IoT-based intelligent monitoring method and system for medical device status

Abstract

This application discloses an intelligent monitoring method and system for the status of medical devices based on the Internet of Things (IoT), relating to the field of medical device monitoring. First, it utilizes information such as the device's service life and operating conditions to train a unique aging model for each device, predicting its normal performance parameters at the current stage of its life cycle. Then, by subtracting this predicted normal aging value from the real-time monitoring data, a clean residual signal is obtained. Subsequent anomaly detection will only target this residual signal; any significant fluctuations will likely point to an impending abnormal failure. This prediction-subtraction-detection model fundamentally solves the problem of aliasing between normal aging and abnormal deviations in data characteristics and adapts to the individual differences of devices, thereby accurately identifying true impending failures and significantly reducing the false alarm rate.

Description

Technical Field

[0001] This application relates to the field of medical device monitoring, and more specifically, to a method and system for intelligent monitoring of the status of medical devices based on the Internet of Things. Background Technology

[0002] With the rapid development and deep application of IoT technology, combining it with the management of medical equipment to build an intelligent status monitoring system has become a key path to ensure the quality of medical services, improve equipment management efficiency, and ensure patient safety. Medical equipment, as high-value and high-precision core assets, can suffer serious clinical risks and economic losses from any unexpected downtime or performance degradation. Therefore, real-time data collection of equipment operation through IoT technology, followed by intelligent analysis and predictive maintenance to identify potential faults in advance and extend equipment lifespan, has significant practical implications and application value.

[0003] However, existing equipment condition monitoring methods face a core technical bottleneck in practice. Current monitoring systems, whether based on simple threshold alarms or employing rudimentary machine learning models, often struggle to accurately distinguish between two fundamentally different types of equipment condition changes: normal aging-related degradation and abnormal pre-failure deviations. This confusion stems primarily from three aspects: First, in terms of data characteristics, the slow performance decline caused by long-term use (such as natural wear and tear of critical components) and the minute parameter drifts caused by early failures exhibit remarkably similar behavior in time-series signals, making traditional methods highly prone to false alarms or missed alarms. Second, existing models are mostly one-size-fits-all group models, ignoring the unique aging trajectories formed by each piece of equipment due to differences in usage frequency, operating environment, and maintenance history, lacking individualized adaptability. Finally, existing technologies generally lack explicit modeling of the dynamic process of aging; they typically treat any signal deviating from the initial static normal state as a potential anomaly, failing to recognize that the equipment's health baseline itself is a dynamically changing function with increasing service life.

[0004] Therefore, how to accurately separate normal aging trends from monitoring data, and thus effectively identify abnormal signals that truly represent the precursors of failure, has become a key technical problem that must be solved to achieve high-precision, low-false-alarm-rate intelligent monitoring. Summary of the Invention

[0005] To address the aforementioned problems in existing technologies, this application provides an intelligent monitoring method for the status of medical devices based on the Internet of Things (IoT). The method includes: cleaning and standardizing the acquired raw sensor data to obtain clean time-series data; calculating the service life and fusing features of the acquired static device information and device context data to obtain a fused feature vector; training and predicting an individualized aging baseline model based on historical clean time-series data and historical fused feature vectors to obtain an aging baseline model and predicted aging values; calculating residuals and constructing sequences from the clean time-series data and predicted aging values ​​to obtain a residual signal sequence; performing context-aware anomaly detection on the residual signal sequence based on the fused feature vector to obtain an anomaly score; and dynamically thresholding the anomaly score to obtain alarm information.

[0006] This application also provides an IoT-based intelligent monitoring system for the status of medical devices, comprising: a raw sensor data preprocessing module for cleaning and standardizing the acquired raw sensor data to obtain clean time-series data; a device data fusion module for calculating service life and fusing features of the acquired static device information and context data to obtain a fused feature vector; a baseline model generation module for training and predicting an individualized aging baseline model based on historical clean time-series data and historical fused feature vectors to obtain an aging baseline model and predicted aging values; a residual signal generation module for calculating residuals and constructing sequences from the clean time-series data and predicted aging values ​​to obtain a residual signal sequence; an anomaly analysis module for performing context-aware anomaly detection on the residual signal sequence based on the fused feature vector to obtain an anomaly score; and an alarm module for dynamically thresholding the anomaly score to obtain alarm information.

[0007] Compared with existing technologies, this application provides an intelligent monitoring method and system for the status of medical devices based on the Internet of Things (IoT). This method establishes a dynamic, individualized device aging baseline model to separate predictable normal aging trends from the raw monitoring signals. Instead of directly judging anomalies in the mixed raw data, this method first uses information such as the device's service life and operating conditions to train a dedicated aging model for each device, predicting its normal performance parameter values ​​at the current stage of its life cycle. Then, by subtracting this predicted normal aging value from the real-time monitoring data, a clean residual signal is obtained. Subsequent anomaly detection will only target this residual signal; any significant fluctuations will likely point to an abnormal fault precursor. This prediction-subtraction-re-detection model fundamentally solves the problem of aliasing between normal aging and abnormal deviations in data characteristics and adapts to the individual differences of devices, thereby accurately identifying true fault precursors and significantly reducing the false alarm rate. Attached Figure Description

[0008] The above and other objects, features and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings.

[0009] Figure 1 This is a flowchart of an IoT-based intelligent monitoring method for the status of medical devices according to an embodiment of this application.

[0010] Figure 2 This is a schematic diagram of data flow in an IoT-based intelligent monitoring method for the status of medical devices according to an embodiment of this application.

[0011] Figure 3 This is a flowchart of step S2 in the IoT-based intelligent monitoring method for the status of medical devices according to an embodiment of this application.

[0012] Figure 4 This is a flowchart of step S6 in the IoT-based intelligent monitoring method for the status of medical devices according to an embodiment of this application.

[0013] Figure 5 This is a block diagram of an IoT-based intelligent monitoring system for the status of medical devices according to an embodiment of this application. Detailed Implementation

[0014] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. It should be understood that the drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0015] In view of the shortcomings in the above-mentioned technical fields, this application proposes an intelligent monitoring method for the status of medical devices based on the Internet of Things. Figure 1 This is a flowchart of an IoT-based intelligent monitoring method for the status of medical devices according to an embodiment of this application. Figure 2 This is a schematic diagram illustrating the data flow of an IoT-based intelligent monitoring method for the status of medical devices according to an embodiment of this application. Figure 1 and Figure 2 As shown, the IoT-based intelligent monitoring method for medical device status according to an embodiment of this application includes: S1, cleaning and standardizing the acquired raw sensor data to obtain clean time-series data; S2, calculating the service life and fusing features of the acquired static information and context data of the device to obtain a fused feature vector; S3, training and predicting an individualized aging baseline model based on historical clean time-series data and historical fused feature vectors to obtain an aging baseline model and predicted aging value; S4, calculating residuals and constructing sequences from the clean time-series data and predicted aging value to obtain a residual signal sequence; S5, performing context-aware anomaly detection on the residual signal sequence based on the fused feature vector to obtain an anomaly score; and S6, performing dynamic threshold judgment on the anomaly score to obtain alarm information.

[0016] In step S1, the acquired raw sensor data is cleaned and standardized to obtain clean time-series data. It should be understood that in the IoT-based intelligent monitoring process for medical device status, data is the core driving force behind all analysis and decision-making. This raw time-series data, originating from sensors deployed on various key components of the device, is inevitably affected by electromagnetic interference, network fluctuations, or sensor instability during transmission and acquisition, resulting in noise, outliers, and missing data. Furthermore, data generated by different types of sensors (such as temperature, pressure, and vibration) often have drastically different physical units and numerical ranges. Directly using this mixed and inconsistent raw data for subsequent aging baseline modeling and anomaly detection will severely interfere with the model's training effect and prediction accuracy, failing to guarantee the reliability of the final monitoring results. Therefore, cleaning and standardizing the acquired raw sensor data to obtain high-quality clean time-series data is a necessary prerequisite and fundamental guarantee for ensuring the effectiveness and accuracy of the entire intelligent monitoring method.

[0017] For example, one feasible implementation of step S1 of this application is as follows: In this embodiment, the monitored object is an X-ray tube assembly of a medical computed tomography (CT) device. The raw sensor data is multi-dimensional time-series data related to the state of the tube, collected in real time based on the Internet of Things gateway. Specifically, it includes the coolant temperature (unit: degrees Celsius) reflecting its thermal management and the vibration amplitude (unit: g) of the anode rotating assembly reflecting its mechanical stability. The acquisition frequency is once per minute. Therefore, the acquired raw sensor data is represented as a multivariate time-series set, where each timestamp corresponds to a data point, such as [timestamp, coolant temperature, vibration amplitude].

[0018] First, a data cleaning process is performed on the raw sensor data. Data cleaning includes two sub-steps: missing value imputation and outlier handling. For missing value imputation, linear interpolation is used. For example, the data acquired at time stamp T is [T, 35.2, 0.15], while the data acquired at time stamp T+2 is [T+2, 35.4, 0.17], but the data point at time stamp T+1 is completely missing. In this case, linear calculation is performed using the two valid data points before and after T+1 to fill in the data at time T+1 with [T+1, 35.3, 0.16]. This method is suitable for scenarios where data changes are relatively gradual and the amount of missing data is small. Next, outlier handling is performed using a window-based moving median filtering method. The core of this method is to set a sliding window size, for example, 5. For each data point in the time series, its original value is replaced by the median of all data points within its window. The window size is determined based on the physical characteristics of the signal. For relatively slow-changing temperature signals, a slightly larger window can be selected to enhance the smoothing effect; for vibration signals that may contain high-frequency information, a smaller window is selected to retain effective details. For example, consider a series of coolant temperature readings [35.1, 35.2, 38.9, 35.4, 35.3]. The value 38.9 is clearly an anomalous spike caused by a transient disturbance. Centering on this value, a window of size 3 is taken [35.2, 38.9, 35.4], and the median is calculated to be 35.4. Thus, the original 38.9 is corrected to 35.4, effectively eliminating noise interference and obtaining a preliminarily clean data sequence.

[0019] After data cleaning, the data is standardized to eliminate the influence of different physical dimensions and numerical ranges, ensuring that different features are on the same scale, facilitating unified processing in subsequent models. This embodiment uses the Z-score standardization method. This method requires calculating the mean and standard deviation of each feature (i.e., coolant temperature and vibration amplitude) based on a representative historical clean data set. For example, by statistically analyzing historical data from the past month, the mean coolant temperature is 35.5 degrees Celsius, with a standard deviation of 1.5 degrees Celsius; the mean vibration amplitude is 0.20g, with a standard deviation of 0.05g. For a new cleaned data point [T, 36.1, 0.208], the standardization calculation process is as follows: the standardized temperature value is (36.1 - 35.5) / 1.5 = 0.4; the standardized vibration amplitude value is (0.208 - 0.20) / 0.05 = 0.16. After this processing, the original physical quantities are converted into dimensionless Z-score values.

[0020] Finally, the output is a clean time series dataset. This dataset is a multivariate time series with the same format as the original dataset but processed content. All missing values ​​have been filled, outlier noise has been smoothed, and all feature dimensions have been standardized to a distribution with a mean of 0 and a standard deviation of 1. For example, the aforementioned data point [T,36.1,0.208], after complete processing, outputs [T,0.4,0.16].

[0021] In step S2, the acquired static equipment information and equipment context data are used to calculate service life and fuse features to obtain a fused feature vector. Correspondingly, in the lifecycle management of medical devices, their performance is not static but undergoes a natural and gradual decline over time; this process is called aging. However, the rate and trajectory of aging are not solely determined by time; they are significantly influenced by dynamic factors such as the actual intensity of equipment use, operating conditions, and the environment in which it operates. To accurately construct an individualized aging baseline model that can predict the normal state of the equipment, it is necessary to quantify and fuse these key information characteristics that represent the equipment's age and experience. Therefore, calculating service life and fusing features on the acquired static equipment information and equipment context data to obtain a fused feature vector that comprehensively describes the current aging stage and operating conditions of the equipment is a core step in achieving subsequent individualized modeling and accurate prediction. Its purpose is to provide the model with key inputs that distinguish individual differences and understand dynamic operating conditions.

[0022] Figure 3 This is a flowchart of step S2 in the IoT-based intelligent monitoring method for the status of medical devices according to an embodiment of this application. Figure 3 As shown, one feasible method of step S2 in this application is to perform service life calculation and feature fusion on the acquired equipment static information and equipment context data to obtain a fused feature vector, including: S21, extracting the installation date from the equipment static information; S22, calculating the equipment service life based on the installation date and the current timestamp; S23, encoding and concatenating the operating condition features of the equipment context data to obtain an operating condition vector; S24, taking the equipment service life as the first element of the vector and concatenating it with the operating condition vector to obtain a fused feature vector.

[0023] For example, step S2 of this application is implemented as follows: Static equipment information refers to the attributes of the equipment that remain essentially unchanged after installation. It is recorded in a centralized asset management database or configuration management database when the equipment asset is put into storage or during installation and commissioning, and indexed by a unique equipment ID. For the aforementioned CT tube assembly, its static equipment information may be recorded as an entry containing multiple fields, such as: {Equipment ID: "CTX-SN12345", Equipment Model: "xxx", Manufacturing Date: "2021-01-15", Installation Date: "2021-03-01", Installation Location: "Room 1, Imaging Center, Hospital A"}. Equipment context data refers to non-sensor information collected synchronously with sensor data, reflecting the current operating status or environment of the equipment. It is generated by the equipment's control unit or external environment monitoring module and uploaded via the Internet of Things (IoT) gateway. For example, while collecting tube temperature and vibration data, the current scanning power, scanning protocol type, etc., are also recorded. In this example, the device context data at that moment is: {scan power: 100, scan protocol type: "head scan", room ambient temperature: 22.5}.

[0024] Step S21, extracting the installation date from the equipment static information, is implemented as follows: When a new cleanroom time-series data point needs to be processed, the process first obtains the equipment ID corresponding to that data point, namely CTX-SN12345. Then, using this ID as the query key, a data retrieval request is initiated to the asset management database. The goal of this request is to retrieve the equipment static information record associated with that ID and specifically parse the value of the installation date field from the returned record. In this application, the installation date in string form is extracted: 2021-03-01, which provides a benchmark for subsequent calculation of the equipment's service life.

[0025] Step S22: Calculate the equipment's service life based on the installation date and the current timestamp. The specific implementation process is as follows: The current timestamp is provided by the system clock of the server or edge computing node executing the calculation task, representing the real-time moment of data processing. Its format must be consistent with the installation date, for example, 2024-04-10T10:30:00Z. The calculation process first parses the installation date string and the current timestamp string into standard date and time objects. Next, a time difference calculation is performed, that is, subtracting the installation date object from the current time object to obtain a total duration representing the time interval. To make the service life feature have a unified and easily understood scale, this time interval needs to be converted into a standardized numerical unit. In this embodiment, days are chosen as the unit. For example, from March 1, 2021 to April 10, 2024, a total of 1137 days have passed. Therefore, the final output equipment service life is a numerical value: 1137. This value dynamically reflects the total time the CT tube has experienced since it was put into use, and is the most basic and important dimension for quantifying its aging degree.

[0026] Step S23 involves encoding and concatenating the operating condition features of the equipment context data to obtain an operating condition vector. The specific implementation process is as follows: This process first performs targeted processing on different types of features. Among the acquired equipment context data, scanning power and room ambient temperature are numerical features, while scanning protocol type is a categorical feature. For the categorical feature, scanning protocol type, a one-hot encoding method is used. This method requires a predefined ordered list containing all possible protocol types, determined based on the analysis of all functions of the equipment model or a large amount of historical data. For example, the predefined protocol type list is: ["Cranial plain scan", "Chest enhanced", "Abdominal plain scan", "Limb scan"]. Since the current value is a cranial plain scan, it corresponds to the first element in the list, therefore its one-hot encoding result is a four-dimensional binary vector [1,0,0,0]. For the numerical features scanning power and room ambient temperature, to eliminate dimensional differences and map them to a unified range, min-max normalization is used. This requires presetting the minimum and maximum values ​​for normal operation for each feature. These boundary values ​​are derived from the equipment's technical specifications manual or statistical analysis of long-term operating data. For example, the scanning power is set to a range of kilowatts, and the room temperature is set to a range of degrees Celsius. Based on this, the scanning power with a current value of 100 is normalized: (100-80) / (120-80) = 0.5. The room temperature with a current value of 22.5 is normalized: (22.5-18) / (26-18) = 0.5625. After encoding and normalizing all features, they are concatenated in a predetermined fixed order to form a condition vector. For example, the concatenation order is set as: normalized scanning power, normalized room temperature, and scan protocol type after thermal encoding. Based on this, the aforementioned calculation results are concatenated to obtain the condition vector [0.5, 0.5625, 1, 0, 0, 0].

[0027] Step S24 involves using the equipment's service life as the first element of the vector and concatenating it with the operating condition vector to obtain a fused feature vector. The process is as follows: The results of the first two steps are integrated, meaning the service life value is directly used as the first dimension of the new vector, and then all elements of the operating condition vector are appended to it sequentially. After the concatenation operation, the final fused feature vector is [1137, 0.5, 0.5625, 1, 0, 0, 0]. This fused feature vector not only contains the equipment's time aging information (service life) but also incorporates its instantaneous operating state description (operating condition vector), forming a high-dimensional, information-rich feature set.

[0028] In step S3, an individualized aging baseline model is trained and predicted based on historical cleanliness time-series data and historical fusion feature vectors to obtain the aging baseline model and predicted aging values. It is understandable that, since each medical device follows its unique aging trajectory influenced by its own usage history and operating conditions, a static threshold or a general group model cannot accurately define the normal state of a specific device at a specific moment. To accurately isolate true precursors of abnormalities from the complex monitoring signals, it is necessary to construct an individualized model for each device that dynamically reflects its health baseline. Therefore, training and predicting an individualized aging baseline model based on the device's own historical cleanliness time-series data and historical fusion feature vectors can learn and solidify the device's unique aging patterns, generating a dynamic baseline that can accurately predict the expected level of its normal performance parameters based on its current service life and operating conditions, providing a high-precision reference standard for subsequent residual calculations and anomaly identification.

[0029] One feasible approach to step S3 of this application is to train and predict an individualized aging baseline model based on historical clean time-series data and historical fusion feature vectors to obtain an aging baseline model and a predicted aging value, including: S31, training the aging baseline model with historical fusion feature vectors as input and historical clean time-series data as labels to obtain a trained aging baseline model; S32, inputting the fusion feature vectors into the trained aging baseline model to obtain a predicted aging value.

[0030] For example, step S3 of this application is implemented as follows: First, in order to perform step S31, i.e., model training, a corresponding historical dataset needs to be prepared. This dataset consists of two parts: historical clean time-series data and historical fusion feature vectors. These data are records accumulated and stored in a historical database since the device was installed. This historical data is obtained because it comprehensively depicts the performance and operating condition changes of the device throughout the entire aging process from a newer state to the current state, and is the sole basis for the model to learn its individualized aging pattern. Specifically, the historical clean time-series data is the collection of all clean time-series data processed by step S1 over the past few months or even years (e.g., from the installation date of March 1, 2021 to the present). Similarly, the historical fusion feature vectors are the collection of fusion feature vectors generated after processing by step S2, which correspond one-to-one with each historical clean time-series data point in terms of timestamp. These two parts of data together constitute a large training sample set, where each sample is a (feature, label) pair.

[0031] Step S31: Using historical fusion feature vectors as input and historical clean time-series data as labels, the aging baseline model is trained to obtain the trained aging baseline model. The specific implementation process is as follows: Since the temperature and vibration of the CT tube are two different physical quantities, their aging modes may also be different. Therefore, in this embodiment, two aging baseline models will be trained independently for the two sensor channels of coolant temperature and vibration amplitude, respectively. In terms of model selection, this embodiment uses Support Vector Regression (SVR) as the aging baseline model. The reason for choosing SVR is that it performs well in handling nonlinear relationships, has good adaptability to high-dimensional feature spaces, and is insensitive to noise by introducing interval boundaries, making it very suitable for modeling aging data with slow nonlinear trends. The model architecture and training process are illustrated using the coolant temperature channel as an example. The core of the SVR model is to find a regression function such that after the feature vectors of all training samples are mapped by this function, the deviation between the predicted value and the true label (i.e., the temperature value in the historical clean time-series data) does not exceed a preset tolerance, while maximizing the boundary of the tube. This embodiment uses the Radial Basis Function (RBF) as the kernel function for SVR, with the form exp(-gamma*||xi-xj||^2), where xi and xj represent the i-th and j-th historical fused feature vectors. This kernel function maps the original seven-dimensional fused feature vector [service age, normalized power, normalized ambient temperature, protocol code 1, protocol code 2, protocol code 3, protocol code 4] to an infinite-dimensional feature space, thereby capturing the complex nonlinear relationship between service age, operating conditions, and temperature. The model training process is essentially a process of solving a constrained quadratic optimization problem, with the goal of finding suitable Lagrange multipliers (i.e., the weights of the support vectors) and bias term b. These parameters are determined by executing the training algorithm on the historical dataset. Before training begins, several key hyperparameters need to be set. First is gamma in the kernel function, which determines the influence range of a single training sample; second is the penalty coefficient C, which balances the complexity of the model with the degree of penalty for samples exceeding the tolerance; and finally, the tolerance, which defines the range of biases that are not considered errors. The settings of these hyperparameters are crucial to model performance. A grid search combined with cross-validation is used to determine the optimal combination. For example, candidate values ​​[0.1, 1, 10] can be set for C, and candidate values ​​[0.01, 0.1, 1] can be set for gamma. Then, cross-validation is performed on all combinations, and the set of hyperparameters that minimizes the average error on the validation set is selected as the final configuration. After training begins, the algorithm uses the historical fused feature vector as input X and the standardized temperature values ​​from the corresponding historical clean time series data as labels y. Through iterative optimization, a set of support vectors and their corresponding weights, as well as a bias term b, are finally obtained. These parameters together define the trained aging baseline model.The model is fixed and saved as a file, bound to the device ID "CTX-SN12345" and the temperature channel, and stored in the model library for online access. Similarly, using historical vibration amplitude data as labels, an SVR model specifically for the vibration channel can be trained and obtained.

[0032] Step S32 involves inputting the fused feature vector into the trained aging baseline model to obtain the predicted aging value; this is the online real-time prediction stage. The specific implementation process is as follows: First, based on the currently processed sensor channel (e.g., coolant temperature), the previously trained and saved SVR model file, bound to the device CTX-SN12345 and the temperature channel, is loaded from the model library. Then, the current fused feature vector [1137,0.5,0.5625,1,0,0,0] is used as the input to the loaded model. Internally, the model utilizes its fixed support vectors, weights, biases, and RBF kernel function to calculate this new input vector and output a predicted value. This output value is the predicted aging value. For example, after inputting the above fused feature vector into the SVR model of the temperature channel, the model may output a predicted aging value of 0.21. The physical meaning of this value is as follows: for a CT tube that has been in service for 1137 days, currently performing a head scan at 100 kW power, and with a room temperature of 22.5 degrees Celsius, the standardized value of its coolant temperature should be 0.21. This value represents the theoretical normal value or healthy baseline under the current service life and operating conditions. Similarly, inputting this fused feature vector into the vibration channel model may yield another predicted aging value, such as 0.17. Ultimately, the output is a predicted aging value [0.21, 0.17] for each sensor channel at the current moment; for the temperature channel, the output is 0.21; for the vibration channel, the output is 0.17. These predicted values ​​are key to distinguishing between normal aging and abnormal offset.

[0033] In step S4, residual calculation and sequence construction are performed on the clean time-series data and predicted aging values ​​to obtain a residual signal sequence. It should be understood that although the preceding steps have yielded a dynamic baseline capable of accurately predicting the normal performance of equipment under specific service life and operating conditions, the original clean time-series data itself is still a mixture of normal aging trends and potential abnormal signals. Directly performing anomaly detection on this mixed data will still face the problem of misjudgment or missed detection due to the slow drift of the aging trend. To highlight weak fault precursor signals from the macroscopic aging background, an effective data separation method is needed. Therefore, this application provides a method of performing residual calculation and sequence construction on clean time-series data and predicted aging values ​​to mathematically remove the predictable normal aging components from the data, thereby purifying and amplifying those unexpected fluctuations that cannot be explained by the aging model.

[0034] For example, one feasible implementation of step S4 in this application is as follows: This calculation is performed independently for each sensor channel, that is, subtracting the theoretical aging value predicted by the model from the actual cleanliness time-series data value. Specifically, the calculation is as follows: For the coolant temperature channel, the residual value is 0.4 - 0.21 = 0.19. For the anode rotation assembly vibration amplitude channel, the residual value is 0.16 - 0.17 = -0.01. These two calculated residual values ​​represent the deviation between the actual measured value and the theoretical normal value. The temperature channel shows a relatively significant positive deviation, while the deviation of the vibration channel is within the normal range.

[0035] After calculating the residuals at a single time point, sequence construction is required. This is because residual values ​​at a single moment may exhibit random fluctuations, insufficient to constitute sufficient evidence for anomaly detection, while residual sequences over a continuous period can reveal meaningful patterns, trends, or volatile changes. Sequence construction is achieved by maintaining a fixed-length sliding window. The length of this window is a parameter that needs to be preset, and its setting depends on the duration of the anomaly pattern to be captured. For example, if experience indicates that a typical precursor to a fault in the X-ray tube manifests as a small parameter shift lasting for tens of minutes, the window length can be set to 60, storing residual values ​​from the past 60 minutes (since the sampling frequency is once per minute). The specific construction process is as follows: a separate queue of length 60 is maintained for each sensor channel. At each new timestamp, when a new residual value (e.g., 0.19) is calculated, this value is added to the end of the corresponding channel's queue. Simultaneously, if the queue is full (i.e., the length has reached 60), the oldest residual value at the head of the queue is removed. In this way, the queue always stores the 60 most recent and consecutive residual values, forming a dynamically updated residual signal sequence. For example, at the current moment, the residual signal sequence of the temperature channel may be in the form of [...,0.02,-0.03,0.01,0.19], where 0.19 is the most recently added value.

[0036] The final output is a residual signal sequence generated for each sensor channel. For the temperature channel, the output is a vector containing 60 floating-point numbers, representing the time series of deviations between the actual and theoretical temperature values ​​over the past hour. For the vibration channel, a similarly formatted vector is output. This residual signal sequence theoretically filters out all trend changes related to normal equipment aging, and the data should fluctuate smoothly around the zero mean.

[0037] In step S5, context-aware anomaly detection is performed on the residual signal sequence based on the fused feature vector to obtain an anomaly score. Correspondingly, although the signal sequence obtained after residual calculation has successfully separated the predictable normal aging trend of the equipment, whether the fluctuations it contains constitute an anomaly cannot be judged by an isolated, absolute scale. This is because the normal fluctuation range of the residual is closely related to the equipment's current operating conditions (such as load, power, and mode). For example, slight residual fluctuations generated under high-power scanning may be completely normal, while the same fluctuations appearing in the equipment's standby state are highly likely to indicate a potential fault. Therefore, this application combines the analysis of the residual signal with the perception of real-time operating conditions to perform context-aware anomaly detection, constructing an intelligent judgment framework that can dynamically understand what kind of residual behavior is reasonable under the current operating conditions. This avoids misjudging reasonable fluctuations under normal operating conditions as anomalies, achieving truly accurate fault identification with low false alarms.

[0038] One feasible approach to step S5 of this application is to perform context-aware anomaly detection on the residual signal sequence based on the fused feature vector to obtain an anomaly score, including: S51, performing feature engineering on the residual signal sequence to obtain residual statistical features; S52, concatenating the residual statistical features with the fused feature vector to obtain a concatenated feature vector; and S53, inputting the concatenated feature vector into the trained anomaly detection model to obtain an anomaly score.

[0039] For example, step S5 of this application is implemented as follows: This embodiment will take the coolant temperature channel as an example to illustrate the processing process in detail. Step S51, feature engineering is performed on the residual signal sequence to obtain residual statistical features. The specific implementation process is as follows: The residual signal sequence of the coolant temperature channel is a vector of length 60, for example [...,0.02,-0.03,0.01,0.19]. Directly using this high-dimensional original sequence as model input is not only computationally expensive, but also makes it difficult to capture its inherent statistical regularity. Therefore, feature engineering is needed to extract a set of statistical features from this sliding time window that can effectively summarize its distribution characteristics and dynamic behavior. In this embodiment, the selected statistical features include: mean, standard deviation, skewness, and kurtosis. The selection of these features all has clear physical meaning: the mean reflects the overall direction and magnitude of the residual signal's shift over the past hour; a mean that consistently deviates from zero may indicate a new, systematic drift not captured by the aging model. The standard deviation measures the dispersion or fluctuation of the residual signal; a significant increase in volatility is usually associated with increased system instability. Skewness describes the asymmetry of the residual distribution; normally, the residuals should be roughly symmetrically distributed on both sides of zero. A significant skewness may indicate a unidirectional shock or fault. Kurtosis measures the sharpness of the distribution and the thickness of the tails; an abnormally high kurtosis value indicates more extreme outliers in the residual sequence, which is often a direct manifestation of sudden faults. The calculation process involves applying standard statistical formulas to a sample set of 60 input residual values. For example, for the input temperature residual sequence, due to the newly added larger residual values, its statistical characteristics will undergo a clear but reasonable shift. The calculated residual statistical characteristics are: mean 0.05, standard deviation 0.08, skewness 0.8, and kurtosis 3.6. This set of values ​​constitutes a four-dimensional residual statistical feature vector [0.05, 0.08, 0.8, 3.6]. This vector condenses a 60-dimensional original time series into a 4-dimensional feature representation with higher information density.

[0040] Step S52 involves concatenating the residual statistical features with the fused feature vector to obtain a concatenated feature vector. This process is crucial for achieving context awareness. The purpose of concatenation is to combine the features describing signal performance (residual statistical features) with the features describing the background (operating condition features), and then feed them into the subsequent anomaly detection model. It is worth noting that the long-term aging trend represented by the first element in the fused feature vector, service age (1137), has already been consumed in steps S3 and S4 by generating and using the aging baseline model. In the current anomaly detection stage, the focus is on instantaneous anomalies; therefore, the correlation between the residual signal and instantaneous operating conditions is more important. Thus, in this concatenation step, the part representing the real-time operating conditions, i.e., the operating condition vector, is extracted from the fused feature vector. This operating condition vector is [0.5, 0.5625, 1, 0, 0, 0], corresponding to the normalized scan power, the normalized room ambient temperature, and the one-hot encoded scan protocol type, respectively. The splicing operation merges two vectors into a longer vector according to a predefined, strictly fixed order. For example, the set order is to place the residual statistical features first, followed by the operating condition vector. Following this rule, [0.05,0.08,0.8,3.6] and [0.5,0.5625,1,0,0,0] are spliced ​​together to obtain a 10-dimensional spliced ​​feature vector: [0.05,0.08,0.8,3.6,0.5,0.5625,1,0,0,0]. This spliced ​​feature vector is a highly information-rich composite; its first four dimensions precisely describe the real-time dynamic behavior of the signal after filtering out aging trends, while the latter six dimensions clearly indicate the specific operating environment of the equipment when these behaviors occur. By inputting such a complete information vector containing both performance and context into the subsequent anomaly detection model, the model can learn and judge whether the current signal performance exceeds its reasonable range under specific working conditions, thereby making a more context-aware and accurate anomaly judgment.

[0041] The processing flow of this embodiment follows the previous steps, and its core is to use a trained anomaly detection model to process the concatenated feature vectors generated in step S52. This embodiment selects an autoencoder as the anomaly detection model. This is an unsupervised model based on deep learning, consisting of an encoder and a decoder. The model is trained using only a large number of historical concatenated feature vectors generated when the device is operating in a confirmed healthy state. The training objective is to minimize the model's reconstruction error, that is, to make the model's output vector as consistent as possible with the input vector. In this way, the model deeply learns and memorizes all the inherent correlations and patterns of normal data. When an anomaly data input that does not conform to the normal pattern, the model will be unable to reconstruct it well, resulting in a large reconstruction error. One feasible approach to step S53 of this application, which involves inputting the concatenated feature vector into a trained anomaly detection model to obtain an anomaly score, includes: S531, inputting the concatenated feature vector into the encoder of the trained anomaly detection model to obtain a concatenated feature latent space representation; S532, inputting the concatenated feature latent space representation into the decoder of the trained anomaly detection model to obtain a reconstructed concatenated feature vector; and S533, calculating the reconstruction error between the reconstructed concatenated feature vector and the concatenated feature vector as the anomaly score.

[0042] First, the specific architecture and training process of the autoencoder model are explained. The model is trained independently for each sensor channel (e.g., coolant temperature). Its input layer has 10 neurons, corresponding to a 10-dimensional concatenated feature vector. The encoder part consists of two fully connected hidden layers with 8 and 4 neurons respectively, compressing the 10-dimensional input into a 4-dimensional latent space representation. The decoder part has a symmetrical structure to the encoder, also consisting of two fully connected hidden layers with 8 and 10 neurons respectively, restoring the 4-dimensional latent space representation to a 10-dimensional reconstructed vector. All hidden layers use ReLU as the activation function. The model's weights and biases, among other parameters, are obtained through backpropagation training using the Adam optimizer on a large dataset containing only historical concatenated feature vectors under normal operating conditions. The loss function is the mean squared error (MSE) between the input and output.

[0043] Step S531 involves inputting the concatenated feature vector into the encoder of the trained anomaly detection model to obtain the concatenated feature latent space representation. The specific implementation process is as follows: The 10-dimensional concatenated feature vector of the coolant temperature channel is fed into the input layer of the trained temperature channel autoencoder model. The data flows through two hidden layers of the encoder (10->8->4). In each layer, the input is multiplied by the weight matrix of that layer and a bias vector is added, then passed through the ReLU activation function. Finally, at the output of the encoder, i.e., the bottleneck layer with four neurons, a 4-dimensional vector is obtained. This vector is a highly condensed representation of the original 10-dimensional input, i.e., the concatenated feature latent space representation. For example, after calculation, the obtained concatenated feature latent space representation might be [1.8, -0.95, 0.8, 3.8].

[0044] Step S532 involves inputting the concatenated feature latent space representation into the decoder of the trained anomaly detection model to obtain a reconstructed concatenated feature vector. The implementation of this step follows immediately. The concatenated feature latent space representation is fed into the decoder. The data flows through two hidden layers (4->8->10) of the decoder, undergoing a decompression process opposite to that of the encoder. The decoder uses its learned weights and biases to attempt to reconstruct the original 10-dimensional vector most likely to have generated it from this compressed latent information. Finally, a 10-dimensional reconstructed concatenated feature vector is obtained at the decoder's output layer. Since the current input concatenated feature vector represents an anomaly state, the model will try its best to reconstruct it into a "normal" pattern that it has learned that is most similar to it. For example, the resulting reconstructed concatenated feature vector might be [0.01, 0.03, 0.526, 3.4, 0.5, 0.562, 0.98, 0.01, 0.02, 0.01]. As can be seen, the values ​​of the first four, representing the statistical characteristics of the residuals, are significantly smaller than the outliers in the input vector, and are closer to the expected performance of a healthy state; while the reconstructed values ​​of the last six dimensions, representing the working conditions, are very close to the original input, especially the last four, which clearly reproduce the structure of the one-hot encoding, and their values ​​are very close to 0 or 1.

[0045] Step S533 calculates the reconstruction error between the reconstructed and concatenated feature vectors as the anomaly score. This is the final step in quantifying the degree of anomaly. The calculation method involves squared the difference between the two vectors element-wise, then summing the squares of all 10 differences to obtain the mean squared error. For example, calculating (0.05-0.01)... 2 +(0.08-0.03) 2 +(0.8-0.526) 2 +(3.6-3.4) 2+... This calculation ultimately yields a single scalar value. In this example, because the original vector represents an anomalous state, the model cannot accurately reconstruct it, resulting in a significant difference between the original and reconstructed vectors. The calculated reconstruction error will be a significantly larger value, such as 0.12. This value is the anomaly score. A high anomaly score means that the current device state deviates severely from its known healthy mode.

[0046] Similarly, for the vibration amplitude channel of the anode rotating component, the entire process of step S5 described above is executed independently and in parallel. Specifically, feature engineering is performed based on its own residual signal sequence, which is then concatenated with the operating condition vector at the same time. The resulting concatenated feature vector is then input into an autoencoder anomaly detection model specifically trained for the vibration channel. This allows for the calculation of an anomaly score specific to the vibration channel.

[0047] In step S6, a dynamic threshold judgment is performed on the anomaly score to obtain alarm information. In other words, after obtaining the quantified anomaly score, the next crucial step is how to make an accurate alarm judgment based on this score. Using a fixed, unchanging threshold for judgment, while simple, ignores the crucial contextual factors in equipment status assessment. The actual risk level of an anomaly score can vary drastically depending on the task the equipment is currently performing, the business environment, and its own importance. Without considering this dynamically changing background information, it is easy to mistake minor fluctuations in non-critical tasks for serious alarms, or to react slowly to early signs of real danger in core business scenarios. Therefore, dynamic threshold judgment of anomaly scores can establish an intelligent, flexible, and business scenario-driven alarm triggering mechanism, ensuring that the sensitivity of alarms can adapt to the current operational risk level in real time, thereby significantly improving the accuracy and effectiveness of alarms.

[0048] Figure 4 This is a flowchart of step S6 in the IoT-based intelligent monitoring method for the status of medical devices according to an embodiment of this application. Figure 4 As shown, one feasible method of step S6 in this application, which involves dynamically judging anomaly scores to obtain alarm information, includes: S61, determining a context adjustment factor based on device context data; S62, determining a business criticality factor based on the device's business criticality; and S63, dynamically adjusting a base threshold based on the context adjustment factor and the business criticality factor to obtain the dynamic threshold. Another feasible method of step S6 in this application, which involves dynamically judging anomaly scores to obtain alarm information, further includes: S64, processing the dynamic threshold based on a predefined multiplier factor to obtain multiple hierarchical boundaries; and S65, classifying the anomaly scores into severity levels based on the multiple hierarchical boundaries to obtain the alarm information.

[0049] For example, step S6 of this application is implemented as follows: Step S61, based on the device context data, determines the context adjustment factor, the specific implementation process of which is as follows: The core of this step is to predefine a context-adjustment factor mapping table. This table assigns a risk adjustment coefficient to each of the various operating modes or protocol types that the device may perform. The establishment of this mapping table is based on the knowledge and experience of domain experts and aims to quantify the alarm sensitivity requirements under different operating conditions. For example, for a CT device, a predefined mapping table may be as follows: {"cardiac interventional surgery guidance":0.8,"chest enhanced scan":0.9,"head plain scan":1.0,"device self-test":1.2}. Among them, a factor less than 1 indicates that higher sensitivity (i.e., a lower threshold) is required, and a factor greater than 1 indicates that the sensitivity can be appropriately relaxed. In the currently input device context data, the scanning protocol type is head plain scan. By querying this mapping table, the adjustment factor corresponding to head plain scan is found to be 1.0. This value is the context adjustment factor.

[0050] Step S62, based on the equipment's business criticality, determines the business criticality factor. The specific implementation process is as follows: Equipment business criticality refers to the importance of the equipment itself within its workflow or department; this is a relatively static attribute. For example, a CT scanner installed in the emergency department for emergency treatment has a much higher business criticality than the same model installed in a teaching and research institution for routine research. Similar to the previous step, this step also relies on a predefined criticality-adjustment factor mapping table. This table is set according to the equipment's installation location, department, or asset level. For example, a predefined mapping table might be: {"Emergency Department":0.7,"Imaging Center":0.9,"Physical Examination Center":1.0,"Teaching and Research Department":1.1}. In this embodiment, the static information of the known equipment CTX-SN12345 includes its installation location as Room 1 of the Imaging Center at Hospital A. By querying this mapping table, the adjustment factor corresponding to the Imaging Center is found to be 0.9. This value is the business criticality factor.

[0051] Step S63: Based on the context adjustment factor and the business criticality factor, the basic threshold is dynamically adjusted to obtain the dynamic threshold. The specific implementation process is as follows: The dynamic threshold is obtained through statistical analysis of a large number of historical abnormal scores generated during operation. A common setting method is to take the 99.9th percentile of the historical normal score distribution to ensure that alarms are almost never triggered under normal circumstances. For example, after analysis, the basic threshold for this temperature channel is determined to be 0.1. The formula for calculating the dynamic threshold is: Dynamic threshold = Basic threshold × Context adjustment factor × Business criticality factor. Substituting the values ​​in this embodiment into the calculation: 0.1 × 1.0 × 0.9 = 0.09. Finally, the output is a dynamically adjusted threshold: 0.09. This value accurately reflects the upper limit of abnormal scores that a CT scanner installed in an imaging center and performing a head scan can tolerate.

[0052] The dynamic threshold for the coolant temperature channel is set to 0.09. First, a preliminary alarm trigger check is performed. The current anomaly score of 0.12 is compared to the dynamic threshold of 0.09. Since 0.12 is greater than 0.09, the alarm condition is confirmed to have been triggered, and the process proceeds to the severity level classification stage. If the anomaly score is less than or equal to 0.09, no alarm is detected, the alarm level is output as 0, and the process terminates.

[0053] Step S64 involves processing the dynamic threshold based on predefined multiplier factors to obtain multiple hierarchical boundaries. The specific implementation process is as follows: A set of predefined multiplier factors, greater than or equal to 1, is used to construct a tiered alarm escalation system. These multiplier factors are determined based on a deep understanding of the fault modes of this type of equipment, statistical analysis of historical fault data, and operational management procedures. They represent the escalation steps of the operational team's attention to different levels of anomalies. In this embodiment, the predefined set of multiplier factors is [1.0, 1.2, 1.5], corresponding to the starting boundaries of the three alarm levels: Attention, Warning, and Danger. The calculation process involves multiplying the dynamic threshold by each multiplier factor to obtain a series of specific numerical boundaries. The first hierarchical boundary (the lower limit of the "Attention" level) is: 0.09 × 1.0 = 0.09. The second hierarchical boundary (the lower limit of the "Warning" level) is: 0.09 × 1.2 = 0.108. The third grading boundary (the lower limit of the "hazard" level) is: 0.09 × 1.5 = 0.135. The output is this set of calculated grading boundaries used to classify the severity of the alarm: [0.09, 0.108, 0.135].

[0054] Step S65: The severity level of the anomaly score is classified based on multiple grading boundaries to obtain the alarm information. The specific implementation process is as follows: The core of the processing is to compare the anomaly score with these grading boundaries to determine its severity level. The specific classification logic is as follows: First, determine whether the anomaly score is greater than or equal to the first grading boundary 0.09 and less than the second grading boundary 0.108. The current anomaly score of 0.12 is not within this range. Next, determine whether the anomaly score is greater than or equal to the second grading boundary 0.108 and less than the third grading boundary 0.135. The current anomaly score of 0.12 satisfies the condition 0.108 ≤ 0.12 < 0.135. Therefore, this anomaly score is determined to belong to the warning level. If the anomaly score is greater than or equal to the third grading boundary 0.135, it will be determined to be the highest danger level. Based on the above judgment, the alarm level of the current device is determined to be level 2 (“Warning”).

[0055] Ultimately, the output is a structured alarm message. This message is not merely a simple alarm signal, but rather contains rich contextual decision support data. For example, the output alarm message might be a data object containing multiple fields: {Alarm Timestamp:"2024-04-10T10:30:00Z", Device ID:"CTX-SN12345", Monitoring Channel:"Coolant Temperature", Anomaly Score: 0.12, Dynamic Threshold: 0.09, Alarm Level: 2, Alarm Description: "Warning"}. This alarm message clearly informs maintenance personnel of the time of the incident, the specific device, the location of the problem, the quantification of the anomaly, the judgment criteria at the time, and its severity rating, providing comprehensive and accurate information for subsequent troubleshooting and maintenance decisions.

[0056] Similarly, for the vibration amplitude channel of the anode rotating component, the entire process of step S6 described above is executed independently and in parallel. Based on its own anomaly score and dynamic threshold, a graded judgment is made, and its exclusive alarm information is generated, thereby realizing a multi-dimensional and comprehensive assessment and alarm of the equipment status.

[0057] In particular, after obtaining the quantified anomaly score, setting an alarm threshold that effectively captures early fault signals while avoiding numerous false alarms under complex and variable operating conditions is crucial to the success of the entire monitoring method. Adjusting using simple fixed thresholds or heuristic rules essentially assumes that the impact of equipment operating conditions on the normal fluctuation range of the anomaly score is simple, independent, and linear, which is severely inconsistent with physical reality. For example, under simultaneous high load and high power, the normal fluctuation range of equipment sensor signals may experience a sharp, non-linear amplification, which cannot be accurately characterized by simple multiplication factors. Therefore, using context-aware conditional probability dynamic threshold judgment for anomaly scores can abandon the approach of adjusting fixed thresholds and instead adopt a data-driven modeling method. This directly learns and predicts the statistical distribution of the anomaly score of a healthy device under any given real-time operating condition, and directly defines a fully adaptive, statistically significant normal boundary from the extremely high quantiles of this distribution, thereby achieving the most accurate and context-specific setting of the alarm threshold.

[0058] Based on this, a preferred method for step S6 of this application involves dynamically thresholding the abnormal scores to obtain alarm information, including: constructing a historical normal operation dataset. This step is the cornerstone of the entire modeling process, providing high-quality, uncontaminated healthy samples for subsequent model training to ensure that the model learns the behavior patterns of the equipment under various normal operating conditions, rather than being interfered with by fault data. In specific implementation, it is necessary to filter all data within the time periods confirmed as normal operation from all historical data of the equipment, such as the CT tube assembly with ID "CTX-SN12345" since its installation. This confirmation can be based on various information, such as excluding all time periods where maintenance work orders have occurred or those marked as abnormal by domain experts. Finally, a dataset containing tens of thousands of records is constructed. Each record in the dataset is a data pair (Ci, Si), where Ci is the operating condition vector generated by step S2 at time i; and Si is the corresponding abnormal score calculated by step S5 at the same time.

[0059] The mean and standard deviation prediction models are trained to obtain well-trained mean and standard deviation prediction models. Then, under a given operating condition C, the anomaly score S follows a model with mean μ(C) and standard deviation σ... 2 (C) represents the core statistical assumption that the variance follows a normal distribution: S|C~N(μ(C),σ 2(C) is instantiated using a machine learning model to obtain two parameterized models capable of predicting the distribution of normal and abnormal scores in real time based on any operating condition. In practice, two regression models need to be trained independently. The first is the mean prediction model Mμ. Its training objective is to learn the complex relationship between the operating condition vector C and the expected central value μ(C) of the abnormal score. In this embodiment, the mean prediction model Mμ uses a gradient boosting decision tree model, which is good at handling tabular data and can automatically learn complex nonlinear interactions between features. In practice, the historical normal operation dataset constructed in the previous step is used, with the operating condition vector C as the model input (feature) and the corresponding abnormal score S as the model output (label), to train a regression model, such as a gradient boosting tree or a neural network. The second is the standard deviation prediction model Mσ. Its training objective is to learn the relationship between the operating condition vector C and the fluctuation range σ(C) of the abnormal score. Its implementation consists of two steps: First, labels need to be prepared for training Mσ. Using the already trained Mμ, for each sample i in the historical dataset, its absolute prediction error Errori=|Si-Mμ(Ci)| is calculated. The Errori value represents the degree to which the true score deviates from its expected center and can serve as an effective approximation of the standard deviation. Then, using the work condition vector C as the model input and the calculated Errori as the model output, a second regression model Mσ, also based on a gradient boosting decision tree, is trained. The hyperparameters of both models are optimized using methods such as cross-validation and grid search to ensure their generalization ability.

[0060] The current context feature vector is input into the trained mean prediction model and standard deviation prediction model, respectively, and a dynamic threshold is calculated based on a preset confidence parameter to obtain the dynamic threshold. Then, a unique, highly contextualized alarm threshold is calculated for the current device state. The effect is that the threshold can be continuously and smoothly adjusted as the operating conditions change smoothly. In specific implementation, the current context feature vector, i.e., the operating condition vector C(t), needs to be obtained first, such as the aforementioned [0.5, 0.5625, 1, 0, 0, 0]. Simultaneously, a confidence parameter k needs to be preset, which represents the statistical standard deviation multiple and directly determines the alarm sensitivity. The larger the k value, the higher the threshold and the higher the system's tolerance for anomalies. The k value is set between 3 and 5 and can be fine-tuned based on business criticality; for example, k=2.8 can be set for critical equipment in the emergency department, while k=3.5 can be set for routine physical examination equipment. Furthermore, the value of k can be dynamically adjusted based on business criticality, for example, by setting it to the product of a base k value and a business criticality factor retrieved from a mapping table. Next, C(t) is input into the pre-trained Mμ and Mσ models, yielding two outputs: the predicted mean Mμ(C(t)) and the predicted standard deviation Mσ(C(t)). For example, the model predicts that under current conditions, the normal abnormal score mean should be 0.02 and the standard deviation should be 0.025. Finally, the dynamic threshold is calculated using the following formula: Dynamic threshold = Mμ(C(t)) + k * Mσ(C(t)). The physical meaning of this formula is that under current conditions, any abnormal score exceeding the expected center value plus k times the fluctuation range will be considered a statistically low-probability event, i.e., a potential anomaly. Taking the aforementioned values ​​as an example, and setting k=3, the dynamic threshold is calculated as: 0.095 = 0.02 + 3 * 0.025.

[0061] Alarm information is obtained based on dynamic thresholds. In other words, the preceding numerical calculation results are transformed into a final output that provides clear guidance for operations and maintenance personnel, enabling differentiated and refined alarm responses.

[0062] In practice, the current actual anomaly score calculated in step S5 (e.g., 0.12) is compared with the dynamically calculated threshold (0.095). Since 0.12 > 0.095, an alarm is triggered. Furthermore, a tiered alarm mechanism can be incorporated. A predefined multiplier factor (e.g., [1.0, 1.2, 1.5]) is used to process the dynamic threshold 0.095, resulting in multiple tiered boundaries [0.095, 0.114, 0.1425]. Because the current anomaly score of 0.12 falls between 0.114 and 0.1425, the final generated alarm information will be defined as a "warning" level (Level 2), containing complete information such as time, device, channel, score, threshold, and level, providing comprehensive support for subsequent maintenance decisions. Specifically, this process is implemented identically to step S65 above, and therefore will not be described in detail.

[0063] To further illustrate the technical advantages of this preferred embodiment over the basic embodiment, a specific scenario can be envisioned. For example, at a certain moment, the CT device is in a low-load standby stable operating condition. At this time, due to the early deterioration of a certain electronic component, its coolant temperature exhibits a very small but continuous abnormal fluctuation, and the actual abnormal score calculated by step S5 is 0.04.

[0064] When using the basic implementation method (i.e., using a fixed base threshold and a multiplied adjustment factor) for judgment, since the device is in standby mode, its context adjustment factor may be set to a value greater than 1 to reduce false alarms, such as 1.2. In this case, the dynamic threshold is calculated as: base threshold (0.1) × context adjustment factor (1.2) × business criticality factor (0.9) = 0.108. Since the current true anomaly score of 0.04 is much smaller than the calculated dynamic threshold of 0.108, this method will determine that the device is in a completely normal state and will not generate any alarm information, thus potentially missing the best opportunity to detect early potential faults.

[0065] However, the situation is quite different when using this preferred embodiment for judgment. As designed in this preferred solution, the model can automatically learn from the data that under stable operating conditions such as device standby or low load, it will predict very low mean and variance of abnormal scores. The already trained mean prediction model Mμ and standard deviation prediction model Mσ, through learning from a large amount of historical data, have mastered that the distribution of normal abnormal scores should have extremely low mean and extremely small variance under stable standby conditions. For example, when the context feature vector representing the standby condition is input, the mean Mμ(C(t)) predicted by the model may be only 0.005, and the predicted standard deviation Mσ(C(t)) may be 0.01. Based on this, and with the confidence parameter k=3, the calculated dynamic threshold is: 0.005+3×0.01=0.035. At this time, the current true abnormal score of 0.04 has exceeded this more sensitive and more consistent dynamic threshold of 0.035, thus triggering an alarm. Furthermore, when classifying alarms, the first classification boundary (“Attention” level) is 0.035 × 1.0 = 0.035, and the second classification boundary (“Warning” level) is 0.035 × 1.2 = 0.042. Since the actual anomaly score of 0.04 falls exactly between 0.035 and 0.042, an alert of the Attention level will ultimately be generated.

[0066] This process generates lower, more sensitive thresholds, successfully capturing minute but potentially fatal anomalies that occur under stable operating conditions and are ignored by basic implementations due to their overly lenient thresholds. This significantly improves the early warning capability and accuracy of the entire monitoring method. Conversely, under normal operating conditions such as high equipment load, which naturally generate large residual fluctuations, this preferred solution also predicts higher mean and variance, thereby generating higher and more reasonable dynamic thresholds and effectively avoiding misjudging normal operating condition fluctuations as anomalies. Furthermore, because it processes continuous contextual features (such as load percentage and ambient temperature), the threshold can be smoothly and continuously adjusted with smooth changes in operating conditions, rather than being adjusted abruptly based on discrete states, thus achieving stronger adaptability.

[0067] In summary, the IoT-based intelligent monitoring method for medical device status, based on embodiments of this application, is explained. By establishing a dynamic and individualized device aging baseline model, predictable normal aging trends are separated from the original monitoring signals. This method no longer directly judges anomalies in the mixed raw data. Instead, it first uses information such as the device's service life and operating conditions to train a dedicated aging model for each device to predict its normal performance parameter values ​​at the current stage of its life cycle. Then, by subtracting this predicted normal aging value from the real-time monitoring data, a clean residual signal is obtained. Subsequent anomaly detection will only target this residual signal; any significant fluctuations are highly likely to point to an abnormal fault precursor. This prediction-subtraction-re-detection mode fundamentally solves the problem of aliasing between normal aging and abnormal deviations in data characteristics and adapts to the individual differences of devices, thereby accurately identifying true fault precursors and significantly reducing the false alarm rate.

[0068] Figure 5 This is a block diagram of an IoT-based intelligent monitoring system for the status of medical devices according to an embodiment of this application. Figure 5 As shown, the IoT-based intelligent monitoring system for medical device status 100 according to an embodiment of this application includes: a raw sensor data preprocessing module 110, used to clean and standardize the acquired raw sensor data to obtain clean time-series data; a device data fusion module 120, used to calculate service life and fuse features of the acquired static device information and device context data to obtain a fused feature vector; a baseline model generation module 130, used to train and predict an individualized aging baseline model based on historical clean time-series data and historical fused feature vectors to obtain an aging baseline model and predicted aging value; a residual signal generation module 140, used to calculate residuals and construct sequences from the clean time-series data and predicted aging value to obtain a residual signal sequence; an anomaly analysis module 150, used to perform context-aware anomaly detection on the residual signal sequence based on the fused feature vector to obtain an anomaly score; and an alarm module 160, used to perform dynamic threshold judgment on the anomaly score to obtain alarm information.

[0069] Here, those skilled in the art will understand that the specific operations of each step in the above-described IoT-based intelligent monitoring system for medical device status have been referenced above. Figures 1 to 4 The method for intelligent monitoring of the status of medical devices based on the Internet of Things has been described in detail, and therefore, its repeated description will be omitted.

Claims

1. A method for intelligent monitoring of the status of medical devices based on the Internet of Things, characterized in that, include: The acquired raw sensor data is cleaned and standardized to obtain clean time-series data; The service life of the acquired equipment static information and equipment context data are calculated and features are fused to obtain a fused feature vector. Individualized aging baseline models are trained and predicted based on historical clean time-series data and historical fusion feature vectors to obtain aging baseline models and predicted aging values. Residual signal sequences are obtained by calculating residuals and constructing sequences from clean time-series data and predicted aging values. Based on the fused feature vector, context-aware anomaly detection is performed on the residual signal sequence to obtain anomaly scores; Dynamic threshold judgment is performed on abnormal scores to obtain alarm information.

2. The method for intelligent monitoring of medical device status based on the Internet of Things according to claim 1, characterized in that, The acquired equipment static information and equipment context data are used to calculate service life and fuse features to obtain a fused feature vector. This includes: extracting the installation date from the equipment static information; calculating the equipment service life based on the installation date and the current timestamp; encoding and concatenating the operating condition features of the equipment context data to obtain an operating condition vector; and concatenating the equipment service life as the first element of the vector with the operating condition vector to obtain the fused feature vector.

3. The method for intelligent monitoring of medical device status based on the Internet of Things according to claim 1, characterized in that, The process of training and predicting an individualized aging baseline model based on historical clean time-series data and historical fused feature vectors to obtain an aging baseline model and predicted aging values ​​includes: training the aging baseline model with historical fused feature vectors as input and historical clean time-series data as labels to obtain a trained aging baseline model; and inputting the fused feature vectors into the trained aging baseline model to obtain predicted aging values.

4. The method for intelligent monitoring of medical device status based on the Internet of Things according to claim 1, characterized in that, Based on the fused feature vector, context-aware anomaly detection is performed on the residual signal sequence to obtain anomaly scores, including: performing feature engineering on the residual signal sequence to obtain residual statistical features; concatenating the residual statistical features with the fused feature vector to obtain a concatenated feature vector; and inputting the concatenated feature vector into the trained anomaly detection model to obtain anomaly scores.

5. The method for intelligent monitoring of medical device status based on the Internet of Things according to claim 4, characterized in that, The process of inputting the concatenated feature vector into the trained anomaly detection model to obtain an anomaly score includes: inputting the concatenated feature vector into the encoder of the trained anomaly detection model to obtain a concatenated feature latent space representation; inputting the concatenated feature latent space representation into the decoder of the trained anomaly detection model to obtain a reconstructed concatenated feature vector; and calculating the reconstruction error between the reconstructed concatenated feature vector and the concatenated feature vector as the anomaly score.

6. The method for intelligent monitoring of medical device status based on the Internet of Things according to claim 1, characterized in that, The abnormal score is dynamically thresholded to obtain alarm information, including: determining a context adjustment factor based on device context data; determining a business criticality factor based on the device's business criticality; and dynamically adjusting the basic threshold based on the context adjustment factor and the business criticality factor to obtain the dynamic threshold.

7. The method for intelligent monitoring of medical device status based on the Internet of Things according to claim 6, characterized in that, The method of dynamically thresholding abnormal scores to obtain alarm information also includes: processing the dynamic threshold based on a predefined multiplier factor to obtain multiple grade boundaries; and classifying the abnormal scores into severity levels based on the multiple grade boundaries to obtain the alarm information.

8. An intelligent monitoring system for the status of medical equipment based on the Internet of Things, characterized in that, include: The raw sensor data preprocessing module is used to clean and standardize the acquired raw sensor data to obtain clean time-series data. The equipment data fusion module is used to calculate the service life and fuse features of the acquired equipment static information and equipment context data to obtain a fused feature vector. Baseline The model generation module is used to train and predict an individualized aging baseline model based on historical clean time-series data and historical fusion feature vectors to obtain the aging baseline model and predicted aging value. The residual signal generation module is used to perform residual calculation and sequence construction on clean time series data and predicted aging values ​​to obtain a residual signal sequence. The anomaly analysis module is used to perform context-aware anomaly detection on the residual signal sequence based on the fused feature vector to obtain anomaly scores; the alarm module is used to perform dynamic threshold judgment on the anomaly scores to obtain alarm information.

Images