced feature representation; S35, according to the attention-enhanced feature, mapping it to the health score space through a fully connected layer, and outputting a device health indicator in the range of 0 to 1. According to an embodiment of the present invention, by converting the original data matrix into multiple subsequence fragments and using a sliding time window for segmentation, the local features in the time series can be effectively captured, helping the model to understand the data from a more fine-grained perspective, avoiding the problem of data loss and improving the accuracy of the model; the use of parallel convolutional layers can simultaneously process multiple different spatial features, significantly improving computational efficiency, and helping to better capture the characteristics of complex signals such as frequency domain energy distribution, thereby improving the performance of the model in multidimensional data analysis; combining multiple technologies such as convolutional layers, LSTM, bidirectional networks and self-attention mechanisms, it can fully extract and fuse spatiotemporal features, and improve the accuracy and stability of equipment health assessment.

[0061] For example, S31: Set a 5-minute sliding time window to move on the data matrix, each time moving a 1-minute step, so that each subsequence segment contains continuous 5-minute data; extract continuous 5-minute segments from the original data matrix to generate a sequence list. Each subsequence fragment will serve as the input of the subsequent steps; S32: Design multiple convolution layers and run them in parallel. These convolution layers use different convolution kernel sizes to capture different spatial features of the data. For example, convolution kernels of different sizes such as 3×3 and 5×5 can be used; in addition to directly applying spatial convolution, the fast Fourier transform (FFT) can also be used to convert the time domain signal into the frequency domain to extract the energy distribution in the frequency domain; multiple feature maps are generated through convolution operations, each feature map contains the expression of different spatial features, and the expressive ability of data features is further enhanced by modeling the frequency domain; S33: Construct a bidirectional LSTM model, which contains two LSTM networks, one processes the forward dependency of the sequence (from front to back), and the other processes the reverse dependency (from back to front); the convolution feature map generated by S32 is passed into the LSTM network as input; LSTM outputs the processed data through the hidden state (hidden layer state), and each time step generates a context-associated hidden state vector, which contains the current time step and The correlation information of the previous and next time steps; S34: The self-attention mechanism is used to weight the hidden state of the LSTM output. The self-attention mechanism calculates the weight value of each time step, which represents the degree of attention of the time step to other time steps; by calculating the Query, Key, and Value matrices, the attention weight of each time step is obtained, and the weighted sum of the hidden state of the LSTM output is performed; the over-weighted feature vectors are synthesized into an enhanced feature representation, which highlights the important time step information and suppresses the irrelevant time step information; S35:; Design one or more layers of fully connected neural networks, and input the attention enhancement features generated by S34 into this layer. The number of nodes in the fully connected layer should be set according to actual needs; Use the Sigmoid activation function at the end of the fully connected layer to compress the output to the range of 0 to 1 to represent the health score of the device; The output of the fully connected layer will represent the health of the device, with a score range of 0 (very poor) to 1 (excellent), and this output can be used as the basis for device status monitoring and prediction.

[0062] In some embodiments, if the health score is lower than J1 and the current harmonic distortion rate exceeds b, it is determined to be a level 3 fault state; if the health score is between J1 and J2 and the temperature rise rate is higher than w, it is determined to be a level 2 warning state; if the health score is higher than J2, it is determined to be a level 1 normal state; wherein J1 is the first health score threshold, J2 is the second health score threshold, J1<J2, b is the current harmonic distortion rate threshold, and w is the temperature rise rate threshold. According to an embodiment of the present invention, by setting multiple health score thresholds and combining them with current harmonics and temperature, it is possible to accurately distinguish different levels of fault states, take maintenance or shutdown measures in advance, and reduce losses caused by faults; the health score thresholds and the current harmonic and temperature rise rate thresholds can be flexibly adjusted according to the actual working conditions and needs of different devices, making the solution highly adaptable and able to optimize monitoring and warning strategies according to specific devices and working environments; based on the different fault and warning state classifications, the specific type of equipment problem can be effectively diagnosed, helping technicians quickly locate the problem and develop a repair plan.

[0063] For example, if the health score is lower than 0.3 and the current harmonic distortion rate exceeds 15%, it is determined to be a level 3 fault state; if the health score is between 0.3 and 0.6 and the temperature rise rate is higher than 2°C / min, it is determined to be a level 2 warning state; if the health score is higher than 0.6, it is determined to be a level 1 normal state.

[0064] In some embodiments, if the temperature rise rate exceeds the threshold k2 within three consecutive detection cycles in the second-level warning state, it is determined to be a third-level fault state, and k2 is the second percentage threshold; if a single current surge occurs in the first-level normal state and exceeds the rated value k1, a re-evaluation is performed, and k1 is the first percentage threshold.

[0065] For example, if the temperature rise rate exceeds 120% within three consecutive detection cycles in the second-level warning state, it is determined to be a third-level fault state; if a single current surge exceeds 200% of the rated value in the first-level normal state, a reassessment is performed.

[0066] In some embodiments, S4 includes: S41, inputting the health score and 12-dimensional time series features into the random forest model based on the feature set; S42, calculating the probability distribution of each state category and the confidence of the decision path based on the model output of S41; S43, dynamically adjusting the classification threshold boundary using the Bayesian update rule based on the confidence; S44, initiating a secondary verification process for data points in the classification boundary area based on the adjusted threshold of S43; S45, generating a final status report containing a timestamp and a fault code based on the verification result of S44. According to an embodiment of the present invention, the model dynamically adjusts the classification threshold based on the confidence using the Bayesian update rule, which enables the model to flexibly adjust the classification criteria based on newly obtained information and feedback to cope with changes in data. The Bayesian update rule provides an update mechanism based on prior and posterior probabilities, enabling the model to self-optimize in a constantly changing environment.

[0067] In some embodiments, if the matching degree between the feature vector and the historical fault library in the secondary verification process exceeds k3, the fault classification is confirmed; k3 is the third percentage threshold; if the LSTM hidden state entropy value of the boundary area data point is higher than Sz, it is determined to be an unknown abnormality type; Sz is the entropy value threshold; if the verification results of two consecutive times are inconsistent, S1 to S3 are re-executed.

[0068] For example, if the matching degree between the feature vector and the historical fault library in the secondary verification process exceeds 85%, the fault classification is confirmed; if the LSTM hidden state entropy value of the boundary area data point is higher than 4.2, it is determined to be an unknown anomaly type.

[0069] In some embodiments, for unknown abnormal type data, if its Mahalanobis distance in the feature space exceeds the range of 3σ, it is marked as a high-risk event; for a high-risk event, if it is repeated more than g times within time t, a new fault category label is automatically created; t is the time threshold, g is the number threshold, and σ is the standard deviation threshold; for data that still has classification ambiguity after sensor calibration, it is forced to enter the manual diagnosis queue and automatic decision-making is suspended.

[0070] For example, for unknown abnormal type data, if its Mahalanobis distance in the feature space exceeds the 3σ range, it will be marked as a high-risk event; for high-risk events, if they occur more than three times within 72 hours, a new fault category label will be automatically created; for data that still has classification ambiguity after sensor calibration, it will be forced to enter the manual diagnosis queue and automatic decision-making will be suspended.

[0071] In some embodiments, S5 includes: S51, generating a JSON instruction set containing device ID, fault code and recommended measures based on the status report of S4; S52, calling a three-dimensional visualization engine to generate a device health status heat map based on the instruction set; S53, pushing key alarm information to a mobile terminal and writing it into blockchain evidence based on the heat map; S54, automatically generating a preventive maintenance work order based on the evidence record of S53 and assigning it to the nearest service site; S55, updating the fault resolution case library in the device knowledge graph based on the maintenance feedback data of S54.

[0072] For example, S51 extracts key information from the status report of S4, such as device ID, fault code, alarm information, and recommended treatment measures; structures the data into JSON format as required. The structure example is as follows:

[0073] ```json

[0074] {

[0075] "device_id":"12345",

[0076] "fault_code":"F123",

[0077] "suggested_actions":[

[0078] "Check power supply",

[0079] "Verify connectionintegrity",

[0080] "Inspect cooling system" ]

[0082] }

[0083] Organize the fault codes and recommended measures and generate a JSON instruction set through preset rules. Each fault code will be mapped to different diagnostic steps and recommended measures; the generated JSON instruction set is stored in a database or message queue and is ready to be pushed to the subsequent processing module; S52: Get the JSON instruction set containing the device ID, fault code and related status information from S51; convert the fault information of the device into a visual numerical form, and generate the color and numerical range required for the heat map according to the severity of the fault; use a three-dimensional visualization engine (such as Unity3D, Three.js, etc.) to generate a three-dimensional model of the device and its health status. The model shows the status of each device component and displays the heat map through color or other visual effects; according to the status of each part of the device, the health level of different components is displayed through a heat map, for example, red indicates a high probability of failure, yellow indicates a medium probability of failure, and green indicates good; the heat map is presented in the form of a graphical interface, and users can view it on the visualization platform to understand the health status of the device; S53: extract key alarm information from the heat map, such as the device fault component, fault type, fault severity, etc. Severity, etc.; push alarm information to the mobile terminal of the administrator or maintenance personnel through mobile push services (such as APNs, Firebase push, etc.); the push content includes device ID, fault description, alarm level, recommended measures, etc., and can be pushed asynchronously through message queues (such as Kafka); use blockchain technology to store fault information to ensure that the data cannot be tampered with. The blockchain can store fault reports, timestamps, related equipment information, etc.; write data to the blockchain in the form of smart contracts to form a traceable fault record. The smart contract can automatically trigger other operations, such as generating work orders, notifying maintenance personnel, etc.; S54: obtain fault records and related alarm information from the blockchain to ensure that the information is up to date and reliable; predefine work order templates, including fields such as device ID, fault type, handling measures, priority, etc.; automatically fill in the work order content based on the fault information in the evidence record, for example, generate a maintenance plan, estimated working hours, required parts, etc. based on the fault code; assign maintenance work orders based on the device's geographic location or the nearest service station. Use the service site's geographic information system (GIS) for distance matching; assign work orders to the corresponding service site through a scheduling system (such as an automated workflow engine) and notify relevant personnel; S55 collects maintenance data feedback from service sites, including processing methods, repair time, replacement parts, etc.; extracts effective fault solutions and lessons learned from maintenance feedback to determine which solutions are effective and which are ineffective; classifies new fault resolution cases according to equipment, fault type, solution, etc., and updates them to the equipment knowledge graph. Graph databases (such as Neo4j) can be used to manage the relationship between equipment faults and solutions; use machine learning algorithms to analyze historical data to discover possible failure modes and solution strategies, thereby optimizing equipment maintenance processes and improving fault diagnosis accuracy.

[0084] In some embodiments, if the same medical device triggers the same fault code more than three times within 24 hours, the spare parts replacement priority upgrade program will be initiated; if the maintenance work order response time exceeds 4 hours without confirmation of receipt, it will be automatically transferred to the superior management center; if similar failure modes occur in related devices after the knowledge graph is updated, group warning notifications will be sent in batches.

[0085] It should be noted that for the aforementioned method embodiments, for simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the present disclosure is not limited by the order of the actions described, because according to the present disclosure, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily required by the present disclosure.

[0086] It should be understood that the various forms of the above-mentioned processes can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions of this disclosure can be achieved. This is not limited herein.

[0087] The above specific embodiments do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure shall be included within the scope of protection of this disclosure.

Claims

1. A medical equipment status monitoring method based on artificial intelligence, characterized in that: include: S1, obtains real-time operating data of medical equipment, including current signals and temperature parameters, through synchronous collection of multi-source sensor networks to form raw data sets; S2, performing data cleaning and standardization processing on the original data set, including noise filtering, missing value interpolation and data normalization, to generate a preprocessed structured data matrix; S3 inputs the structured data matrix into a hybrid model that integrates a convolutional neural network and a long short-term memory network, extracts multidimensional time series feature vectors, and outputs a feature set containing the device health score; S4, predicting the medical device status category using a random forest classifier based on the feature set, generating a preliminary determination result including three levels of labels: normal, warning, and fault, and calculating the corresponding confidence probabilities; S5, making a determination based on the preliminary determination result and the corresponding confidence probability. If the confidence is lower than a preset threshold, a manual review process is initiated; otherwise, a device maintenance instruction is directly generated and transmitted to the terminal execution unit.

2. The medical device status monitoring method based on artificial intelligence according to claim 1, characterized in that S3 include: S31, dividing the structured data matrix into a plurality of subsequence segments through a sliding time window, each segment covering a continuous sampling period of 5 minutes; S32, extracting spatial local features using parallel convolution layers according to the subsequence segments, and generating a convolution feature map including frequency domain energy distribution; S33, inputting a bidirectional LSTM network into the convolutional feature map to capture temporal dependencies and outputting a hidden state vector with context association; S34, performing weighted fusion of key features through a self-attention mechanism according to the hidden state vector to form an attention-enhanced feature representation; S35 , mapping the attention enhancement feature to a health score space through a fully connected layer, and outputting a device health index in the range of 0 to 1.

3. The artificial intelligence-based medical device status monitoring method according to claim 2, characterized in that: If the health score is lower than J1 and the current harmonic distortion rate exceeds b, it is determined to be a level 3 fault state; If the health score is between J1 and J2 and the temperature rise rate is higher than w, it is determined to be a level 2 warning state; If the health score is higher than J2, it is determined to be in a level 1 normal state; where J1 is the first health score threshold, J2 is the second health score threshold, J1<J2, b is the current harmonic distortion rate threshold, and w is the temperature rise rate threshold.

4. The artificial intelligence-based medical device status monitoring method according to claim 3, characterized in that: If a single current surge occurs in the first-level normal state and exceeds the rated value k1, a reassessment is performed, where k1 is the first percentage threshold; If the temperature rise rate exceeds the threshold value k2 for three consecutive detection cycles in the second-level warning state, it is determined to be a third-level fault state; k2 is the second percentage threshold.

5. The medical device status monitoring method based on artificial intelligence according to claim 1, characterized in that S4 include: S41, inputting the health score and the 12-dimensional time series features into a random forest model according to the feature set; S42, based on the model output of S41, calculate the probability distribution of each state category and the confidence of the decision path; S43, dynamically adjusting the classification threshold boundary using a Bayesian update rule according to the confidence level; S44, starting a secondary verification process for data points in the classification boundary area according to the adjusted threshold value of S43; S45 , generating a final status report including a timestamp and a fault code based on the verification result of S44 .

6. The artificial intelligence-based medical device status monitoring method according to claim 5, characterized in that: If the matching degree between the feature vector and the historical fault library in the secondary verification process exceeds k3, the fault classification is confirmed; k3 is the third percentage threshold; If the LSTM hidden state entropy value of the boundary area data point is higher than Sz, it is judged as an unknown anomaly type; Sz is the entropy value threshold; If the verification results are inconsistent for two consecutive times, S1 to S3 are executed again.

7. The artificial intelligence-based medical device status monitoring method according to claim 6, characterized in that: For unknown abnormal type data, if its Mahalanobis distance in the feature space exceeds the 3σ range, it is marked as a high-risk event; For high-risk events, if they occur more than g times within time t, a new fault category label is automatically created; t is the time threshold, g is the number threshold, and σ is the standard deviation threshold; For data that still has classification ambiguity after sensor calibration, it is forced to enter the manual diagnosis queue and the automatic decision-making is suspended.

8. The medical device status monitoring method based on artificial intelligence according to claim 1, characterized in that S5 include: S51, based on the status report of S4, generates a JSON instruction set containing the device ID, fault code and recommended measures; S52, calling a three-dimensional visualization engine to generate a device health status heat map according to the instruction set; S53, based on the heat map, push key alarm information to the mobile terminal and write it into the blockchain for evidence storage; S54, based on the evidence records in S53, automatically generates a preventive maintenance work order and assigns it to the nearest service station; S55: Update the fault resolution case library in the equipment knowledge graph based on the maintenance feedback data of S54.

9. The artificial intelligence-based medical device status monitoring method according to claim 8, characterized in that: If the same medical device triggers the same fault code more than three times within 24 hours, the spare parts replacement priority escalation procedure will be initiated; If the maintenance work order response time exceeds 4 hours and is not confirmed, it will be automatically transferred to the higher-level management center; If similar failure modes occur in related devices after the knowledge graph is updated, group warning notifications will be sent in batches.

Images