Medical metrology node self-learning dynamic calibration system

Abstract

The application discloses a medical metrology node self-learning dynamic calibration system. The system comprises: a data acquisition module acquires a multi-view entropy vector of a target detection device. A data processing module performs change point detection on the multi-view entropy vector to obtain a change point intensity. An abnormal score calculation module calculates a maximum difference value between the multi-view entropy vector and a first baseline and a second baseline difference value, and calculates an abnormal score by taking the maximum difference value and the change point intensity. A first determination module determines whether the abnormal score is greater than or equal to a first threshold value. If yes, a full calibration instruction is issued; otherwise, a consumable health index is calculated, and the health index is fitted to a consumable remaining service life. A second determination module determines whether the consumable health index and the consumable remaining service life are less than a corresponding preset threshold value. If one or both satisfy the condition, a short-range calibration instruction is issued; if neither satisfies the condition, a delayed calibration instruction is issued. The application dynamically distinguishes between device abnormalities and consumable degradation problems by combining multi-view entropy vectors with double baseline comparison.

Description

Technical Field

[0001] This invention relates to the field of medical testing setup calibration technology, specifically a medical metrology node self-learning dynamic calibration system. Background Technology

[0002] With the development of modern medicine, various medical testing equipment (such as clinical biochemical analyzers, blood analyzers, immunoassay analyzers, and imaging diagnostic equipment) have been widely used in disease screening, clinical diagnosis, and efficacy evaluation. The accuracy of medical test results is directly related to the scientific nature of treatment decisions and the health and safety of patients. Therefore, precise monitoring and management of the operating status of testing equipment and the health status of key consumables are core aspects of ensuring testing quality.

[0003] In actual medical testing, testing equipment typically relies on multiple sensors, detection modules, and various consumables (such as reagent kits, optical probes, electrodes, and filter membranes) working together. As usage time increases, testing equipment may generate abnormal signals due to factors such as consumables and differences in operating habits. If these abnormalities are not identified and handled in a timely manner, they may lead to deviations in test results or even misjudgments.

[0004] Currently, the calibration of testing equipment is mostly based on replacing consumables at fixed intervals or calibrating the entire equipment, without taking into account the degradation trend of consumables and the real-time changes in the operating status of the equipment. The calibration strategy is not flexible enough, which can easily lead to increased maintenance costs or delayed intervention that affects the quality of testing. Summary of the Invention

[0005] To address the lack of flexibility in existing calibration strategies that rely on fixed-cycle consumable replacements or full-scale equipment calibration, this invention provides a self-learning dynamic calibration system for medical metrology nodes.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] In the first aspect, this application discloses a medical metrology node self-learning dynamic calibration system, including a data acquisition module, a data processing module, an anomaly score calculation module, a first judgment module, and a second judgment module.

[0008] The data acquisition module is used to acquire the multi-view entropy vector of the target detection device. The multi-view entropy vector is calculated by the edge device based on touch timing data.

[0009] The data processing module is used to preprocess the multi-view entropy vector, detect change points on the processed multi-view entropy vector, and then calculate the change point intensity.

[0010] The anomaly score calculation module is used to calculate the difference between the preprocessed multi-view entropy vector and the first baseline and the second baseline, respectively, and to take the maximum difference and the change point intensity for weighted calculation to obtain the anomaly score; wherein, the first baseline is the reference value of the target detection device under normal health conditions, and the second baseline is the reference mean of all devices of the same model as the target detection device under normal health conditions.

[0011] The first determination module is used to determine whether the abnormal score is greater than or equal to the first threshold. If so, a full calibration command is issued; otherwise, the processed multi-view entropy vector is mapped to the consumable health index in an unsupervised manner, and the health index is degraded and fitted to the remaining service life of the consumable.

[0012] The second determination module is used to determine whether the health indicators of consumables and the remaining service life of consumables are less than the corresponding preset thresholds. If one or both are satisfied, a short-range calibration command is issued; if neither is satisfied, a delayed calibration command is issued.

[0013] Secondly, this application discloses a self-learning dynamic calibration method for medical metrology nodes, including the following steps:

[0014] The multi-view entropy vector of the target detection device is obtained. The multi-view entropy vector is calculated by the edge device based on touch timing data.

[0015] The multi-view entropy vector is preprocessed, and the change point is detected after processing. Then, the change point intensity is calculated.

[0016] The preprocessed multi-view entropy vector is compared with the first baseline and the second baseline. The maximum difference is taken and weighted with the change point intensity to obtain the anomaly score. The first baseline is the reference value of the target detection device under normal and healthy conditions, and the second baseline is the reference mean of all devices of the same model as the target detection device under normal and healthy conditions.

[0017] If the abnormal score is greater than or equal to the first threshold, a full calibration command is issued; otherwise, the processed multi-view entropy vector is mapped to the consumable health index in an unsupervised manner, and the health index is degraded and fitted to the remaining service life of the consumable.

[0018] Determine whether the health indicators and remaining service life of the consumables are less than the corresponding preset thresholds. If one or both are met, issue a short-range calibration command; otherwise, issue a delayed calibration command.

[0019] Thirdly, this application discloses a self-learning dynamic calibration method for medical metrology nodes, including the following steps:

[0020] Collect operational behavior data, metering signal data, consumable data, and test result data of the target detection equipment;

[0021] After preprocessing the collected data, information entropy, permutation entropy, sample entropy, and Markov transition entropy were calculated respectively.

[0022] The information entropy, permutation entropy, sample entropy, and Markov transition entropy are combined into a multi-view entropy vector and sent to the central processing unit.

[0023] After the central processing unit generates a calibration command based on the multi-view entropy vector, it receives and executes the calibration command and sends back the operation log after execution.

[0024] Fourthly, this application discloses a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned medical metrology node self-learning dynamic calibration method.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0026] 1. Based on the anomaly score and consumable life calculated by multi-view entropy vector, this application generates maintenance instructions of different levels such as full calibration, short-range calibration, and delayed calibration, which match the high precision and high stability requirements of medical metrology nodes such as hospital testing centers, laboratory laboratories, and POCT testing points, making equipment maintenance more flexible and efficient, reducing unnecessary downtime, and improving equipment availability;

[0027] 2. During long-term operation, the system in this application continuously adjusts model parameters through multi-view entropy vectors and calibration effects, achieving self-learning and strategy optimization, reducing the risk of detection result deviation, and ensuring the accuracy of medical metrology. Attached Figure Description

[0028] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0029] Figure 1 This is a structural block diagram of the medical metrology node self-learning dynamic calibration system described in Embodiment 1 of the present invention;

[0030] Figure 2 Based on Figure 1 A system block diagram with a baseline update module;

[0031] Figure 3 This is a time series diagram illustrating the relationship between the location of change points and entropy jumps;

[0032] Figure 4 Degeneracy fitting and RUL estimation plot;

[0033] Figure 5A flowchart of the overall process of a self-learning dynamic calibration system for medical metrology nodes;

[0034] Figure 6 Based on Figure 1 Application scenario illustration diagram;

[0035] Figure 7 This is a flowchart of the self-learning dynamic calibration method for medical metrology nodes described in Example 2;

[0036] Figure 8 This is a flowchart of the self-learning dynamic calibration method for medical metrology nodes described in Example 3. Detailed Implementation

[0037] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0038] Application Overview

[0039] In current technologies, with the development of modern medicine, various medical testing devices have been widely used in disease screening, clinical diagnosis, and efficacy evaluation. The accuracy of medical test results is directly related to the scientific nature of diagnostic and treatment decisions and the health and safety of patients. Testing devices typically rely on the collaborative work of multiple sensors, detection modules, and various consumables. However, with increased usage time, factors such as consumable degradation and differences in operating habits may cause abnormal signals from the equipment. Existing calibration methods are mostly based on fixed-cycle replacement of consumables or full-scale calibration of the equipment, without considering the real-time changes in consumable degradation trends and equipment operating status. This leads to increased maintenance costs or delayed interventions that affect test quality. For example, a hospital's biochemical analyzer experienced test result deviations due to reagent kit performance degradation. However, traditional methods could not promptly identify the degradation status of consumables, and calibration was still performed according to the fixed cycle, increasing reagent consumption and potentially delaying the handling of abnormalities.

[0040] To address the aforementioned issues, this research revealed a limitation of existing calibration strategies: their inability to dynamically perceive changes in the correlation between device status and consumable health. Analysis of device operational data uncovered multi-dimensional status information hidden within touch timing data. Further, a method was proposed to utilize multi-view entropy vectors to characterize device operational features, combined with change point detection technology to capture abrupt changes. Simultaneously, considering both individual device differences and group commonalities, a dual baseline comparison mechanism was designed, dynamically triggering a tiered calibration strategy through anomaly scoring. The resulting self-learning calibration system achieves coordinated monitoring of device status and consumable lifespan.

[0041] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0042] Example 1

[0043] like Figure 1 As shown in the figure, this embodiment introduces a medical metrology node self-learning dynamic calibration system, including a data acquisition module 100, a data processing module 200, an anomaly score calculation module 300, a first judgment module 400, and a second judgment module 500.

[0044] The data acquisition module 100 is used to acquire the multi-view entropy vector of the target detection device. The multi-view entropy vector is calculated by the edge device based on the touch timing data.

[0045] Among them, the multi-view entropy vector refers to the multi-dimensional feature that characterizes the device's operating status through a combination of information entropy, permutation entropy, sample entropy and Markov transition entropy. It is implemented by using edge computing devices to perform multi-entropy parallel computation on touch timing data, and is used to comprehensively reflect the complexity of the device's operating status.

[0046] Touch timing data primarily consists of touch events (press / swipe data from touchscreens, touchpads, or virtual buttons) generated during the operation of the target detection device. These touch events are recorded by the device's built-in sensor interface driver at a millisecond-level sampling frequency. Edge computing devices utilize high-speed buses (such as SPI, I / O) to record these events. 2 (C, USB) Real-time read. The caching strategy is to store the touch timing data of the most recent N seconds (e.g., 5-10 seconds) to ensure that the calculation latency is below the set threshold.

[0047] The collected data undergoes preprocessing, including time synchronization and interpolation, followed by normalization and noise filtering. Then, edge computing devices perform parallel computation.

[0048] Information entropy:

[0049] Suppose that the touch timing data is discretized to obtain a finite set of states. Its probability distribution is Then the information entropy is:

[0050]

[0051] Where n is the total number of states.

[0052] Permutation entropy:

[0053] Given time series The embedding dimension is m, and the time delay is Construct vectors:

[0054]

[0055] Will The components are arranged in order of size, forming an arrangement pattern. Calculate the probability of all possible permutations. Then the permutation entropy is:

[0056]

[0057] Sample entropy:

[0058] Let the time series length be N, the embedding dimension be m, and the tolerance threshold be r (usually taken as 0.1 to 0.25 times the standard deviation).

[0059] Construct an m-dimensional vector:

[0060]

[0061] Define distance function This represents the maximum difference between vectors. Count the number of matches:

[0062]

[0063] Similarly, construct an m+1 dimensional vector and calculate the number of matches:

[0064]

[0065] The sample entropy is then:

[0066]

[0067] Markov transition entropy:

[0068] Suppose that the time series is discretized into a set of states {s}.

[0069] The Markov transition entropy is:

[0070]

[0071] in, These are the discretized state values ​​of the touch timing signal at times t+1, t, and t-1, respectively. Represents a state triple The probability of occurrence; Indicates that given the current state and the previous state In the case of the next state The probability of occurrence; This means that, given only the current state... In the case of the next state The probability of occurrence.

[0072] The four types of entropy values ​​are output according to the calculation window to form a four-dimensional vector. Each multi-view entropy vector is appended with a timestamp to ensure synchronization with subsequent change point detection and baseline comparison. Different entropy types are computed in parallel using multiple cores to reduce latency.

[0073] The data processing module 200 is used to preprocess the multi-view entropy vector, perform change point detection on the processed multi-view entropy vector, and then calculate the change point intensity. Specifically:

[0074] The multi-view entropy vectors undergo preprocessing operations such as time alignment, denoising, standardization, and feature fusion. Then, change point detection is performed. First, the sequence is divided into sliding windows of length w, and statistical features such as mean and variance are calculated within each window.

[0075] Compare the distribution differences between adjacent windows:

[0076]

[0077] in, The probability distribution function of the reference window (the previous window); The probability distribution function for the comparison window (the next window); Let P be the probability of state i occurring under distribution P; Let Q be the probability of state i occurring under distribution Q.

[0078] like Exceeding the threshold is considered a change point. The KL threshold can be determined empirically or through permutation. Empirical setting is preferred, as it allows for rapid implementation even without a large amount of labeled data. Specifically, a baseline set is constructed using multiple data segments from the device's "healthy and stable period," the KL divergence sequence for the baseline period is calculated, and the threshold is set as a quantile or mean standard deviation control line to obtain the KL threshold.

[0079] For each entropy component, detect the change point to obtain a candidate set. Then, use a voting mechanism or weighted summation to output the final change point position.

[0080] At the detected change point location t∗, calculate the change point intensity and quantify the degree of abrupt change. For example... Figure 3 The diagram shown illustrates the relationship between t∗ and entropy jumps. The intensity of the change point can be calculated using either the mean difference method or the variance ratio method. Taking the mean difference method as an example:

[0081]

[0082] in, and These are the mean values ​​of the window before and after the point change, respectively.

[0083] The comprehensive intensity index is calculated by weighting multiple metrics and used as the change point intensity.

[0084]

[0085] in, , These are empirical weights or obtained through training. Let KL divergence be the probability distribution of the front and back windows. This represents the probability distribution of the multi-view entropy vector within the window before the point of change. This represents the probability distribution of the multi-view entropy vector within the window after the point of change.

[0086] The anomaly score calculation module 300 calculates the difference between the preprocessed multi-view entropy vector and the first and second baselines, respectively, and takes the maximum difference and weighted calculation based on the change point intensity to obtain the anomaly score; wherein, the first baseline is the reference value of the target detection device under normal and healthy conditions, and the second baseline is the reference mean of all devices of the same model as the target detection device under normal and healthy conditions. The specific steps are as follows:

[0087] First, calculate the dimensional standardized difference between the multi-view entropy vector and the first and second baselines. Then, the vectorized distance formula is calculated to obtain... The vectorized distance formula is:

[0088]

[0089] in, Let p be the dimension weight, with a norm order of p=2, and d represent the d-th dimension in the multi-view entropy vector.

[0090] Maximum difference selection: .

[0091] Maximum difference and change point intensity Normalization process, anomaly score calculation:

[0092]

[0093] in, , The maximum difference and change point strength after normalization; weights The preferred value is 0.6 to 0.7.

[0094] The first judgment module 400 is used to determine whether the abnormal score is greater than or equal to the first threshold. If so, a full calibration command is issued; otherwise, the processed multi-view entropy vector is mapped to a consumable health indicator in an unsupervised manner, and the health indicator is degraded and fitted to the remaining service life of the consumable. The specific steps are as follows:

[0095] like ∈(0,1] (first threshold), issue a full calibration command.

[0096] like This involves unsupervised mapping of entropy vectors to consumable health indicators, specifically estimating the health mean and covariance based on the equipment's "health stable period" data. Calculate the Mahalanobis distance:

[0097]

[0098] in, This is the processed multi-view entropy vector; for The inverse matrix; This is for the transpose operation.

[0099] Map the distance to HI (0 = fault / range, 1 = healthy):

[0100]

[0101] Among them, the normalization scale Take the average of the healthy and stable period 1–3 times; requires a sample size of more than 50 during the period of stable health.

[0102] Can be Short-term smoothing (anti-burr) is obtained Fitting RUL requires at least 10 historical HI samples.

[0103] Set fault thresholds For unacceptable health levels of the equipment (e.g., 0.3 or 0.5), or as defined by the operational strategy and testing criteria.

[0104] Degenerate fitting can employ linear, exponential, or power-law degenerate methods, taking linear fitting as an example:

[0105] Assume that HI is approximately linear with time: The initial health level estimate is obtained through least squares fitting. and degradation rate estimation .

[0106] If the current time is Then predict the time point when the threshold will be reached. Calculate the remaining useful life (RUL) at the current moment: .

[0107] Constructing prediction intervals using residual variance: Perform bootstrap resampling B (≥5) times on the HI time series residuals, repeatedly fitting to obtain the RUL estimated sequence, and taking the quantiles to give the confidence interval. Use a sliding time window (e.g., 50–200 HI points) to fit the model, with the window shift frequency matched to the sampling rate to handle non-stationarity and drift. Each estimation simultaneously outputs the current... , Confidence intervals and fit quality indices (residual standard deviation). For example... Figure 4 The figure shown is a graph of degenerate fitting and RUL estimation. Figure 4 As can be seen from the graph, the vertical axis represents the health index. The closer the index is to 1, the healthier the equipment or consumables are. The graph shows that the health index of the target equipment is decreasing. Figure 4 The smaller plot at the bottom center is the residual (fitting interval) plot, reflecting the deviation between the model and the actual degradation.

[0108] The second determination module 500 is used to determine whether the health indicators of the consumables and the remaining service life of the consumables are less than the corresponding preset thresholds. If one or both are satisfied, a short-range calibration command is issued; if neither is satisfied, a delayed calibration command is issued.

[0109] The threshold for consumable health indicators is the fault threshold. The threshold for the remaining service life of consumables is defined as the lifespan threshold. The threshold value is determined based on business needs or statistical distribution. In practical applications, multi-level thresholds can be used, for example:

[0110] : Remind maintenance personnel (e.g., 20% remaining lifespan).

[0111] Short-range calibration trigger (e.g., 10% remaining lifetime).

[0112] Minimum margin before forced shutdown.

[0113] like and / or If the condition is met, a short-range calibration command will be issued. Otherwise, a delayed calibration command will be issued, along with a suggested re-inspection interval. .

[0114] Maximum time interval; weight ∈[0.05,0.2].

[0115] It is important to emphasize that if only one indicator is available, that indicator should be used for judgment. If both are missing, maintain the previous instruction and report a "pending confirmation" status, and perform extreme value checks. , If it exceeds the reasonable range, it will be manually reviewed.

[0116] Furthermore, estimation can be based on the degradation rate. Dynamic adjustment : .

[0117] in, The threshold before adjustment, As a regulating factor, The degradation rate estimated at the current moment. The historical average degradation rate is typically obtained based on population data or historical equipment cycles and is used for normalization.

[0118] Therefore, this application can monitor the operating status of medical testing equipment and the health of consumables in real time, and dynamically trigger graded calibration commands through an anomaly scoring mechanism. When a sudden anomaly occurs in the equipment, full calibration is quickly initiated; during the gradual degradation stage of consumables, short-term or delayed calibration is selected based on the predicted remaining lifespan, effectively reducing unnecessary maintenance operations and avoiding testing errors caused by calibration delays. It also achieves efficient data processing and dynamic baseline updates, enhancing the system's adaptive capabilities.

[0119] The main scheme of this embodiment has been introduced above. The following is a detailed description of the second determination module 500 determining whether the degradation score is greater than a preset degradation threshold and issuing corresponding instructions, as follows:

[0120] After determining that the health indicators and remaining service life of the consumables are both greater than the corresponding preset thresholds, the second determination module 500 also determines whether the degradation score is greater than the preset degradation threshold. If yes, it issues a rapid self-test command; otherwise, it issues a delayed calibration command.

[0121] The degradation score is calculated by the data processing module using 200 pairs of viewpoint entropy vectors to determine the entropy change rate, combined with the retry rate of the target detection device, the number of process interruptions, and the degradation acceleration, all weighted together. The retry rate of the target detection device... Number of process interruptions Data was collected by data acquisition module 100. Before calculating the degradation score, all data needs to be normalized. The formula for calculating the degradation score is:

[0122]

[0123] in, The entropy change rate is calculated by the difference between the multi-view entropy vectors in adjacent time windows; degradation acceleration. That is, the degeneracy slope The rate of change; The weighting coefficients can be determined based on experience or statistics. For example, the weights are related to the volatility of features in historical samples, and weights can be allocated accordingly. Alternatively, they can be determined empirically. For example, the entropy change rate (a core health indicator) can be given a larger weight, the retry rate and the number of process interruptions can reflect moderate operational stability, and the degradation acceleration can be used as a trend indicator to assist in judgment. This method is suitable for the initial stage.

[0124] like If the degradation threshold is met, a fast self-check command is output; otherwise, a delayed calibration command is output. The degradation threshold can be determined using statistical / quantile methods or supervised learning. For example, the degradation score distribution can be calculated on a healthy sample set, and the tail probability can be selected. Calculated : , For the sample quantiles, You can choose 0.05, 0.1, etc.

[0125] The retry rate refers to the frequency at which the equipment automatically restarts after an error occurs during the execution of the testing process. It is quantified by counting the number of restart commands triggered in the anomaly log per unit time and is used to characterize the wear and tear of mechanical components. The number of process interruptions refers to the number of unplanned terminations of the testing process due to hardware failures or software anomalies. It is counted by preset abnormal termination markers in the operation log and is used to identify system-level reliability degradation.

[0126] This application addresses the problem of undetected potential degradation trends in equipment when consumables are in normal condition, enabling early detection of latent performance degradation caused by mechanical wear and decreased system stability. By quantitatively assessing the overall degree of equipment degradation, it avoids calibration delays or excessive maintenance due to misjudgment of a single indicator, optimizes maintenance resource allocation while ensuring testing accuracy, and prevents test result deviations caused by calibration strategy failures.

[0127] The above describes in detail the specific steps of the second judgment module 500 in making judgments based on the degradation score. The following section details the additional judgments made by the first judgment module 400 based on change point intensity and quality control samples. The specific steps are as follows:

[0128] The first determination module 400 is also used to determine whether the change point intensity is greater than the strong change point threshold and whether the quality control sample deviation exceeds the preset allowable range. If so, a full calibration command is issued.

[0129] Among them, the quality control sample deviation is calculated by the data processing module 200 based on the deviation between the actual value and the nominal value of the quality control sample obtained by the data acquisition module 100.

[0130] Actual value of quality control sample Data is collected by data acquisition module 100, and the nominal value of the quality control sample is... From equipment calibration settings. Strong change point threshold. The change point intensity statistic is obtained by processing data from the target detection equipment during its healthy and stable period. We obtain the upper quantile: ; For the sample Quantiles of 1, first probability 1. Initially, a value of 0.02~0.05 can be used, and then fine-tuned during backtesting. Regarding the threshold for strong change points... Alternatively, ROC curve analysis can be used to determine this.

[0131] When performing deviation calculations, both absolute and relative deviation calculations can be performed, and corresponding preset allowable ranges are also set for absolute deviations. and relative deviation preset allowable range The preferred method is to determine the optimal approach based on data from the equipment's stable health period.

[0132] Right now .

[0133] in, For the sample Quantiles of 2, first probability 2. Take 0.01~0.05. It is a constant greater than 0, used to avoid when When the denominator is close to 0, it becomes too small, leading to distorted results.

[0134] We can first determine the intensity at the change point, and then... At that time, during the quality control deviation calculation and the determination of the absolute deviation Or relative deviation If so, a full calibration command will be issued.

[0135] For example, if the electrode performance of the blood analyzer undergoes a sudden change, causing the change point intensity to reach 0.85 (the strong change point threshold is set to 0.8), and the deviation of the hemoglobin detection value of its quality control sample exceeds ±5%, the system will immediately initiate a full calibration process.

[0136] This application, through the synergistic verification of change point intensity and quality control sample deviation, can accurately identify composite faults involving sudden equipment anomalies and test result deviations. For example, in a scenario where the optical module of a biochemical analyzer is aging, the system can simultaneously capture the abrupt change characteristics of the entropy vector and the absorbance detection deviation of the quality control sample, thereby triggering calibration in a timely manner when the equipment experiences substantial performance degradation, avoiding delays in calibration due to relying solely on entropy vector analysis.

[0137] The above describes in detail the process by which the first judgment module 400 makes judgments based on change point intensity and quality control samples. The following section details the time-scale detection of the processed multi-view entropy vector by the data processing module 200 when calculating change point intensity through change point detection. The specific steps are as follows:

[0138] When the data processing module 200 performs change point detection and calculates the change point intensity, it performs first time scale detection and second time scale detection on the processed multi-view entropy vector. If both detections find a change point and the time difference does not exceed the time alignment threshold, the change point is identified as a true change point, and its change point intensity is calculated.

[0139] The window length for the second timescale detection is at least 10 times that for the first timescale detection.

[0140] The specific method for time-scale detection is as follows:

[0141] First, construct a single-scale statistic, using time point t as the dividing point, and take the front / back window. Then, calculate the standardized differences within the perspective, and finally fuse the standardized differences of multiple perspectives to obtain the multi-perspective normalized difference statistic.

[0142] First time scale (short window): Window length (Number of sample points), step size / slip ( Second time scale (long window): window length (Number of sample points), requirements ≥10 Step length ( ).

[0143] Local maxima are detected on the multi-view normalized difference statistics of the two time windows, and each is thresholded, i.e., greater than or equal to a threshold (which can be taken as the P95–P99 quantile of the healthy segment or ROC / cost method calibration), to obtain two sets of candidate time points. .

[0144] For each ,exist Find the nearest neighbor ,like Then a pair is formed. And record the true turning point. If a candidate point has no matching at another scale, it is considered a spurious change point.

[0145] Among them, the time alignment threshold ,default The value is 1. The time interval for a single sampling point.

[0146] This application effectively solves the problem of misjudgment of change points caused by fluctuations in equipment operating status or temporary interference. Specifically, it achieves this by rapidly capturing potential abnormal signals at short timescales and verifying the persistence of abnormal signals through long-term analysis. The synergistic effect of these two methods filters out instantaneous fluctuations caused by operational interference and sensor noise, thereby improving the accuracy of change point detection. Simultaneously, the proportional design of the window length ensures that the two detection perspectives complement each other in terms of temporal resolution and statistical stability, avoiding detection blind spots or overfitting problems caused by improper window size settings.

[0147] The above describes in detail the steps of the data processing module 200 in performing time-scale detection on the processed multi-view entropy vector when calculating the change point intensity. The following section provides a detailed explanation of the corresponding adjustments made to the first and second baselines based on the sample types detected by the target detection device.

[0148] Sample types include blood, urine, saliva, swabs, reagent batches, etc. If a first and second baseline for the corresponding sample are stored, they are retrieved directly from the baseline database after the sample type is identified. If the type identification is uncertain, a pooled baseline is created based on the confidence level. If there is no record for the type, the database reverts to the parent type, uses a cross-type baseline for the population, and establishes / updates the first and second baselines after collecting sufficient healthy samples, gradually converging the baselines.

[0149] This application can automatically match the corresponding reference baseline for different sample types, ensuring that the calibration benchmark is consistent with the physical characteristics of the current testing scenario, thereby reducing the risk of misjudgment due to differences in sample types. For example, when testing new samples after the reagent kit is changed, the system can accurately distinguish between abnormal equipment status and changes in sample characteristics by updating the baseline reference value, avoiding unnecessary full calibration operations, and improving the triggering accuracy of short-range calibration commands.

[0150] The baseline update module 600, which is also included in the system, is described in detail below:

[0151] like Figure 2 As shown, the baseline update module is used to dynamically update the first and second baselines; the baseline update module includes:

[0152] The first baseline update unit is used to add the current multi-view entropy vector to the historical dataset, recalculate the reference value and use it as the first baseline when the target detection device returns to normal after executing the full calibration command or short-range calibration command.

[0153] The second baseline update unit is used to periodically obtain the reference values ​​of all devices of the same model as the target detection device under normal health conditions, and the average value is calculated as the second baseline.

[0154] Specifically as follows:

[0155] Before updating, the first and second baseline update units perform admission checks, namely:

[0156] 1. The most recent or current batch quality control deviation meets the requirements. .

[0157] 2. Variable point strength .

[0158] 3. Anomaly Scoring .

[0159] 4. The sample types are consistent.

[0160] The admission verification criteria are not limited to the examples above.

[0161] First baseline update unit: After calibration, continuously If each sampling period meets the "access verification" requirement, it is marked as "return to normal". A cooldown period can be set. (e.g., 5–15 minutes or 50–200 samples), sampling should only begin after the cooling period to avoid transients.

[0162] Add the current multi-view entropy vector to the historical dataset, controlling the size and retaining the most recent ones. Item or most recent Time window, decay of old samples.

[0163] A robust estimate is performed based on the historical dataset after sampling, and a new first baseline is obtained through smooth updates. This baseline is then versioned and recorded. If the false alarm rate increases during a safe period (e.g., the most recent week), the system rolls back to the previous version.

[0164] The second baseline update unit only includes devices that meet the admission criteria – for a given time period. Device weights are positively correlated with device activity / sample size and negatively correlated with historical false alarm rate / online stability (unstable devices have lower weights), and are finally normalized.

[0165] The sample set is collected by category. First, the mean is calculated for each device (to reduce correlation), and then a weighted mean is calculated across devices. A second baseline is output periodically and recorded in a versioned manner. If the false alarm rate increases during a safe period (e.g., the most recent week), the system is rolled back to the previous version.

[0166] This application achieves dynamic adaptation and continuous optimization of calibration baselines for medical testing equipment. The individual baseline update mechanism ensures that the reference values ​​of the equipment accurately reflect the current health status under different sample types, avoiding misjudgments caused by component aging or environmental changes. The group baseline update mechanism eliminates the impact of common degradation among equipment groups on the benchmark through statistical learning, maintaining the uniformity of calibration standards. The synergistic effect of these dual update strategies solves the adaptation problem of traditional fixed baselines in complex application scenarios, improving the robustness and accuracy of the calibration system.

[0167] The baseline update module has been described in detail above. The following section provides a detailed explanation of how the data acquisition module 100 adjusts its acquisition frequency according to the frequency of sample detection by the target detection device:

[0168] The acquisition frequency refers to the time interval at which the data acquisition module 100 acquires multi-view entropy vectors from the target detection device. The acquisition interval automatically shortens when the sample detection frequency increases and lengthens when the sample detection frequency decreases. The sample detection frequency refers to the number of times the target detection device completes sample detection per unit time, which can be obtained in real-time through device operation logs or sensor signals.

[0169] By dynamically linking the acquisition frequency with the sample detection frequency, it is possible to increase the data acquisition density to capture transient abnormal signals when the equipment is under high load, and reduce redundant data storage when the equipment is under low load.

[0170] To facilitate a better understanding of this solution, please refer to the table below for the execution coverage of different instructions:

[0171] Table 1: Comparison of the differences among the four types of calibration instructions

[0172] Instruction type Triggering conditions Scope and content Time / Resource Impact Data and baseline impact Risk Preference Level Application Scenario Description Full calibration command Anomaly score ≥ first threshold or strong change point threshold exceeds limit and quality control sample bias exceeds limit. System-level end-to-end calibration, including zero / range recalibration, multi-view synchronization, and sample type mapping verification. This process is the most time-consuming, requiring a complete calibration procedure and potentially necessitating a short downtime. Data during the calibration period is marked as "calibrated," and the first baseline update is triggered upon restoration to normal; the second baseline is not directly affected. Highest New reagent batch launched; strong inflection point detected and QC exceeded limits; systematic drift. Short-range calibration instructions Consumable health index < threshold or RUL ≤ lifespan threshold Rapid calibration of local cells, such as consumable-related channels, background subtraction, and drift compensation. Medium execution time, moderate resource consumption, can be executed during runtime intervals. The first baseline update is triggered after the data returns to normal; data markers are locally adjusted. higher Consumables are nearing the end of their lifespan; health indicators are declining but have not yet reached a systemic abnormality. Quick self-test command HI > threshold and RUL > threshold, but degradation score ≥ degradation threshold Short-term verification and consistency checks do not modify system parameters. Shortest processing time, no impact on production capacity Only generate a self-test report, do not update the baseline. middle An increase in short-term retry rates and a rise in the number of process interruptions necessitates risk assessment. Delay calibration instructions HI > threshold and RUL > threshold and degradation score < degradation threshold Without calibration, only observation is performed, and a re-inspection interval is set. Zero time consumption, only scheduling Do not update data and baseline lowest All indicators are normal or fluctuate slightly; a follow-up test is scheduled.

[0173] In summary, this embodiment uses multi-view entropy vector analysis combined with dual baseline comparison to dynamically distinguish between equipment malfunctions and consumable degradation issues, and generates differentiated calibration instructions based on change point intensity and malfunction scores to achieve precise triggering of calibration strategies and optimized resource allocation.

[0174] To facilitate understanding of the above embodiments, a specific application scenario of the above embodiments will be used as an example for illustration below:

[0175] like Figure 5-6 As shown, taking a blood analyzer as the target detection device, the detection frequency is set to once every 10 minutes. The edge device collects touch timing data and calculates the multi-view entropy vector.

[0176] Data processing module 200 performs time series detection on the entropy vector:

[0177] A sudden change was detected at a timescale (window = 20 points) of t1 = 100 min.

[0178] The mutation time t2 = 102 min was detected on a long time scale (window = 200 points);

[0179] If the time difference between the two is less than the time alignment threshold, it is determined to be a true change point, and the change point intensity is calculated. =0.65.

[0180] Compare the entropy vector with the baseline:

[0181] First baseline (historical average health value of this device): 0.4.

[0182] Second baseline (average health value of the same model of equipment): 0.35.

[0183] Difference from the first baseline The difference between the second baseline and the second baseline .

[0184] Take the maximum value .

[0185] Anomaly rating: λ=0.65.

[0186] Since 0.845 < 0.95 (first threshold), full calibration was not triggered.

[0187] Unsupervised mapping of the health index HI: 0.76 (out of 1.0, the higher the score, the healthier).

[0188] The degradation trend decreases approximately linearly, and it is expected to drop to the fault threshold of 0.3 at 580 minutes. The current time is 200 minutes, so RUL=380 minutes.

[0189] Since HI=0.76>0.3 (not exceeding the threshold);

[0190] RUL=380min>300min (not exceeding the threshold).

[0191] Therefore, short-range calibration was not triggered.

[0192] Degradation score calculation:

[0193] Assume the entropy change rate is 0.015 / min; the retry rate is 4%; the number of process interruptions is 2 times / hour; and the degradation acceleration is 0.003 / min².

[0194] If the weighted degradation score exceeds a threshold, a rapid self-check command is triggered. After the device completes the rapid self-check and confirms its health, the system incorporates the current entropy vector into historical data and updates the first baseline. Simultaneously, it periodically aggregates data from all devices of the same model across the hospital and updates the second baseline.

[0195] Example 2

[0196] like Figure 7 As shown, this embodiment introduces a self-learning dynamic calibration method for medical metrology nodes, which is executed by a central processing unit and includes the following steps:

[0197] S701. Obtain the multi-view entropy vector of the target detection device. The multi-view entropy vector is calculated by the edge device based on touch timing data.

[0198] S702. Preprocess the multi-view entropy vector, perform change point detection on the processed multi-view entropy vector, and then calculate the change point intensity;

[0199] S703. Calculate the difference between the preprocessed multi-view entropy vector and the first baseline and the second baseline respectively, and take the maximum difference and the change point intensity to perform weighted calculation to obtain the anomaly score; where the first baseline is the reference value of the target detection device under normal health conditions, and the second baseline is the reference mean of all devices of the same model as the target detection device under normal health conditions.

[0200] S704. Determine whether the abnormal score is greater than or equal to the first threshold. If yes, issue a full calibration command; otherwise, map the processed multi-view entropy vector to consumable health indicators in an unsupervised manner, and fit the health indicators to the remaining service life of the consumables.

[0201] S705. Determine whether the health indicators and remaining service life of the consumables are less than the corresponding preset thresholds. If one or both are met, issue a short-range calibration command. If neither is met, issue a delayed calibration command.

[0202] This embodiment has the same beneficial effects as Embodiment 1.

[0203] Example 3

[0204] like Figure 8 As shown, this embodiment introduces a self-learning dynamic calibration method for medical metrology nodes, which is executed by an edge device and includes the following steps:

[0205] S801. Collect operational behavior data, metering signal data, consumable data, and test result data of the target detection equipment;

[0206] S802. After preprocessing the collected data, calculate the information entropy, permutation entropy, sample entropy and Markov transition entropy respectively;

[0207] S803. Combine the information entropy, permutation entropy, sample entropy, and Markov transition entropy into a multi-view entropy vector and send it to the central processing unit.

[0208] S804. After the central processing unit generates a calibration command based on the multi-view entropy vector, it receives and executes the calibration command, and sends back the operation log after execution.

[0209] In application, multi-view entropy vectors can be encapsulated in JSON format, and the central processing device can use a random forest classifier to parse the vectors to generate calibration instructions. Operation log feedback can be achieved using the MQTT protocol.

[0210] Example 4

[0211] This embodiment introduces a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the aforementioned medical metrology node self-learning dynamic calibration method.

[0212] When applying the medical metrology node self-learning dynamic calibration method of Example 2 or Example 3, it can be applied in the form of software, such as a program designed to run independently on a computer-readable storage medium, which can be a USB flash drive, designed as a USB security token, and designed to start the entire method through an external trigger.

[0213] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A medical metrology node self-learning dynamic calibration system, characterized in that, The method comprises the following steps: a data acquisition module is configured to acquire a multi-view entropy vector of a target detection device, wherein the multi-view entropy vector is calculated by an edge device based on touch timing data; a data processing module is configured to pre-process the multi-view entropy vector, detect a change point of the pre-processed multi-view entropy vector, and calculate a change point intensity based on a change trend; when the data processing module detects the change point and calculates the change point intensity, the data processing module performs first time scale detection and second time scale detection on the pre-processed multi-view entropy vector respectively, both of which detect a change point, and the time difference is not more than a time alignment threshold, so that the change point is determined as a true change point, and the change point intensity is calculated; an abnormal score calculation module is configured to calculate a difference between the pre-processed multi-view entropy vector and a first baseline and a second baseline respectively, take a maximum difference, and perform weighted calculation on the maximum difference and the change point intensity to obtain an abnormal score, wherein the first baseline is a reference value of the target detection device in a normal and healthy state, and the second baseline is a reference mean value of all devices of the same type as the target detection device in the normal and healthy state; a first determination module is configured to determine whether the abnormal score is greater than or equal to a first threshold, and if yes, an all-amount calibration instruction is issued; otherwise, the pre-processed multi-view entropy vector is mapped into a consumable health index in an unsupervised manner, and a health index degradation is fitted into a consumable remaining service life; the first determination module is further configured to determine whether the change point intensity is greater than a strong change point threshold and a quality control sample deviation is more than a preset allowable range, and if yes, the all-amount calibration instruction is issued; wherein the quality control sample deviation is calculated by the data processing module based on a deviation between an actual value and a nominal value of a quality control sample acquired by the data acquisition module; a second determination module is configured to determine whether the consumable health index and the consumable remaining service life are less than corresponding preset thresholds, and if one or both of them are satisfied, a short-range calibration instruction is issued, and if neither of them is satisfied, a delayed calibration instruction is issued; a baseline updating module is configured to dynamically update the first baseline and the second baseline, and the baseline updating module comprises: a first baseline updating unit is configured to add a current multi-view entropy vector into a historical data set when the target detection device returns to normal after executing the all-amount calibration instruction or the short-range calibration instruction, and recalculate a reference value as the first baseline; 2. The metrology node self-learning dynamic calibration system of claim 1, wherein, a second baseline updating unit is configured to periodically acquire reference values of all devices of the same type as the target detection device in the normal and healthy state, and calculate a mean value of the reference values as the second baseline. After the second determination module determines that the consumable health index and the consumable remaining service life are both greater than the corresponding preset thresholds, the second determination module further determines whether a degradation score is greater than a preset degradation threshold, and if yes, a rapid self-check instruction is issued, and if not, the delayed calibration instruction is issued; wherein the degradation score is obtained by the data processing module by analyzing and calculating an entropy change rate of the multi-view entropy vector, combining a retry rate and a process interruption number of the target detection device, and performing weighted calculation on a degradation acceleration, and the retry rate and the process interruption number of the target detection device are acquired by the data acquisition module.

3. The metrology node self-learning dynamic calibration system of claim 1, wherein, The first baseline and the second baseline are adjusted correspondingly based on a sample type detected by the target detection device.

4. The metrology node self-learning dynamic calibration system of claim 1, wherein, The data acquisition module has an adaptive adjustment of the collection frequency according to a frequency of sample detection by the target detection device.

5. The calibration method of a medical metrology node self-learning dynamic calibration system according to any of claims 1-4, characterized in that, It comprises the following steps: obtaining a multi-view entropy vector of the target detection device, the multi-view entropy vector being calculated by the edge device based on touch timing data; preprocessing the multi-view entropy vector, performing change point detection on the processed multi-view entropy vector, and then calculating a change point strength; calculating a difference between the preprocessed multi-view entropy vector and a first baseline and a second baseline, respectively, taking a maximum difference, and performing weighted calculation on the maximum difference and the change point strength to obtain an abnormal score; wherein the first baseline is a reference value of the target detection device in a normal and healthy state, and the second baseline is a reference mean value of all devices of the same model as the target detection device in a normal and healthy state; judging whether the abnormal score is greater than or equal to a first threshold value, and if yes, issuing a full-scale calibration instruction; otherwise, mapping the processed multi-view entropy vector to a consumable health index in an unsupervised manner, and fitting the health index degradation to a consumable remaining service life; judging whether the consumable health index and the consumable remaining service life are less than corresponding preset threshold values, and if one or both of them are satisfied, issuing a short-range calibration instruction, and if neither of them is satisfied, issuing a delayed calibration instruction.

6. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the calibration method of claim 5.

Images