CN114548259B - A PISA fault identification method based on semi-supervised Semi-KNN model - Google Patents

Abstract

The invention relates to a PISA fault identification method based on the semi-supervised Semi-KNN model, which includes: S10. Obtain the blood glucose information to be measured within a preset time period and preprocess it to obtain the preprocessed blood glucose information to be measured; S20. Based on The pre-established PISA constraint set and the pre-processed blood glucose information to be measured are used to obtain the constraint relationship using similarity measurement processing; S30, input the pre-processed blood glucose information to be measured and the constraint relationship into the pre-trained semi-supervised Semi-KNN In the model, the semi-supervised Semi-KNN model outputs the classification results of the blood glucose information to be measured; the semi-supervised Semi-KNN model uses the training data set and the PISA constraint set to train the KNN model, and the semi-supervised model is used to identify abnormal blood glucose information. way model. The above method improves the reliability of blood glucose information detection, improves the accuracy of fault diagnosis results, and improves processing efficiency.

Description

A PISA fault identification method based on semi-supervised Semi-KNN model

Technical Field

The present invention relates to a PISA fault recognition technology, and in particular to a PISA fault recognition method based on a semi-supervised Semi-KNN model.

Background Art

In recent years, continuous glucose monitoring systems (CGM) have received more and more attention. Continuous glucose monitoring signals are used as an auxiliary means to diagnose and guide various types of diabetes. Continuous glucose monitoring signals are usually analyzed using data-driven methods, which have problems such as blood glucose signals being easily affected by noise, blood glucose monitors being prone to failure, and blood glucose prediction alarms being affected by data errors and having low accuracy. Most fault identification methods for continuous glucose monitoring systems are plagued by poor performance and high false positive rates, which limits the clinical use of the auxiliary.

Digital signal processing has developed rapidly in recent years, and the noise problem of CGM signals has been solved through finite and infinite impulse response filters. However, CGM fault detection is still a challenge that needs attention and is a very active research application field.

When the skin around the sensor used in the continuous glucose monitoring system CGM is under great pressure, the CGM reading will drop rapidly. Algorithms based on CGM readings, such as predictive pump shut-off, rely on the estimation of the rate of change of the glucose sensor to shut down the insulin pump to avoid hypoglycemia. However, since PISA events cannot be paid attention to in time when they occur at night, inappropriate pump shutdown will occur; in addition, the prediction algorithm will also cause low prediction data due to PISA failure, and have a more serious impact on forecasts and warnings. Therefore, how to distinguish PISA failures from other signal value low events such as insulin events, hypoglycemia events, and exercise events has become a technical problem that needs to be solved urgently. Therefore, a semi-supervised PISA fault identification method with a fast enough execution time for real-time operation is needed.

Summary of the invention

1. Technical issues to be resolved

In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a PISA fault identification method based on a semi-supervised Semi-KNN model.

(II) Technical solution

In order to achieve the above object, the main technical solutions adopted by the present invention include:

In a first aspect, an embodiment of the present invention provides a PISA fault identification method based on a semi-supervised Semi-KNN model, which includes:

S10, obtaining blood sugar information to be measured within a preset time period, and preprocessing the blood sugar information to be measured to obtain the preprocessed blood sugar information to be measured;

S20, based on the pre-established PISA constraint set and the pre-processed blood glucose information to be measured, a similarity measurement processing method is used to obtain the constraint relationship to which the blood glucose information to be measured belongs;

The PISA constraint set is a set with ML constraints and CL constraints constructed based on prior knowledge in the training phase of the semi-supervised Semi-KNN model, and each element in the set is information on the first-order differential features of the blood glucose subsequence;

S30, inputting the pre-processed blood glucose information to be measured and the constraint relationship into a pre-trained semi-supervised Semi-KNN model, and the semi-supervised Semi-KNN model outputs a classification result of the blood glucose information to be measured;

The semi-supervised Semi-KNN model is a semi-supervised model for identifying abnormal blood glucose information obtained by training the KNN model using a training data set and the PISA constraint set, and the training data set includes blood glucose data processed by first-order difference.

Optionally, before S10, the method further includes:

S01. Acquire multiple historical blood glucose data with the help of a CGM device, and pre-process each historical blood glucose data to obtain a blood glucose sequence; each blood glucose sequence includes blood glucose data with a PISA timestamp label and blood glucose data without a PISA timestamp label;

S02, dividing each blood glucose sequence into multiple subsequences, and performing first-order difference calculation on each subsequence to obtain a training data set;

S03, based on prior knowledge and training data with PISA timestamp labels in the training data set, forming rules according to semi-supervised constraints to generate a PISA constraint set;

S04, training the semi-supervised Semi-KNN model using the training data set and the PISA constraint set to obtain a trained semi-supervised Semi-KNN model;

The semi-supervised Semi-KNN model is constructed by improving the KNN model in a semi-supervised manner.

Optionally, the S04 includes:

S04-1, traverse all subsequences of the training data set, construct an offline K-dimensional search binary tree, and obtain a K-D tree;

S04-2. Based on the K-D tree, traverse the PISA constraint set to obtain an abnormal threshold σ of the semi-supervised Semi-KNN model, where the boundary thresholds of the abnormal threshold are σ1 and σ2, expressed as σ=[σ1, σ2];

Among them, the DTW similarity measurement function is used to calculate the average distance between each PISA event and other events in the PISA constraint set to obtain a distance set;

Then, according to the following formula (1), the abnormal threshold σ = [σ1, σ2] is obtained;

σ1＝Q3+1.5(Q3-Q1), formula (1)

σ2＝Q1-1.5(Q3-Q1),

If the sample distance between the blood glucose data to be tested and the ML relationship in the PISA constraint is dist<σ2, it is determined to be an abnormal sample; if the sample distance between the blood glucose data to be tested and the CL relationship in the PISA constraint is dist>σ1, it is determined to be an abnormal sample;

Q3 is the upper quartile in the distance set, and Q1 is the lower quartile in the distance set.

Optionally, the S02 includes:

Perform sliding window processing on each blood glucose sequence. In the blood glucose sequence X = {x1, x2, ..., xn}, a number of subsequences qi = { _xi , xi ₊₁ , ..., xi _+k } are formed after sliding window of size w.

A sequence subset is D = {q1, q2, ..., qm}. For each subsequence qi, the first-order difference is calculated according to formula (2).

h is the change in the first-order difference formula, and the value of h is 0.8-1.2;

After calculating the first-order differences for all subsequences, the first-order difference value of each subsequence is used as the training data set.

Optionally, the S10 includes:

Obtain blood sugar information for 30-45 minutes or more with the help of CGM equipment;

Filtering is performed, and the blood glucose information to be measured is preprocessed by a sliding window method to remove isolated noise points in the blood glucose information to be measured and to fill missing values, so as to obtain a blood glucose sequence to be measured of the blood glucose information to be measured. In other words, the blood glucose information to be measured can be filtered, and the blood glucose information to be measured can be traversed by a sliding window of appropriate size, and when such tiny areas exist within the sliding window range, they are averaged to remove quasi-isolated noise points of the blood glucose information to be measured. When there are missing values in the blood glucose information to be measured due to sensor problems, the number of missing values that exist continuously can be determined first, and then the missing values of the blood glucose information to be measured can be filled by a general linear interpolation method to obtain the blood glucose sequence to be measured after preprocessing of the blood glucose information to be measured.

Optionally, the S20 includes:

When the SBD distance between the blood glucose data A of each sequence in the blood glucose sequence to be measured and a PISA event B in the PISA constraint set is less than the threshold λ, that is, f _SBD (A, B) <λ, a constraint relationship ML (A, B) is determined, and the PISA constraint set is updated; λ is a preset value greater than 0; f represents a function for calculating the SBD distance between two sequences;

When the SBD distance between the blood glucose data A of each sequence in the blood glucose sequence to be measured and the CL constraint relationship of a PISA event B in the PISA constraint set is less than a threshold λ, that is, f _SBD (A, B) <λ, a constraint relationship CL (A, B) is determined, and the PISA constraint set is updated;

Each sequence in the blood glucose sequence to be tested is traversed, and the updated PISA constraint set is used as the constraint relationship to which the blood glucose information to be tested belongs.

Optionally, the S30 includes:

Based on the K-D tree and the blood glucose sequence to be measured, the K data points closest to each data in the blood glucose sequence to be measured are obtained in a cyclic iteration manner, and the K-D tree of the use stage is obtained.

Based on the K-D tree of the use phase, the constraint relationship is traversed to obtain the classification results of the PISA abnormal information of the blood glucose sequence to be tested;

The DTW similarity measurement function is used to calculate the actual distance between each PISA event and other events in the constraint relationship, and the actual distance is compared with the abnormal threshold σ=[σ1, σ2] to obtain the classification results of PISA events and non-PISA events.

Optionally, after comparing the actual distance with the abnormal threshold, the amount of data belonging to the ML constraint in the constraint relationship is determined, and the abnormal level value in the PISA event is determined based on the data amount.

In a second aspect, an embodiment of the present invention further provides an electronic device, comprising: a memory and a processor, wherein the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory and execute the steps of the PISA fault identification method based on the semi-supervised Semi-KNN model as described in any one of the first aspects above.

In a third aspect, an embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the PISA fault identification method based on the semi-supervised Semi-KNN model as described in any one of the first aspects above are implemented.

(III) Beneficial effects

The method of the embodiment of the present invention performs anomaly classification based on the Semi-KNN model, which solves the uncertainty problem of the KNN model when performing anomaly detection; for the first time, prior knowledge is introduced in the form of constraints to form a semi-supervised anomaly detection method, which maximizes the use of effective prior knowledge and improves the reliability of the detection results; by grading the results, the credibility of the results is improved, thereby helping doctors to make clinical judgments.

The embodiments of the present invention propose for the first time to perform CGM sensor fault diagnosis by a semi-supervised method, and improve the accuracy of the fault diagnosis result by introducing prior knowledge (such as expert experience); compared with the effect of traditional unsupervised fault identification methods applied in the field of CGM sensor fault diagnosis, the detection result of the present invention has higher accuracy, and the PISA constraint set introduced into the semi-supervised model can ensure a high recognition rate for PISA faults and confidence in the detection of uncalibrated anomalies (such as PISA abnormal events occurring at night).

In addition, the semi-supervised model proposed in the present invention can update the distance measurement method from the original Euclidean distance to the DTW and SBD similarity measurement methods for time series data types, thereby improving the measurement accuracy of the time series data type and accelerating the running speed of the entire calculation program.

The method of this embodiment can be applied in a continuous glucose monitoring system CGM to enhance CGM data. Fault detection not only enhances the safety of CGM, but also avoids the reduction of the credibility of tasks such as changes in treatment plans or forecast warnings due to faults. The CGM using the method of the present invention detects pressure-sensitive sensor attenuation (PISA) false signals to improve the confidence of detection.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a PISA fault identification method based on a semi-supervised Semi-KNN model provided by an embodiment of the present invention;

FIG2(a) is a schematic diagram of a process of constructing a sample of a k-dimensional search binary tree K-d tree;

Figure 2(b) Schematic diagram of K-d tree;

FIG3 is a representation diagram of a new sample;

FIG4 is a schematic diagram showing the guiding effect of constraint relationships on the iterative process of anomaly detection using KNN;

FIG5 is a flow chart of a PISA fault identification method based on a semi-supervised Semi-KNN model provided by another embodiment of the present invention.

DETAILED DESCRIPTION

In order to better explain the present invention and facilitate understanding, the present invention is described in detail below through specific implementation modes in conjunction with the accompanying drawings.

Embodiment 1

As shown in FIG. 1 , FIG. 1 shows a flow chart of a PISA fault identification method based on a semi-supervised Semi-KNN model. The execution subject of the method may be any computer/electronic device/CGM. The method may include the following steps:

S10, obtaining blood sugar information to be measured within a preset time period, and preprocessing the blood sugar information to be measured to obtain the preprocessed blood sugar information to be measured.

In this embodiment, CGM can be used to obtain blood glucose information to be tested for greater than or equal to 30-45 minutes; the blood glucose information to be tested is filtered and pre-processed through a sliding window method to remove isolated noise points in the blood glucose information to be tested and fill in missing values, thereby obtaining a blood glucose sequence to be tested of the blood glucose information to be tested.

It should be noted that the time period of the blood sugar information to be measured is adjustable according to the parameter value of the sliding window in the preprocessing.

S20, based on the pre-established PISA constraint set and the pre-processed blood glucose information to be measured, a similarity measurement processing method is used to obtain the constraint relationship to which the blood glucose information to be measured belongs;

The PISA constraint set is a set with ML constraints and CL constraints constructed based on prior knowledge in the training phase of the semi-supervised Semi-KNN model, and each element in the set is the information of the first-order difference feature of the blood glucose subsequence. The PISA constraint set must contain PISA information, that is, the blood glucose data in the training phase is preprocessed according to the PISA timestamp label, and then the PISA constraint set is created.

S30, inputting the pre-processed blood glucose information to be measured and the constraint relationship into a pre-trained semi-supervised Semi-KNN model, and the semi-supervised Semi-KNN model outputs a classification result of the blood glucose information to be measured;

The semi-supervised Semi-KNN model is a semi-supervised model for identifying abnormal blood glucose information obtained by training the KNN model using a training data set and the PISA constraint set, and the training data set includes blood glucose data processed by first-order difference. It is understandable that the training data set is a data set obtained by performing first-order difference processing on the subsequence set of the blood glucose data obtained in the training phase after being processed by Huachuang.

The method of this embodiment performs anomaly classification based on the Semi-KNN model, which solves the uncertainty problem of the KNN model when performing anomaly detection; for the first time, prior knowledge is introduced in the form of constraints to form a semi-supervised anomaly detection method, which maximizes the use of effective prior knowledge and improves the reliability of the detection results; by grading the results, the credibility of the results is improved, thereby helping doctors to make clinical judgments.

In practical applications, before the above step S10, the method shown in FIG1 may further include the following steps which are not shown in the figure:

S01. Acquire multiple historical blood glucose data by means of a continuous blood glucose monitoring system CGM (i.e., a CGM device), and pre-process each historical blood glucose data to obtain a blood glucose sequence; each blood glucose sequence includes blood glucose data with a PISA timestamp label and blood glucose data without a PISA timestamp label;

S02, dividing each blood glucose sequence into multiple subsequences, and performing first-order difference calculation on each subsequence to obtain a training data set;

For example, the S02 may include:

Perform sliding window processing on each blood glucose sequence. In the blood glucose sequence X = {x1, x2, ..., xn}, a number of subsequences qi = { _xi , xi ₊₁ , ..., xi _+k } are formed after sliding window of size w.

A sequence subset is D = {q1, q2, ..., qm}. For each subsequence qi, the first-order difference is calculated according to formula (2).

h is the change in the first-order difference formula, and the value of h is 0.8-1.2, preferably 1;

After calculating the first-order differences for all subsequences, the first-order difference value of each subsequence is used as the training data set.

S03, based on prior knowledge and training data with PISA timestamp labels in the training data set, forming rules according to semi-supervised constraints to generate a PISA constraint set;

S04, training the semi-supervised Semi-KNN model using the training data set and the PISA constraint set to obtain a trained semi-supervised Semi-KNN model;

The semi-supervised Semi-KNN model is constructed by improving the KNN model in a semi-supervised manner.

For example, S04 may include:

S04-1, traverse all subsequences of the training data set, construct an offline K-dimensional search binary tree, and obtain a K-D tree;

S04-2. Based on the K-D tree, traverse the PISA constraint set to obtain an abnormal threshold σ of the semi-supervised Semi-KNN model, where the boundary thresholds of the abnormal threshold are σ1 and σ2, expressed as σ=[σ1, σ2];

Among them, the DTW similarity measurement function is used to calculate the average distance between each PISA event and other events in the PISA constraint set to obtain a distance set;

Then, according to the following formula (1), the abnormal threshold σ = [σ1, σ2] is obtained;

σ1＝Q3+1.5(Q3-Q1), formula (1)

σ2＝Q1-1.5(Q3-Q1),

If the sample distance between the blood glucose data to be tested and the ML relationship in the PISA constraint is dist<σ2, it is determined to be an abnormal sample; if the sample distance between the blood glucose data to be tested and the CL relationship in the PISA constraint is dist>σ1, it is determined to be an abnormal sample;

Q3 is the upper quartile in the distance set, and Q1 is the lower quartile in the distance set.

To better understand the above step S20, the step S20 can be specifically described as follows:

When the SBD distance between the blood glucose data A of each sequence in the blood glucose sequence to be measured and a PISA event B in the PISA constraint set is less than a threshold value λ, that is, f _SBD (A, B) <λ, a constraint relationship ML (A, B) is determined, and the PISA constraint set is updated; λ is a preset value greater than 0; f _SBD represents a function for calculating the SBD distance between two sequences;

When the SBD distance between the blood glucose data A of each sequence in the blood glucose sequence to be measured and the CL constraint relationship of a PISA event B in the PISA constraint set is less than a threshold λ, that is, f _SBD (A, B) <λ, a constraint relationship CL (A, B) is determined, and the PISA constraint set is updated;

Each sequence in the blood glucose sequence to be tested is traversed, and the updated PISA constraint set is used as the constraint relationship to which the blood glucose information to be tested belongs.

Accordingly, the above step S30 may include:

Based on the K-D tree and the blood glucose sequence to be measured, the K data points closest to each data in the blood glucose sequence to be measured are obtained in a cyclic iteration manner, and the K-D tree of the use stage is obtained.

Based on the K-D tree of the use phase, the constraint relationship is traversed to obtain the classification results of the PISA abnormal information of the blood glucose sequence to be tested;

The DTW similarity metric function is used to calculate the actual distance between each PISA event and other events in the constraint relationship, and the actual distance is compared with the abnormal threshold σ = [σ1, σ2] to obtain the classification results of PISA events and non-PISA events. In particular, after comparing the actual distance with the abnormal threshold, the amount of data belonging to the ML constraint in the constraint relationship is determined, and the abnormal level value of the PISA event is determined based on the data amount.

The method of this embodiment can be integrated into an electronic device such as an anomaly detector, which can identify abnormal problems of PISA. Therefore, the above-mentioned anomaly detector is used in clinical emergency patient care to effectively monitor and reliably quantify the patient's abnormal state, thereby solving the lag in the existing technology and realizing real-time monitoring and analysis.

This embodiment proposes for the first time to perform CGM sensor fault diagnosis using a semi-supervised method, and by introducing prior knowledge (such as expert experience), the accuracy of the fault diagnosis result is improved, and the uncertainty problem of the KNN algorithm when performing anomaly detection is solved.

Embodiment 2

The method of this embodiment can be described in detail in the order of a preparation stage, a training stage, and a use stage, as shown in FIG. 5 .

1. Preparation stage - historical CGM blood glucose data acquisition and preprocessing

The continuous glucose monitoring system CGM (Continuous Glucose Monitoring) is one of the key components of the artificial pancreas. Through this device, the patient's blood glucose level can be continuously monitored, thereby helping patients with type 1 diabetes (T1DM) maintain blood glucose concentration within a safe range.

1.1CGM blood glucose data acquisition

CGM indirectly reflects blood sugar levels by monitoring the glucose concentration of the interstitial fluid of the subcutaneous tissue through a glucose sensor, and can provide continuous, comprehensive and reliable blood sugar information throughout the day. The historical blood sugar data obtained should include three complete days of data, in which blood sugar values are collected every five minutes, for a total of 3*288 blood sugar values. In addition to normal physiological activities such as meals, exercise, and sleep, it should include several PISA typical fault information obtained by experimental pressing. Each PISA typical fault information can include an accurate pressing label for establishing the subsequent semi-supervised Semi-KNN model.

1.2 CGM blood glucose data preprocessing

The historical blood glucose data collected by the above-mentioned device, i.e., CGM, is stored in a storage device and can be pre-processed by means of digital signal analysis, including filtering, missing value filling, labeling, etc. Among them, filtering is to remove isolated noise points on the CGM blood glucose sequence. Isolated noise points mean data that deviate too much from the blood glucose values before and after that moment. A sliding window is used to traverse the sequence. When such tiny areas exist within the sliding window range, they are averaged. This step is an optional step. The filtered blood glucose sequence is smoother and convenient for subsequent processing. Missing value filling is to prevent the existence of blood glucose value vacancies. Usually, the blood glucose sequence used is the blood glucose sequence after missing value filling. Otherwise, it is easy to have a fault in the blood glucose curve in the subsequent processing, making the model output result inaccurate.

In actual processing, the filled blood glucose sequence also needs to be marked. The experimental interval segments with pressing actions are marked, and the rest of the time is not marked. This is used to distinguish PISA events from other unknown events. In this way, the accuracy of the semi-supervised Semi-KNN model can be effectively guaranteed.

The above is an explanation of the preprocessing process of blood glucose data collected by CGM. After preprocessing, the continuously collected CGM blood glucose sequence containing PISA compression experiment events is filtered, the missing value filled blood glucose sequence and PISA timestamp label are obtained.

2. Preparation phase - constructing features and adding initial seeds based on prior knowledge

It can be seen from the PISA event time period pressed in the experiment that part of the data in the blood glucose sequence can be labeled. The PISA timestamp label obtained by the above preprocessing, such as the 9:00-9:45 interval, is used to obtain 9 blood glucose values in the blood glucose sequence based on the PISA timestamp label as the blood glucose values when the PISA event occurs. The corresponding first-order differential features also correspond to the PISA event features.

2.1 Structural features

Usually, there are many blood sugar data in the blood sugar sequence, and it is impossible to use a few numerical values to express the characteristics of the entire sequence. In this embodiment, a sliding window method is used to process the blood sugar sequence. Specifically, in the blood sugar sequence X = {x1, x2, ..., xn}, a number of subsequences qi = { _xi , xi ₊₁ , ..., xi _+k } are formed after a sliding window of size w. A sequence subset is defined as D = {q1, q2, ..., qm}. A first-order difference calculation is performed on each subsequence qi. The calculation formula (1) is as follows, where h is the change in the first-order difference formula. In this embodiment, h takes a value of 0.8 to 1.2 in the blood sugar sequence feature construction, and preferably takes 1;

After calculating the first-order difference for all subsequences, it is used as the input of the subsequent model, namely the semi-supervised Semi-KNN model. In this embodiment, the first-order difference feature is used to avoid the impact of the different forms of the original blood glucose sequence on the universality of the model. On the other hand, the first-order difference feature has a good inhibitory effect on the volatility of the time series, which can make the input of the semi-supervised Semi-KNN model simpler.

2.2 Adding initial seeds based on expert experience (prior knowledge)

The representative samples of several data points in the dataset of the current semi-supervised operation are called seeds, which can be expressed as sample points themselves or in the form of constraints. Pairwise constraints consist of must-link (ML) and cannot-link (CL). The two data points in the ML constraint must be in the same cluster, while the two data points declared in the CL constraint must be in different clusters.

In this embodiment, the known PISA event labels are defined as the initial seeds of the semi-supervised model, and CL constraints are constructed with blood sugar fluctuations caused by other physiological events (such as eating and exercising), and different PISA events constitute ML constraints.

In this embodiment, features are constructed for the blood glucose sequence and an initial seed portion is added, that is, after preprocessing the originally collected blood glucose sequence and the PISA timestamp label, the first-order differential features of the blood glucose sequence and the pairwise constraints included in PISA are obtained.

That is, after the above preprocessing, a training data set and a PISA constraint set are obtained.

3. Training phase - training semi-supervised Semi-KNN model

The preprocessed training data set and the pairwise constraints contained in the PISA are input into the Semi-KNN model to obtain a semi-supervised fault detection model;

3.1 Basic Model KNN

The existing KNN is a non-parametric supervised classifier that determines the category of the test sample by the majority class of the closest training sample. Taking the binary classification problem under the blood glucose sequence as an example, the formal definition of the binary classification problem and the solution process is as follows: the sample set of the known blood glucose sequence is S = (x ₁ , y ₁ ), (x ₂ , y ₂ ), ..., (x _N , y _N ), where x _i ∈ R ² is a point in the blood glucose measurement value, and y _i ∈ {c ₁ , c ₂ } represents the category to which the blood glucose value sample belongs. For a new blood glucose value sample x, the category y of the blood glucose value sample can be solved using formula (2):

Where N _K (x) represents the set of K samples closest to the blood glucose value sample x, and f is the indicator function about _yi

The problems with using supervised binary classification algorithms for anomaly detection are: 1) the ratio of abnormal samples to normal samples is extremely unbalanced, resulting in imbalanced classification results; 2) there are many types of abnormal samples, and classifying unknown anomalies and known anomalies into one category is not explanatory, and simple binary classification cannot meet the needs; 3) the cost of data labeling is too high. Therefore, semi-supervised models that only require a small amount of prior knowledge are more suitable for application needs.

3.2 Semi-KNN Model

In this embodiment, the semi-supervised anomaly detection model is more suitable for the identification of PISA faults. The semi-supervised model is based on the basic model, and the label input is changed to the partial PISA constraint set input obtained with prior knowledge. The data object is the preprocessed CGM blood glucose sequence data.

3.2.1 Traverse all data objects in the training dataset and build an offline K-dimensional search binary tree for the training dataset

By fitting the samples in the training data set, a k-dimensional search binary tree K-d tree (K-D tree) is constructed. The specific construction process is as follows. Assume a simple two-dimensional sequence data: {(2,3), (4,7), (5,4), (7,2), (8,1), (9,6)}. First, the variance of the data in the x and y directions is calculated respectively. It is found that the variance in the x direction is larger, so the partition domain is determined to be the x-axis direction; secondly, the median value is 7 according to the x-axis value 2, 5, 9, 4, 8, 7, so the splitting hyperplane is (7,2) and perpendicular to the x-axis; then the left subspace and the right subspace are determined. The splitting hyperplane divides the entire space into two parts, as shown in Figure 2 (a). The left subspace contains 3 nodes {(2,3), (4,7), (5,4)}; the right subspace contains 2 nodes {(8,1), (9,6)}, and then continue to recurse until each space contains only one data point, generating the final K-d tree, as shown in Figure 2 (b).

3.2.2 Traversing the PISA constraint set and determining the PISA anomaly threshold

The above training data set contains PISA event data caused by experimental pressing. Therefore, by traversing the PISA constraint set, the PISA abnormalities in the training data set can be distinguished from other normal physiological events. Therefore, the abnormality threshold σ of the semi-supervised Semi-KNN model can be obtained. The boundary thresholds of the abnormality threshold are σ1 and σ2, expressed as σ=[σ1, σ2].

Calculate the average distance between each PISA event and other normal physiological events. This distance usually uses the Euclidean distance. However, in order to fit the blood glucose sequence data type and better express the shape similarity of the time series, the DTW similarity metric function is used here, which is warped under the time axis to achieve a better alignment effect. After obtaining all the average distances, calculate the upper quartile Q3 and lower quartile Q1 of the distance set. Then the abnormal threshold of the model is σ1 = Q3 + 1.5 (Q3-Q1), σ2 = Q1-1.5 (Q3-Q1). When the distance from the normal sample is dist>σ1 or dist<σ2, the sample is considered to be an abnormal sample.

That is to say, if the sample distance between the blood glucose data to be tested and the ML relationship in the PISA constraint is dist<σ2, it is determined to be an abnormal sample; if the sample distance between the blood glucose data to be tested and the CL relationship in the PISA constraint is dist>σ1, it is determined to be an abnormal sample.

4. Use phase - get new blood glucose data and perform constraint propagation

After the semi-supervised Semi-KNN model is trained, the real-time collected blood glucose data can be tested. When the newly received blood glucose sequence contains PISA failure events (i.e., non-PISA events), accurate identification can be achieved.

4.1 Blood glucose data to be measured

Obtain the continuous blood glucose values to be analyzed that are within at least 45 minutes, that is, 9 blood glucose measurement points, and perform preprocessing, such as filtering, and use the sliding window method to remove isolated noise points in the blood glucose information to be measured and fill in missing values to obtain the blood glucose sequence to be measured. 4.2 Constraint Propagation

The constraint propagation algorithm is executed on the blood glucose sequence to be tested, with the aim of expanding the PISA constraint set in the original training data set, thereby assisting judgment when using the semi-supervised Semi-KNN model and making the results more accurate.

The specific implementation process is as follows:

4.2.1 Traverse the PISA constraint set in the original training data set and calculate the SBD distance

When a sequence subset is D = {q1, q2, ..., qm}, any qi in D regards other sequences as nearest neighbors. The nearest neighbor means that the distance from the current sequence to other subsequences ql is less than the threshold λ, which can usually be artificially defined. The above pairwise constraints include the following properties:

1) For any q∈D, ML(q,r) can be generated, where r∈D;

2) Given ML(p, q) and (q, r), we can get ML(p, r);

3) Given ML(p, q), we can generate ML(c, p) and ML(p, r), where c∈D, r∈D;

4) Given CL(p, q), CL(c, p) and CL(p, r) can be generated, where c∈D, r∈D.

The distance metric selected in this embodiment is SBD distance. The SBD algorithm is a shape similarity metric based on cross-correlation. Its high efficiency and parameter-free characteristics are incomparable to the DTW metric. DTW is a measurement method with high accuracy but high computational cost. In addition, the accuracy of the SBD algorithm is close to that of the DTW algorithm, so SBD can be better used to measure the similarity between CGM curves and facilitate online similarity calculation.

4.2.2 Execution Constraint Propagation

When the SBD distance between the blood glucose data A of each sequence in the blood glucose sequence to be tested and a PISA event B in the training data set is less than the threshold λ, that is, f _SBD (A, B) <λ, a constraint relationship ML (A, B) can be determined, and the constraint set can be updated according to the above constraint propagation properties.

When the SBD distance between the blood glucose data A in each sequence of the blood glucose sequence to be tested and the CL constraint relationship of a PISA event B in the training data set is less than the threshold λ, that is, f _SBD (A, B) <λ, a constraint relationship CL(A, B) can be determined. Similarly, the constraint set can be updated by the above-mentioned constraint propagation property.

5. Use stage - use the semi-supervised Semi-KNN model to detect anomalies and determine the anomaly level

After executing step 4, the input of the semi-supervised Semi-KNN model is the blood glucose data of each sequence in the blood glucose sequence to be tested and the constraint relationship expanded to the PISA constraint set. The semi-supervised Semi-KNN model established in step 3.2.2 is used for anomaly detection, and the anomaly level is determined according to the constraint relationship. The specific execution steps are as follows:

5.1 Search for k nearest neighbors and calculate the average distance

The blood glucose data of each sequence in the blood glucose sequence to be tested (i.e., new blood glucose data) is input into the semi-supervised Semi-KNN model. Through the k-dimensional search binary tree established in step 3.2.1, the k data points closest to the blood glucose data of each sequence can be obtained. For example, based on the k-dimensional search binary tree sample established in the above step 3.2.1, it is assumed that the blood glucose data of each sequence is (2.1, 3.1), as shown in Figure 3. First, through binary search, find the nearest neighbor approximate point (2,3), and calculate the distance is 0.1414; then backtrack to the parent node (5,4), with (2.1,3.1) as the center and 0.1414 as the radius, and find that it does not intersect with the hyperplane of y=4, so there is no need to enter the right subspace search; then backtrack to the parent node (7,2), the circle does not intersect with the hyperplane of x=7, so there is no need to enter the right subspace search of (7,2); so far, the backtracking is completed, and the nearest sample is (2,3). After k iterations, the nearest k sample points can be obtained, which are called k nearest neighbors. After the k nearest neighbors are obtained, the average distance with each sequence of blood glucose data is calculated, and the distance is used as the score for abnormal judgment. When the score meets the threshold of 3.2.2, the sample point is regarded as a PISA failure event.

It should be noted that during the iteration process, when the new sample has a constraint relationship with PISA, the k-nearest neighbor samples should not contain CL relationship. That is, if the distance between the sample with CL relationship and the new sample meets the requirement of k-nearest neighbor during the iteration process, the sample should be discarded and the iteration should continue. The guiding role of the constraint relationship on the iterative process of KNN anomaly detection can be represented by Figure 4.

5.2 Output abnormal level according to constraint relationship

While the above steps output the anomaly detection results, we can consider whether the PISA constraint relationship is included in the statistical k-nearest neighbors. That is, after obtaining the distance of each sequence of blood glucose data in the semi-supervised Semi-KNN model, in addition to comparing with the anomaly threshold, we also need to consider whether there is a constraint relationship, that is, the result obtained in the aforementioned constraint propagation stage.

The specific operation is as follows: when no constraint relationship is included, it is judged as abnormal level 1; when the ML relationship of the expanded PISA constraint set is included, a higher abnormal level 2, 3, ... n is assigned according to the number of ML relationships including the expanded PISA constraint set. The more ML constraints of the expanded PISA constraint set exist, the higher the abnormal level.

The aforementioned DTW (Dynamic Time Warping) algorithm is used to detect the similarity between two time series and stretch or compress the time series to align them as much as possible. In most cases, the two series have very similar shapes as a whole, but these shapes are not aligned on the x-axis. Before comparing the similarity, one (or both) of the series needs to be warped on the time axis to achieve better alignment. DTW is an effective way to achieve this warping distortion.

The SBD algorithm is a shape similarity measure based on cross-correlation. Its high efficiency and parameter-free characteristics are incomparable to the DTW measure. DTW is a measurement method with high accuracy but high computational cost. The accuracy of the SBD algorithm is close to that of the DTW algorithm, so SBD can be better used to measure the similarity between CGM curves and facilitate online similarity calculation.

The sliding inner product between two time series data is calculated by cross-correlation, which has native robustness to phase offset. For two given time series x = (x ₁ , x ₂ , …, x _m ) and y = (y ₁ , y ₂ , …, y _m ), and the corresponding phase difference s, the inner product of the two curves is as follows:

The standard cross-correlation NCC and distance metric SBD can be calculated as follows:

This embodiment experiments on the above similarity measurement methods. The experimental results show that the SBD similarity measurement algorithm has strong noise resistance and can effectively distinguish the waveform difference between the PISA fault event and the normal sequence, while the noise difference between other sequences can be effectively avoided; the DTW similarity measurement algorithm is sensitive to shape features and can amplify small differences in shape after the distance output is normalized; the Euclidean distance is sensitive to the amplitude of the blood glucose curve, that is, the absolute distance difference in the original sequence distribution can be fully expressed.

Embodiment 3

This embodiment also provides an electronic device, including: a memory and a processor; the processor is used to execute the computer program stored in the memory to implement the steps of the PISA fault identification method based on the semi-supervised Semi-KNN model described in any of the above-mentioned embodiments 1 and 2.

It should be noted that in the claims, any reference numerals placed between brackets shall not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention may be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In the claims enumerating several means, several of these means may be embodied by the same hardware. The use of the words first, second, third, etc., is for convenience of expression only and does not indicate any order. These words may be understood as part of the component name.

In addition, it should be noted that, in the description of this specification, the description of the terms "one embodiment", "some embodiments", "embodiment", "example", "specific example" or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, unless they are contradictory.

Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments after knowing the basic creative concept. Therefore, the claims should be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present invention.

Obviously, those skilled in the art can make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention should also include these modifications and variations.

Claims (8)

1. A PISA fault identification method based on the semi-supervised Semi-KNN model, which is characterized by: S10. Obtain the blood glucose information to be measured within the preset time period, and preprocess the blood glucose information to be measured to obtain the preprocessed blood glucose information to be measured; S20. Based on the pre-established detection pressure sensor attenuation PISA constraint set and the preprocessed blood glucose information to be measured, use a similarity measurement processing method to obtain the constraint relationship to which the blood glucose information to be measured belongs; The PISA constraint set is a set with ML constraints and CL constraints constructed based on prior knowledge during the training phase of the semi-supervised Semi-KNN model. Each element in the set is information on the first-order difference characteristics of the blood glucose subsequence; The S20 includes: when the SBD distance between the blood glucose data A of each sequence in the blood glucose sequence to be measured and a PISA event B in the PISA constraint set is less than the threshold λ, that is, f _SBD (A, B) < λ, then determine a constraint. Relationship ML(A,B), and update the PISA constraint set; λ is a preset value greater than 0; f _SBD represents the function of calculating the SBD distance between two sequences; When the SBD distance between the blood glucose data A of each sequence in the blood glucose sequence to be measured and the CL constraint relationship of a PISA event B in the PISA constraint set is less than the threshold λ, that is, f _SBD (A, B) < λ, then a constraint relationship is determined CL(A,B), and update the PISA constraint set; Traverse each sequence in the blood glucose sequence to be measured, and use the updated PISA constraint set as the constraint relationship to which the blood glucose information to be measured belongs; S30. Input the preprocessed blood glucose information to be measured and the constraint relationship into the pre-trained semi-supervised Semi-KNN model, and the semi-supervised Semi-KNN model outputs the classification result of the blood glucose information to be measured; The semi-supervised Semi-KNN model is a semi-supervised model for identifying abnormalities in blood glucose information obtained by training the KNN model using a training data set and the PISA constraint set, and the training data set includes a first-order Differentially processed blood glucose data; The S30 includes: based on the K-D tree and the blood glucose sequence to be measured, obtain the nearest K data points to each data in the blood glucose sequence to be measured in a loop iterative manner, and obtain the K-D tree of the use stage, Based on the K-D tree of the usage stage, traverse the constraint relationship and obtain the classification results of the PISA abnormal information of the blood glucose sequence to be measured; The DTW similarity measure function is used to calculate the actual distance between each PISA event and other events in the constraint relationship, and the actual distance is compared with the abnormal threshold σ = [σ1, σ2] to obtain the classification results of PISA events and non-PISA events; σ is the abnormal threshold of the semi-supervised Semi-KNN model, and the boundary thresholds of the abnormal threshold σ are σ1 and σ2. 2. The method according to claim 1, characterized in that before said S10, the method further includes: S01. Obtain multiple historical blood glucose data with the help of the continuous blood glucose monitoring system CGM, preprocess each historical blood glucose data, and obtain a blood glucose sequence; each blood glucose sequence includes blood glucose data with PISA time stamp tags and non-PISA time Glucose data by stamping the label; S02. Divide each blood glucose sequence into multiple subsequences, and perform first-order difference calculation on each subsequence to obtain a training data set; S03. Based on prior knowledge and training data with PISA timestamp labels in the training data set, form rules according to semi-supervised constraints and generate a PISA constraint set; S04. Use the training data set and the PISA constraint set to train the semi-supervised Semi-KNN model to obtain the trained semi-supervised Semi-KNN model; The semi-supervised Semi-KNN model is an improved KNN model and is constructed in a semi-supervised manner. 3. The method according to claim 2, characterized in that said S04 includes: S04-1. Traverse all subsequences of the training data set, construct an offline K-dimensional search binary tree, and obtain a K-D tree; S04-2. Based on the K-D tree, traverse the PISA constraint set to obtain the abnormal threshold σ of the semi-supervised Semi-KNN model. The boundary thresholds of the abnormal threshold are σ1 and σ2, expressed as σ = [σ1, σ2]; Among them, the DTW similarity measure function is used to calculate the average distance between each PISA event and other events in the PISA constraint set to obtain the distance set; Then the abnormal threshold σ = [σ1, σ2] is obtained according to the following formula (1); σ1＝Q3+1.5(Q3-Q1), formula (1) σ2=Q1-1.5(Q3-Q1), If the distance between the blood glucose data to be measured and the sample distance of the ML relationship in the PISA constraint is dist<σ2, it is determined to be an abnormal sample; if the distance between the blood glucose data to be measured and the sample distance of the CL relationship in the PISA constraint is dist>σ1, it is determined to be an abnormal sample; Q3 is the upper quartile in the distance set, and Q1 is the lower quartile in the distance set. 4. The method according to claim 2, characterized in that said S02 includes: Sliding window processing is performed on each blood glucose sequence. In the blood glucose sequence X={x1,x2,…,xn}, several subsequences qi={ _xi , _xi+1 ,…, are formed after sliding windows of size w. x _i+k }, A sequence subset is D = {q1, q2,...,qm}. For each subsequence qi, the first-order difference calculation is performed according to formula (2). n represents the total length of the blood glucose sequence, and i is any one from 1 to n-w. The value represents the i-th sequence, and qi is the i-th sequence in the sequence subset; h is the change amount of the first-order difference formula, and the value of h is 0.8-1.2; After calculating the first-order difference for all subsequences, the first-order difference value of each subsequence is used as the training data set. 5. The method according to claim 1, characterized in that said S10 includes: Use CGM to obtain blood glucose information to be measured for 30-45 minutes or more; Filtering is performed, and the blood glucose information to be measured is preprocessed through a sliding window method to remove isolated noise points in the blood glucose information to be measured and to fill in missing values, thereby obtaining the blood glucose sequence to be measured of the blood glucose information to be measured. 6. The method according to claim 1, characterized in that, after comparing the actual distance with the abnormal threshold, the amount of data belonging to the ML constraint in the constraint relationship is determined, and the abnormal level value in the PISA event is determined based on the amount of data. 7. An electronic device, characterized in that it includes a memory and a processor, a computer program is stored in the memory, the processor executes the computer program stored in the memory, and executes any one of the above claims 1 to 6. The steps of the PISA fault identification method based on the semi-supervised Semi-KNN model are described. 8. A computer-readable storage medium, characterized in that a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the half-based method as described in any one of claims 1 to 6 is implemented. Steps of the PISA fault identification method supervised by the Semi-KNN model.