A sampling system for blood collection safety needle detection

Abstract

The present application relates to the technical field of medical instrument detection, in particular to a sampling system for blood collection safety needle detection, comprising: a data acquisition unit, a cleaning unit, a change identification unit, a component analysis unit and a strategy generation unit. The data acquisition unit synchronously collects monitoring data of needle body deformation, surface texture image and internal stress; the data cleaning unit eliminates abnormal jumps in deformation data; the change identification unit marks the surface texture change area through image frame comparison; the component analysis unit decomposes the internal stress into long-term baseline component and short-term fluctuation component. The strategy generation unit inputs the above multi-dimensional features into a dynamic sampling strategy generator to generate and real-time adjust the adaptive sampling strategy. This scheme can distinguish the cumulative fatigue and instantaneous impact state of the needle body, realize the dynamic optimization of the detection process, and thus improve the monitoring accuracy and efficiency of the performance degradation and failure risk of the blood collection needle.

Description

Technical Field

[0001] This invention relates to the field of medical device testing technology, and in particular to a sampling system for testing blood collection safety needles. Background Technology

[0002] With the continuous growth in clinical blood collection demand, blood collection safety needles, as high-risk disposable medical devices, are directly related to the safety of medical staff and patients due to their structural integrity and performance reliability. In recent years, the widespread application of new materials such as high-strength medical nickel-titanium alloys, antibacterial nanocomposite coatings, and carbon fiber reinforced medical plastics in the manufacture of blood collection safety needles has significantly improved needle toughness, puncture smoothness, and antibacterial performance. However, it has also brought new challenges to product quality testing: the failure modes of new materials are fundamentally different from those of traditional 304 stainless steel needles. For example, there is phase transformation-induced fatigue in nickel-titanium alloys, interfacial micro-peeling in nanocoatings, and interlaminar stress concentration in composite materials. Their early degradation characteristics are more hidden and their evolution is more complex, making it difficult for conventional testing methods to accurately capture them.

[0003] Trend decomposition of the internal stress reading sequence separates the long-term stress baseline component and the short-term stress fluctuation component. The long-term stress baseline component characterizes the slow time-varying process of creep or relaxation of the material under repeated loading, reflecting the cumulative fatigue damage of the structure; the short-term stress fluctuation component captures the transient impact loads caused by each puncture, operation, or external disturbance. This decomposition technique allows for a more refined assessment of needle failure modes, enabling the independent quantification of long-term material degradation trends and the frequency and intensity of short-term overload events. It provides key, independent yet interrelated criteria for predicting structural failure and identifying improper use.

[0004] The smooth deformation data stream, the surface change region record set, the long-term stress baseline component, and the short-term stress fluctuation component are all input into the dynamic sampling strategy generator. The generator performs real-time analysis and decision-making based on a multivariate state vector composed of these four types of features. When the system detects a steady increase in the long-term stress baseline while the surface change region intensifies, it may determine that the needle body has entered an accelerated wear period, thus automatically increasing the sampling frequency for intensive monitoring. If only isolated short-term stress fluctuations occur while other parameters remain stable, the system may determine it as a single, accidental impact, maintaining or reducing the sampling rate to conserve resources. This closed-loop decision-making mechanism based on multi-dimensional feature fusion achieves a fundamental shift in the sampling strategy from a fixed program to dynamic adaptation, ensuring that the system's focus is always automatically concentrated on the sensitive stages of state evolution, maximizing the targeting and timeliness of detection with limited resources.

[0005] Furthermore, in clinical practice, factors such as the standardized operation of blood collection needles by nurses, individual patient differences, and sudden external forces introduce a large number of uncertain instantaneous impact loads. Existing static, fixed-frequency sampling and detection modes are unable to capture these transient risks that are strongly related to clinical operations, and cannot effectively distinguish between structural cumulative fatigue of the needle body and instantaneous overload caused by nurse misoperation or accidents. This results in a lag in early warning of blood collection needle failure risks, making it difficult to ensure the ultimate safety of nurses and patients. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the existing technology and to propose a sampling system for blood collection safety needle detection.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a sampling system for blood collection safety needle detection, comprising: The data acquisition unit acquires the original monitoring dataset synchronously recorded by multiple sensors during the continuous monitoring period of the blood collection safety needle. The original monitoring dataset integrates the time-varying needle deformation measurement value sequence, surface texture image frame sequence, and internal stress reading sequence. The data cleaning unit cleans the needle deformation measurement value sequence, removes abnormal jump data points caused by instantaneous interference from the sensor, and generates a smooth deformation data stream. The change recognition unit compares the surface texture image frame sequence frame by frame, identifies and marks the regions where pixel-level changes occur between adjacent frames, and generates a record set of surface change regions. The component analysis unit performs trend decomposition on the internal stress reading sequence to separate the long-term stress baseline component and the short-term stress fluctuation component. The strategy generation unit inputs the smooth deformation data stream, the surface change region record set, the long-term stress baseline component, and the short-term stress fluctuation component into the dynamic sampling strategy generator.

[0008] Preferably, the dynamic sampling strategy generator performs the following operations: Based on the overall change rate of the smooth deformation data stream, a corresponding reference sampling frequency is matched in the preset slow, medium and fast intervals. Based on the area change rate and spatial distribution dispersion of the concentrated marked areas of the surface change region, the image feature dynamic index is calculated; By combining the stability assessment results of the long-term stress baseline component with the amplitude statistical characteristics of the short-term stress fluctuation component, a composite index of stress state is calculated. A three-dimensional sampling frequency control model is constructed, which uses the reference sampling frequency as the base, the image feature dynamic index as the horizontal adjustment factor, and the stress state composite index as the vertical adjustment factor. Using the aforementioned three-dimensional sampling frequency control model, a real-time sampling instruction sequence for the next stage of the current monitoring cycle is synthesized.

[0009] Preferably, the generation of the real-time sampling instruction sequence further includes: The reference sampling frequency, image feature dynamic index and stress state composite index are normalized and converted into scale-consistent control parameters. The normalized reference sampling frequency is mapped to the base layer of the three-dimensional sampling frequency control model; Above the base layer, a first frequency modulation layer controlled by the normalized image feature dynamic index is superimposed, the first frequency modulation layer introducing a frequency perturbation proportional to the image change dispersion. After the first frequency modulation layer is superimposed, a second frequency modulation layer controlled by the normalized stress state composite index is further superimposed. The second frequency modulation layer introduces frequency compensation that is inversely proportional to the suddenness of stress fluctuation. The multidimensional frequency vector generated after being modulated layer by layer by the base layer, the first frequency modulation layer and the second frequency modulation layer is fused and rounded to finally output the real-time sampling command sequence.

[0010] Preferably, the system further includes a sampling execution and feature extraction unit, used for: The system receives the real-time sampling instruction sequence and, based on the sampling frequency specified at each time point in the instruction sequence, triggers the image sensor and the mechanical sensor to synchronously collect data from the target blood collection safety needle. For each surface texture image acquired during a trigger, texture feature calculation based on the gray-level co-occurrence matrix is ​​performed to extract three texture parameters: contrast, correlation, and entropy. For each mechanical signal acquired during a trigger, a time-frequency joint analysis is performed, simultaneously extracting the root mean square value of the signal in the time domain and the dominant frequency component in the frequency domain. The three texture parameters, the root mean square value in the time domain, and the dominant frequency component in the frequency domain corresponding to each sampling event are packaged to form a single sampling feature data packet. Collect all single-sample feature data packets within a complete monitoring period in chronological order to construct a time-series sampling feature library.

[0011] Preferably, the system further includes a defect mode determination unit, used for: Load a predefined knowledge graph of blood collection safety needle defects, wherein the knowledge graph defines the mapping relationship between various defect types and multi-dimensional feature manifestations; Traverse the temporally sequenced sampling feature library and extract the feature combinations from each single sampling feature data packet in turn; The extracted feature combinations are matched with the feature representation templates in the defect knowledge graph through multiple rounds of progressive matching. In the first round of matching, coarse-grained matching is performed to filter out a set of candidate defect types that match the feature trends; In subsequent rounds of matching, fine-grained matching is performed, and the most matching target defect type is determined from the candidate defect type set by calculating the weighted Euclidean distance between feature values. Each single-sample feature data packet is assigned a defect label containing the target defect type and the corresponding matching confidence level.

[0012] Preferably, the specific steps for the defect mode determination unit to perform fine-grained matching include: For each candidate defect type in the candidate defect type set, its standard feature template is retrieved from the defect knowledge graph. The standard feature template contains the expected value range of the candidate defect type defect in each feature dimension. For the current single-sampled feature data packet to be judged, calculate the deviation between each feature value in its feature combination and the corresponding value in the expected value range; Based on predefined expert experience rules, different weight coefficients are assigned to different feature dimensions. Among them, feature dimensions that reflect the abrupt changes in surface texture are given higher weights, while feature dimensions that reflect the stability of mechanical signals are given medium weights. Based on the deviation of each feature dimension and its corresponding weight coefficient, calculate the comprehensive difference between the current feature combination and the standard feature template of each candidate defect type; The candidate defect type with the smallest overall difference is selected as the target defect type, and the overall difference is converted into the matching confidence score.

[0013] Preferably, the system further includes a sampling strategy feedback optimization unit, used for: Collect all single-sample feature data packets and their corresponding defect labels generated at the end of a complete monitoring cycle; Analyze the distribution changes of the defect labels in different time periods to identify warning periods when the frequency of defect occurrence increases; The original control parameters corresponding to the warning period are traced back, namely the historical values ​​of the normalized baseline sampling frequency, image feature dynamic index, and stress state composite index. The historical control parameters of the warning period are input into a strategy correction network. After training, the strategy correction network can output the adjustment amount of the modulation rule in the three-dimensional sampling frequency control model. Based on the adjustment amount, the calculation rules for frequency disturbance and frequency compensation in the first frequency modulation layer and the second frequency modulation layer are dynamically updated, so that the system can automatically increase the sampling density when facing similar operating conditions in subsequent monitoring.

[0014] Preferably, the workflow of the policy correction network includes: A training sample set is constructed, and each sample consists of a set of historical control parameters and a set of expected sampling frequency adjustment amounts. The expected sampling frequency adjustment amounts are manually labeled by domain experts based on the situation during historical warning periods. The policy correction network is constructed using a multilayer perceptron structure, with historical control parameters as the input layer and the desired sampling frequency adjustment as the output layer. The policy correction network is trained in a supervised manner using the training sample set until the error between the predicted adjustment amount output by the network and the expected adjustment amount labeled by the expert is lower than a preset threshold. During the deployment phase, the historical control parameters of the early warning period obtained in real time are input into the trained policy correction network to directly obtain the corresponding modulation rule adjustment amount.

[0015] Preferably, the system further includes a data visualization and report synthesis unit, used for: Receive a sequence of defect tags with time stamps from the defect mode determination unit; The temporally sequenced sampling feature library is received from the sampling execution and feature extraction unit. The defect label sequence is aligned and fused with the time-series sampling feature library along the time axis to generate complete timeline data containing original features, defect judgment results and confidence levels. Based on the complete timeline data, a two-dimensional chart is automatically generated, where the horizontal axis represents time and the vertical axis simultaneously displays the key feature value curves and the distribution of defect types. Based on the preset report template, the two-dimensional charts, key statistical data, and conclusive text descriptions are integrated to output the final electronic test report document.

[0016] Preferably, performing trend decomposition on the internal stress reading sequence to separate the long-term stress baseline component and the short-term stress fluctuation component specifically includes: Establish a coordinate system with sampling time as the horizontal axis and stress reading as the vertical axis, and plot the internal stress reading sequence as the original stress curve; The original stress curve is processed using the sliding window averaging method. The window length is set to a preset proportion of the total monitoring period, and the arithmetic mean of the stress readings in each window is calculated. Connect the arithmetic mean points of all sliding windows to form a smooth trend line, which is used as the long-term stress baseline component. The residual sequence is obtained by subtracting the long-term stress baseline component values ​​at corresponding times from the original stress curve point by point. The standard deviation of the residual sequence is calculated, and residual points with absolute values ​​exceeding three times the standard deviation are identified as abnormal fluctuation points and removed. The residual sequence after removing abnormal fluctuation points is used as the short-term stress fluctuation component, and the amplitude sign of each fluctuation point is recorded.

[0017] Compared with the prior art, the advantages and positive effects of the present invention are as follows: Trend decomposition was performed on the internal stress reading sequence to separate the long-term stress baseline component and the short-term stress fluctuation component. The long-term stress baseline component characterizes the slow time-varying process of creep or relaxation of the material under repeated loading, reflecting the cumulative fatigue damage of the structure; the short-term stress fluctuation component captures the transient impact loads caused by each puncture, operation, or external disturbance. This decomposition technique allows for a more refined assessment of needle failure modes, enabling independent quantification of the long-term material degradation trend and the frequency and intensity of short-term overload events, providing a scientific criterion for failure prediction of all types of new material blood collection safety needles.

[0018] The smooth deformation data stream, the surface change region record set, the long-term stress baseline component, and the short-term stress fluctuation component are all input into the dynamic sampling strategy generator. The generator performs real-time analysis and decision-making based on a multivariate state vector composed of these four types of features. When the system detects a steady increase in the long-term stress baseline while the surface change region intensifies, it may determine that the needle body has entered an accelerated wear period, thus automatically increasing the sampling frequency for intensive monitoring. If only isolated short-term stress fluctuations occur while other parameters remain stable, the system may determine it as a single, accidental impact, maintaining or reducing the sampling rate to conserve resources. This closed-loop decision-making mechanism based on multi-dimensional feature fusion achieves a fundamental shift in the sampling strategy from a fixed program to dynamic adaptation, ensuring that the system's focus is always automatically concentrated on the sensitive stages of state evolution, maximizing the targeting and timeliness of detection with limited resources.

[0019] The system's built-in defect pattern determination unit supports offline updates to the defect knowledge graph, rapidly incorporating defect types and characteristic templates unique to new materials such as microcracks in nickel-titanium alloys, peeling off nano-coating interfaces, and carbon fiber fracture. The sampling strategy feedback optimization unit, through a strategy correction network, can autonomously learn the characteristic evolution patterns of defects in different materials and dynamically adjust the modulation rules of the three-dimensional sampling frequency control model. This mechanism enables the system to continuously evolve, seamlessly adapting to the application of future novel medical materials in blood collection safety needles without requiring large-scale modifications to the core architecture. Attached Figure Description

[0020] Figure 1This is a timing diagram of the sampling system for detecting blood collection safety needles according to the present invention; Figure 2 A flowchart for generating real-time sampling command sequences; Figure 3 The flowchart for sampling execution and feature extraction; Figure 4 A graph showing the sliding window statistics of defect frequency and identification of early warning periods; Figure 5 This is a dynamic correlation diagram between the number of defects and the adjustment of the sampling frequency. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0022] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0023] See Figure 1 The system acquires raw monitoring datasets from multiple sensors synchronously recorded by the blood collection safety needle during a continuous monitoring period through a data acquisition unit. This dataset integrates time-varying needle deformation measurement sequences, surface texture image frame sequences, and internal stress reading sequences. A data cleaning unit cleans the needle deformation measurement sequence, removing abnormal jump data points caused by instantaneous sensor interference, thus generating a smooth deformation data stream. A change identification unit compares the surface texture image frame sequence frame by frame, identifying and marking regions where pixel-level changes occur between adjacent frames, thereby generating a surface change region record set. A component analysis unit performs trend decomposition on the internal stress reading sequence, separating the long-term stress baseline component and the short-term stress fluctuation component. A strategy generation unit inputs the smooth deformation data stream, the surface change region record set, the long-term stress baseline component, and the short-term stress fluctuation component into a dynamic sampling strategy generator.

[0024] In one embodiment of the present invention, see [reference] Figure 2The dynamic sampling strategy generator matches corresponding reference sampling frequencies in preset slow, medium, and fast intervals based on the overall change rate of the smooth deformation data stream. For example, when the overall change rate of the smooth deformation data stream is less than 0.5 mm / min, the reference sampling frequency for the slow interval is matched once per second; when the overall change rate is between 0.5 mm / min and 2 mm / min, the reference sampling frequency for the medium interval is matched five times per second; and when the overall change rate is greater than 2 mm / min, the reference sampling frequency for the fast interval is matched ten times per second. Based on the area change rate and spatial distribution dispersion of the surface change region, an image feature dynamic index is calculated. The calculation of the image feature dynamic index integrates the normalized values ​​of the area change rate and the spatial distribution dispersion and uses a weighted summation method. The normalized value of the area change rate is obtained by dividing the actual area change rate by the maximum allowable area change rate, and the normalized value of the spatial distribution dispersion is obtained by dividing the actual spatial distribution variance by the maximum allowable spatial distribution variance. By combining the stability assessment results of the long-term stress baseline component with the amplitude statistical characteristics of the short-term stress fluctuation component, a stress state composite index is calculated. The stability assessment results of the long-term stress baseline component are obtained by calculating the reciprocal of the standard deviation of the long-term stress baseline component, and the amplitude statistical characteristics of the short-term stress fluctuation component are obtained by calculating the average amplitude of the short-term stress fluctuation component. The stress state composite index is obtained by normalizing the product of the stability assessment results and the amplitude statistical characteristics.

[0025] In some embodiments, a three-dimensional sampling frequency control model is constructed, using a reference sampling frequency as the base, an image feature dynamic index as the horizontal adjustment factor, and a stress state composite index as the vertical adjustment factor. This three-dimensional sampling frequency control model is then used to synthesize a real-time sampling command sequence for the next stage of the current monitoring cycle. The generation of the real-time sampling command sequence involves normalizing the reference sampling frequency, image feature dynamic index, and stress state composite index, converting them into scale-consistent control parameters. The normalization process uses a linear transformation method to linearly map the reference sampling frequency from its original frequency range to the [0,1] interval, the image feature dynamic index from its original index range to the [0,1] interval, and the stress state composite index from its original index range to the [0,1] interval. The normalized reference sampling frequency is mapped to the base layer of the three-dimensional sampling frequency control model. A first frequency modulation layer, controlled by a normalized image feature dynamic index, is superimposed on this base layer. This first frequency modulation layer introduces a frequency perturbation proportional to the image variation dispersion. The amount of this perturbation is determined by multiplying the normalized image feature dynamic index by a proportional coefficient. The image variation dispersion is characterized by the spatial distribution variance of the marked areas in the surface variation region record. After superimposing the first frequency modulation layer, a second frequency modulation layer, controlled by a normalized stress state composite index, is further superimposed. This second frequency modulation layer introduces frequency compensation inversely proportional to the stress fluctuation suddenness. The amount of this compensation is determined by multiplying the reciprocal of the normalized stress state composite index by an inverse proportional coefficient. The stress fluctuation suddenness is characterized by the amplitude change rate of the short-term stress fluctuation component.

[0026] It can be understood that the three-dimensional sampling frequency control model calculates the real-time sampling frequency using the following formula:

[0027] in: Indicates the real-time sampling frequency. This represents the normalized reference sampling frequency. Represents the dynamic index of normalized image features. This represents the normalized composite index of stress state. This represents the perturbation coefficient, which is proportional to the dispersion of image changes. This represents a compensation coefficient that is inversely proportional to the suddenness of stress fluctuations. This represents the rounding operator.

[0028] In practice, The value is calibrated based on historical data of image variation dispersion. The value is calibrated and set based on historical data of sudden stress fluctuations. The multi-dimensional frequency vector generated after layer-by-layer modulation through the base layer, the first frequency modulation layer, and the second frequency modulation layer is fused and rounded to finally output a real-time sampling command sequence. The fusion operation combines the multi-dimensional frequency vectors into a single frequency value using the above formula, and the rounding operation ensures that the sampling frequency is a positive integer. Optionally, the normalized reference sampling frequency, the normalized image feature dynamic index, and the normalized stress state composite index are processed by a constraint function before being input into the three-dimensional sampling frequency control model. The constraint function constrains the parameter values ​​within the [0,1] interval to prevent overflow. Optionally, the preset range of the reference sampling frequency is dynamically adjusted according to the material properties of the blood collection safety needle, and the thresholds for the slow, medium, and fast ranges are periodically updated based on historical monitoring data.

[0029] In one embodiment of the present invention, see [reference] Figure 3 The sampling execution and feature extraction unit receives a real-time sampling instruction sequence from the dynamic sampling strategy generator. This sequence contains a series of time points and a specified sampling frequency for each time point. For example, within a 10-second monitoring sub-cycle, time points 0, 3, and 7 correspond to sampling frequencies of 2 times per second, 5 times per second, and 1 time per second, respectively. Based on the sampling frequency specified for each time point in the instruction sequence, the sampling execution and feature extraction unit synchronously triggers the image sensor and mechanical sensor to acquire data from the target blood collection needle. The image sensor captures surface texture images, and the mechanical sensor captures stress or strain signals. Synchronous triggering ensures strict time alignment between the image and mechanical data at the same sampling moment. For each surface texture image acquired during triggering, texture feature calculation based on the gray-level co-occurrence matrix is ​​performed. The gray-level co-occurrence matrix is ​​constructed based on the co-occurrence probability of image pixel gray values ​​at a specified spatial distance and direction. In this implementation, the spatial distance is set to 1 pixel, and the direction considers four directions: 0 degrees, 45 degrees, 90 degrees, and 135 degrees, and the average value is taken. Three texture parameters—contrast, correlation, and entropy—are extracted from the calculated gray-level co-occurrence matrix. Contrast reflects the sharpness of the image and the depth of its texture ridges and grooves; correlation measures the similarity of image textures along rows or columns; and entropy represents the richness of information contained in the image. The formula for calculating the entropy parameter is as follows:

[0030] in: Represents the entropy value. This indicates that the gray-level co-occurrence matrix located at the th position in the normalized gray-level co-occurrence matrix is... Line number The element values ​​of the column, i.e., the gray levels. and The probability of co-occurrence under a specified spatial relationship is calculated by summing all rows and columns of the gray-level co-occurrence matrix.

[0031] In some embodiments, time-frequency joint analysis is performed on the mechanical signal acquired each time it is triggered. This joint analysis processes both the time-domain and frequency-domain characteristics of the signal. In the time domain, the root mean square (RMS) value of the signal is extracted. The RMS value is obtained by squaring, averaging, and then taking the square root of the instantaneous values ​​of the mechanical signal within a sampling window, representing the average energy level of the signal. In the frequency domain, the dominant frequency component of the signal is extracted. The dominant frequency component is obtained by applying a Fast Fourier Transform (FFT) to the mechanical signal within the same sampling window to obtain the amplitude spectrum, and then identifying the frequency component with the highest energy in the amplitude spectrum, representing the main oscillation mode of the signal. The FFT uses a Hanning window, with the window length maintaining a fixed multiple relationship with the current sampling frequency, specifically covering 100 consecutive data points at the current sampling frequency, to achieve a balance between frequency resolution and spectral leakage. The three texture parameters, the time-domain RMS value, and the dominant frequency component corresponding to each sampling event are packaged to form a single-sampling feature data packet. The single-sampling feature data packet is a structured data record containing five specific feature values ​​and a precise timestamp. All single-sample feature data packets within a complete monitoring period are collected in chronological order to construct a time-series sampling feature library. The time-series sampling feature library organizes data with time as the index and fully records the continuous evolution of feature parameters within the monitoring period.

[0032] Optionally, the image sensor uses fixed exposure parameters and light source conditions when triggering the acquisition of surface texture images to ensure the comparability of images acquired at different times. The mechanical sensor performs zero-point calibration before triggering acquisition to eliminate baseline drift. The specific fixed exposure parameters are: exposure time 1 / 100 second, aperture F8, and ISO 100. The light source conditions are: a red LED surface light source with a wavelength of 650 nm, uniformly illuminating the needle surface at a 45-degree angle, with an illumination intensity of 1000 lux. The camera lens used has a magnification of 2x, enabling the acquisition of high-resolution images of local areas on the needle surface on the sensor target. Optionally, the gray-level co-occurrence matrix in texture feature calculation is quantized to 16 levels to balance computational accuracy and processing speed. The sampling window length used in the time-frequency joint analysis of the mechanical signal maintains a fixed ratio with the current sampling frequency; for example, the window length is fixed to the time span of 100 data points at the current sampling frequency. In some embodiments, the five feature values ​​of a single sampling feature data packet are normalized before packaging. The normalization process is based on the historical statistical maximum and minimum values ​​of each feature parameter and is linearly scaled so that all feature values ​​fall within the value range of 0 to 1. The time-series sampling feature library stores the normalized feature values.

[0033] In one embodiment of the present invention, the defect mode determination unit loads a predefined blood sampling safety needle defect knowledge graph. The blood sampling safety needle defect knowledge graph is stored in a graph structure, where nodes represent defect types or feature dimensions, and edges represent the association between defect types and feature manifestations. The blood sampling safety needle defect knowledge graph defines various defect types, including needle tip microcracks, needle body plastic bending, surface coating peeling, and internal material fatigue. The unit traverses the temporally sequenced sampling feature library from the sampling execution and feature extraction unit, sequentially extracting feature combinations from each single-sample feature data package. Each feature combination includes three texture parameters (contrast, correlation, and entropy) and five specific values: the root mean square value in the time domain, the dominant frequency component in the frequency domain, and so on. The extracted feature combinations are then progressively matched with feature manifestation templates in the blood sampling safety needle defect knowledge graph. These feature manifestation templates define, in vector form, the typical numerical manifestations or trends of each defect type across various feature dimensions. In the first round of matching, coarse-grained matching is performed. This compares the trend direction of each feature value in the current feature combination with the trend direction of the corresponding feature dimension in the feature representation template, filtering out a set of candidate defect types with matching trend directions. Trend directions are categorized as rising, falling, or stable. In subsequent rounds of matching, fine-grained matching is performed. By calculating the weighted Euclidean distance between feature values, the best-matching target defect type is determined from the set of candidate defect types. The weighted Euclidean distance calculation assigns different weights to different feature dimensions. Each single-sample feature data packet is assigned a defect label containing the target defect type and corresponding matching confidence level. This defect label is appended as metadata to the original single-sample feature data packet.

[0034] The specific steps of fine-grained matching include retrieving the standard feature template from the defect knowledge graph for each candidate defect type in the candidate defect type set. The standard feature template contains the expected value range for each feature dimension of the candidate defect type. For the current single-sampled feature data packet to be judged, the deviation of each feature value in its feature combination from the corresponding midpoint of the expected value range is calculated. The deviation is defined as the absolute value of the current feature value minus the midpoint of the expected value range, divided by the width of the expected value range. Different weight coefficients are assigned to different feature dimensions according to predefined expert experience rules. These expert experience rules are stored in a configuration table, where feature dimensions reflecting surface texture abrupt changes are given higher weights. Based on the deviation of each feature dimension and its corresponding weight coefficient, the comprehensive difference between the current feature combination and the standard feature template of each candidate defect type is calculated. The formula for calculating the comprehensive difference is:

[0035] in: Indicates the current feature combination and the first The overall degree of difference among candidate defect types Indicates the first Pre-assigned weight coefficients for each feature dimension Indicates the th element in the current feature combination The actual values ​​of each feature dimension Indicates the first Candidate defect types in the first The median of the expected value range across each feature dimension. Indicates the first Type of defect in the first The expected value range width across each feature dimension is calculated by summing all five feature dimensions.

[0036] In some embodiments, the candidate defect type with the smallest overall difference is selected as the target defect type, and the overall difference is converted into a matching confidence score. The conversion method involves inputting the overall difference into a preset linear transformation function, which maps the overall difference to a confidence score range of 0% to 100%. The smaller the overall difference, the higher the matching confidence score. It can be understood that coarse-grained matching and fine-grained matching constitute a two-stage screening and precise judgment process. Coarse-grained matching quickly narrows the judgment range to improve efficiency, while fine-grained matching performs precise calculations within the narrowed range to improve accuracy. Optionally, in fine-grained matching, if the calculated minimum overall difference is still higher than a preset absolute threshold, the state corresponding to the single sampled feature data packet is determined to be "unknown" or "normal," and a lower default matching confidence score is assigned. Optionally, the standard feature templates in the defect knowledge graph support offline updates. By injecting new historical detection data and expert annotation results, the expected value range of existing defect types can be adjusted, or new defect type nodes and their feature representation templates can be added.

[0037] In one embodiment of the present invention, the sampling strategy feedback optimization unit collects all single-sample feature data packets and their corresponding defect tags generated after a complete monitoring cycle, such as a continuous eight-hour production batch monitoring cycle. The distribution changes of defect tags at different time periods are analyzed to identify warning periods where the frequency of defect occurrence increases. The analysis process divides the entire monitoring cycle into continuous fixed-length time windows, counts the number of various defect tags appearing in each time window, and calculates their ratio to the average frequency of the entire cycle. Time windows where the ratio exceeds a preset threshold are marked as warning periods. The original control parameters corresponding to the warning periods are backtracked, namely the historical values ​​of the normalized baseline sampling frequency, the normalized image feature dynamic index, and the normalized stress state composite index. The historical values ​​are extracted from the log records of the dynamic sampling strategy generator by timestamp. The historical control parameters of the warning periods are input into a strategy correction network. After training, the strategy correction network can output the adjustment amount of the modulation rules in the three-dimensional sampling frequency control model. Based on the adjustment amount, the calculation rules of frequency disturbance and frequency compensation in the first and second frequency modulation layers are dynamically updated, enabling the system to automatically increase the sampling density when facing similar working conditions in subsequent monitoring. The workflow of the policy correction network includes constructing a training sample set. Each sample consists of a set of historical control parameters and a set of expected sampling frequency adjustments. The expected sampling frequency adjustments are manually labeled by domain experts based on the historical warning periods, and the corresponding adjustments need to be calculated in reverse. A multilayer perceptron structure is used to construct the policy correction network, with historical control parameters as the input layer and expected sampling frequency adjustments as the output layer. The input layer has 3 nodes, corresponding to the three normalized control parameters, and the output layer has 2 nodes, corresponding to the adjustments of the two coefficients. The policy correction network is supervisedly trained using the training sample set until the error between the network's predicted adjustment and the expert-labeled expected adjustment is below a preset threshold. During the deployment phase, the historical control parameters of the warning periods obtained through real-time backtracking are input into the trained policy correction network to directly obtain the corresponding modulation rule adjustments. The data visualization and report synthesis unit receives a sequence of defect labels with time stamps from the defect pattern determination unit and a time-series sampling feature library from the sampling execution and feature extraction unit. The defect label sequence is aligned and fused with the time-series sampling feature library along the time axis to generate complete timeline data containing original features, defect judgment results, and confidence levels. The fusion is based on timestamps accurate to milliseconds. Based on the complete timeline data, a two-dimensional chart is automatically generated, where the horizontal axis represents time, and the vertical axis simultaneously displays key feature value curves and defect type distribution. Key feature values ​​include, for example, entropy and root mean square value in the time domain. Following a preset report template, the two-dimensional chart, key statistical data, and conclusive text descriptions are integrated to output the final electronic inspection report document.

[0038] In some embodiments, a sliding time window statistical method is used to analyze the changes in defect label distribution. The window length is set to 15 minutes, and the sliding step size is set to 5 minutes. The ratio of the number of occurrences of a specific defect type within each window to the average number of occurrences of that defect per hour throughout the entire monitoring period is calculated. When the ratio exceeds 1.5 for two consecutive windows, the time period is determined to be an early warning period for that defect type. Optionally, the historical control parameters are taken as the average value of the control parameters of all sampling points within the corresponding early warning period. The hidden layer of the policy correction network adopts a two-layer structure, with 8 and 4 neurons in each layer, respectively. The ReLU function is used as the activation function, and the linear function is used as the activation function of the output layer. The training objective of the policy correction network is to minimize the mean square error between the predicted adjustment and the expert-annotated adjustment. The policy correction network adopts a multilayer perceptron structure. Its input layer contains 3 nodes, corresponding to the normalized baseline sampling frequency, the image feature dynamic index, and the stress state composite index, respectively. The hidden layer has two layers: the first hidden layer contains 8 nodes, and the second hidden layer contains 4 nodes, both using the linear rectified function as the activation function. The output layer contains two nodes and uses a linear function as the activation function. Its loss function is... Represented as:

[0039] in: This represents the value of the loss function. Indicates the number of training samples. and These represent the policy correction network's response to the first... The adjustment amount of the perturbation coefficient and the adjustment amount of the compensation coefficient for each sample prediction. and They represent the first The expected disturbance coefficient adjustment and expected compensation coefficient adjustment amount annotated by experts in each sample.

[0040] It is understandable that the trained policy correction network can learn the complex mapping relationship between historical control parameter patterns, defect frequency, and required sampling policy adjustments. Optionally, to address different defect type patterns, multiple policy correction network sub-models focusing on warnings for different defect types can be trained. During application, the corresponding sub-model outputs the adjustment amount based on the main defect type during the warning period. In some embodiments, the two-dimensional charts generated by the data visualization and report synthesis unit adopt a dual-axis design. The left axis displays the normalized feature value scale, and the right axis displays the defect type classification code. The defect type distribution is marked on the time axis with discrete points of different colors. The report template is a structured document containing sections such as a test overview, data timeline charts, defect statistics summary tables, and conclusions. Key statistical data includes information such as the total number of defects, the proportion of each defect type, and the period with the highest defect frequency. Refer to Table 1 for an example defect statistics summary table.

[0041] Table 1: Summary Table of Defect Statistics During Monitoring Cycle

[0042] It is understandable that the sampling strategy feedback optimization unit realizes closed-loop control by dynamically optimizing the future sampling strategy based on the defect occurrence history through the strategy correction network, while the data visualization and report synthesis unit provides the interface for human-computer interaction and result delivery.

[0043] See Figure 4 This is a chart showing the defect frequency sliding window statistics and early warning period identification. This chart presents the defect frequency statistics for blood collection safety needles within an 8-hour production monitoring cycle, using a 15-minute sliding window (5-minute step) to calculate the number of defects per hour. This chart directly supports the sampling strategy feedback optimization process: when the defect frequency exceeds the threshold for two consecutive windows, the system automatically marks the early warning period and backtracks the control parameters for that period, inputting the strategy correction network to generate sampling frequency adjustment instructions. Accurate identification of high defect frequency periods can guide the automatic increase of sampling density under similar operating conditions, thereby detecting defects earlier and reducing the risk of defective products leaving the system. Combined with defect type distribution data, specific investigations can be conducted on production equipment and raw material batches during the two high-risk periods to pinpoint the root cause. Furthermore, early warning thresholds can be linked to defect types, setting differentiated thresholds for different defects such as needle tip micro-cracks and coating peeling, improving early warning accuracy.

[0044] In one embodiment of the invention, the process of performing trend decomposition on the internal stress reading sequence to separate the long-term stress baseline component and the short-term stress fluctuation component establishes a coordinate system with sampling time as the horizontal axis and stress reading as the vertical axis. The internal stress reading sequence is plotted as the original stress curve. The internal stress reading sequence is collected from a mechanical sensor at a fixed frequency, for example, at a frequency of 10 times per second within a monitoring period of 600 seconds, resulting in a sequence containing 6000 stress readings. The original stress curve is processed using the sliding window averaging method. The window length is set as a preset proportion of the total monitoring period, for example, 60 seconds, which is one-tenth of the total monitoring period. The arithmetic mean of the stress readings within each window is calculated. For the first window (time 0 seconds to 60 seconds), the arithmetic mean is the sum of the 600 stress readings within that window divided by 600. The sliding window moves forward with a fixed step size, set to one sampling interval, i.e., 0.1 seconds. The sliding window averaging method uses a window length set to one-tenth of the total monitoring period and a sliding step size set to one sampling interval to ensure that the long-term baseline component smoothly covers the original data curve. When calculating texture features, the gray-level co-occurrence matrix is ​​constructed using 16 gray levels, with a spatial distance of 1 pixel. The orientation is considered at 0 degrees, 45 degrees, 90 degrees, and 135 degrees, and the average value is taken. Connecting the arithmetic mean points of all sliding windows forms a smooth trend line, which is used as the long-term stress baseline component. The long-term stress baseline component is a smoothed sequence that corresponds point-to-point with the original stress curve, with the same number of data points as the arithmetic mean points calculated by the sliding windows, and is aligned with the original timestamps through linear interpolation. The long-term stress baseline component values ​​at corresponding times are subtracted point-by-point from the original stress curve to obtain a residual sequence. Each point in the residual sequence represents the difference between the original stress value and the long-term stress baseline component value at that time.

[0045] The standard deviation of the residual sequence is calculated, and residual points with absolute values ​​exceeding three times the standard deviation are identified as abnormal fluctuation points and removed. The standard deviation calculation is based on all points in the entire residual sequence, and three times the standard deviation constitutes a threshold boundary. The residual sequence after removing abnormal fluctuation points is taken as the short-term stress fluctuation component, and the amplitude sign of each fluctuation point is recorded. A positive amplitude sign indicates that the fluctuation direction is the same as the original stress change direction, and a negative sign indicates the opposite. In some embodiments, the window length of the sliding window averaging method is adjusted according to the total duration of the monitoring period and the expected slowness of the stress change. For monitoring periods of several hours, the window length can be set to 5% of the total duration; for short-term monitoring of several minutes, the window length can be set to 20% of the total duration. The formula used to calculate the arithmetic mean of stress readings within each window is:

[0046] in: Indicates the first The arithmetic mean calculated by the sliding window This indicates the total number of stress readings contained within the sliding window. Indicates the first The start time of each sliding window Indicates the duration of the window. Indicates time The collected raw stress readings are summed and the calculation is performed across the window from... arrive All sampling times.

[0047] It is understandable that when connecting all the arithmetic mean points of the sliding window to form the long-term stress baseline component, for time points on the original stress curve located in the edge region of the sliding window, the corresponding long-term stress baseline component value is obtained by linear interpolation of the arithmetic mean of two adjacent windows. Optionally, after identifying and removing abnormal fluctuation points, the short-term stress fluctuation component sequence can be further smoothed and filtered, for example, by using a three-point moving average method to suppress residual high-frequency noise. In some embodiments, while recording the amplitude sign of each fluctuation point of the short-term stress fluctuation component, its absolute timestamp and normalized amplitude relative to the long-term stress baseline component are also recorded. The normalized amplitude is obtained by dividing the actual amplitude of the fluctuation point by the value of the long-term stress baseline component at that time point. Optionally, the sliding window averaging method can be replaced by the exponentially weighted moving average method or the low-pass digital filtering method to extract the long-term stress baseline component, the choice depending on the requirements of real-time calculation and smoothness for the specific application scenario. It is understandable that the separated long-term stress baseline component reflects the steady-state load or slow drift that the blood collection safety needle experiences during the monitoring period, while the short-term stress fluctuation component reflects the dynamic response caused by transient impacts, vibrations, or microstructural changes. Together, they provide mechanical state input for the generation of dynamic sampling strategies.

[0048] See Figure 5This is a dynamic correlation graph showing the relationship between the number of defects and the adjustment of the sampling frequency. The graph illustrates the correspondence between the changes in the number of different defect types and the system's automatic adjustment of the sampling frequency within a 10-minute monitoring period. This dynamic change perfectly reflects the sampling strategy feedback optimization mechanism. When the number of defects increases during a certain period, the system calculates and increases the sampling frequency through a strategy correction network to capture more defect features and improve detection sensitivity. When the number of defects decreases or disappears, the system automatically reduces the sampling frequency to balance detection accuracy and efficiency. For the two high-incidence periods of defect type 1 (1-2 minutes and 8-9 minutes), the corresponding control parameters (such as stress state and image feature index) can be further analyzed to optimize the warning threshold. For the high-incidence period of defect type 3 (2-4 minutes), the sampling density of surface texture images can be increased to focus on monitoring coating peeling-related features.

[0049] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A sampling system for detecting blood collection safety needles, characterized in that, The system includes: The data acquisition unit acquires the original monitoring dataset synchronously recorded by multiple sensors during the continuous monitoring period of the blood collection safety needle. The original monitoring dataset integrates the time-varying needle deformation measurement value sequence, surface texture image frame sequence, and internal stress reading sequence. The data cleaning unit cleans the needle deformation measurement value sequence, removes abnormal jump data points caused by instantaneous interference from the sensor, and generates a smooth deformation data stream. The change recognition unit compares the surface texture image frame sequence frame by frame, identifies and marks the regions where pixel-level changes occur between adjacent frames, and generates a record set of surface change regions. The component analysis unit performs trend decomposition on the internal stress reading sequence to separate the long-term stress baseline component and the short-term stress fluctuation component. The strategy generation unit inputs the smooth deformation data stream, the surface change region record set, the long-term stress baseline component, and the short-term stress fluctuation component into the dynamic sampling strategy generator.

2. The sampling system for blood collection safety needle detection as described in claim 1, characterized in that, The dynamic sampling strategy generator performs the following operations: Based on the overall change rate of the smooth deformation data stream, a corresponding reference sampling frequency is matched in the preset slow, medium and fast intervals. Based on the area change rate and spatial distribution dispersion of the concentrated marked areas of the surface change region, the image feature dynamic index is calculated; By combining the stability assessment results of the long-term stress baseline component with the amplitude statistical characteristics of the short-term stress fluctuation component, a composite index of stress state is calculated. A three-dimensional sampling frequency control model is constructed, which uses the reference sampling frequency as the base, the image feature dynamic index as the horizontal adjustment factor, and the stress state composite index as the vertical adjustment factor. Using the aforementioned three-dimensional sampling frequency control model, a real-time sampling instruction sequence for the next stage of the current monitoring cycle is synthesized.

3. The sampling system for blood collection safety needle detection as described in claim 2, characterized in that, The generation of the real-time sampling instruction sequence further includes: The reference sampling frequency, image feature dynamic index and stress state composite index are normalized and converted into scale-consistent control parameters. The normalized reference sampling frequency is mapped to the base layer of the three-dimensional sampling frequency control model; Above the base layer, a first frequency modulation layer controlled by the normalized image feature dynamic index is superimposed, the first frequency modulation layer introducing a frequency perturbation proportional to the image change dispersion. After the first frequency modulation layer is superimposed, a second frequency modulation layer controlled by the normalized stress state composite index is further superimposed. The second frequency modulation layer introduces frequency compensation that is inversely proportional to the suddenness of stress fluctuation. The multidimensional frequency vector generated after being modulated layer by layer by the base layer, the first frequency modulation layer and the second frequency modulation layer is fused and rounded to finally output the real-time sampling command sequence.

4. The sampling system for blood collection safety needle detection as described in claim 3, characterized in that, It also includes a sampling execution and feature extraction unit, used for: The system receives the real-time sampling instruction sequence and, based on the sampling frequency specified at each time point in the instruction sequence, triggers the image sensor and the mechanical sensor to synchronously collect data from the target blood collection safety needle. For each surface texture image acquired during a trigger, texture feature calculation based on the gray-level co-occurrence matrix is ​​performed to extract three texture parameters: contrast, correlation, and entropy. For each mechanical signal acquired during a trigger, a time-frequency joint analysis is performed, simultaneously extracting the root mean square value of the signal in the time domain and the dominant frequency component in the frequency domain. The three texture parameters, the root mean square value in the time domain, and the dominant frequency component in the frequency domain corresponding to each sampling event are packaged to form a single sampling feature data packet. Collect all single-sample feature data packets within a complete monitoring period in chronological order to construct a time-series sampling feature library.

5. The sampling system for blood collection safety needle detection as described in claim 4, characterized in that, It also includes a defect mode determination unit, used for: Load a predefined knowledge graph of blood collection safety needle defects, wherein the knowledge graph defines the mapping relationship between various defect types and multi-dimensional feature manifestations; Traverse the temporally sequenced sampling feature library and extract the feature combinations from each single sampling feature data packet in turn; The extracted feature combinations are matched with the feature representation templates in the defect knowledge graph through multiple rounds of progressive matching. In the first round of matching, coarse-grained matching is performed to filter out a set of candidate defect types that match the feature trends; In subsequent rounds of matching, fine-grained matching is performed, and the most matching target defect type is determined from the candidate defect type set by calculating the weighted Euclidean distance between feature values. Each single-sample feature data packet is assigned a defect label containing the target defect type and the corresponding matching confidence level.

6. The sampling system for blood collection safety needle detection as described in claim 5, characterized in that, The specific steps for the defect mode determination unit to perform fine-grained matching include: For each candidate defect type in the candidate defect type set, its standard feature template is retrieved from the defect knowledge graph. The standard feature template contains the expected value range of the candidate defect type defect in each feature dimension. For the current single-sampled feature data packet to be judged, calculate the deviation between each feature value in its feature combination and the corresponding value in the expected value range; Based on predefined expert experience rules, different weight coefficients are assigned to different feature dimensions. Among them, feature dimensions that reflect the abrupt changes in surface texture are given higher weights, while feature dimensions that reflect the stability of mechanical signals are given medium weights. Based on the deviation of each feature dimension and its corresponding weight coefficient, calculate the comprehensive difference between the current feature combination and the standard feature template of each candidate defect type; The candidate defect type with the smallest overall difference is selected as the target defect type, and the overall difference is converted into the matching confidence score.

7. The sampling system for blood collection safety needle detection as described in claim 6, characterized in that, It also includes a sampling strategy feedback optimization unit, used for: Collect all single-sample feature data packets and their corresponding defect labels generated at the end of a complete monitoring cycle; Analyze the distribution changes of the defect labels in different time periods to identify warning periods when the frequency of defect occurrence increases; The original control parameters corresponding to the warning period are traced back, namely the historical values ​​of the normalized baseline sampling frequency, image feature dynamic index, and stress state composite index. The historical control parameters of the warning period are input into a strategy correction network. After training, the strategy correction network can output the adjustment amount of the modulation rule in the three-dimensional sampling frequency control model. Based on the adjustment amount, the calculation rules for frequency disturbance and frequency compensation in the first frequency modulation layer and the second frequency modulation layer are dynamically updated, so that the system can automatically increase the sampling density when facing similar operating conditions in subsequent monitoring.

8. The sampling system for blood collection safety needle detection as described in claim 7, characterized in that, The workflow of the policy correction network includes: A training sample set is constructed, and each sample consists of a set of historical control parameters and a set of expected sampling frequency adjustment amounts. The expected sampling frequency adjustment amounts are manually labeled by domain experts based on the situation during historical warning periods. The policy correction network is constructed using a multilayer perceptron structure, with historical control parameters as the input layer and the desired sampling frequency adjustment as the output layer. The policy correction network is trained in a supervised manner using the training sample set until the error between the predicted adjustment amount output by the network and the expected adjustment amount labeled by the expert is lower than a preset threshold. During the deployment phase, the historical control parameters of the early warning period obtained in real time are input into the trained policy correction network to directly obtain the corresponding modulation rule adjustment amount.

9. The sampling system for blood collection safety needle detection as described in claim 1, characterized in that, It also includes a data visualization and report synthesis unit, used for: Receive a sequence of defect tags with time stamps from the defect mode determination unit; The temporally sequenced sampling feature library is received from the sampling execution and feature extraction unit. The defect label sequence is aligned and fused with the time-series sampling feature library along the time axis to generate complete timeline data containing original features, defect judgment results and confidence levels. Based on the complete timeline data, a two-dimensional chart is automatically generated, where the horizontal axis represents time and the vertical axis simultaneously displays the key feature value curves and the distribution of defect types. Based on the preset report template, the two-dimensional charts, key statistical data, and conclusive text descriptions are integrated to output the final electronic test report document.

10. The sampling system for blood collection safety needle detection as described in claim 1, characterized in that, The trend decomposition of the internal stress reading sequence to separate the long-term stress baseline component and the short-term stress fluctuation component specifically includes: Establish a coordinate system with sampling time as the horizontal axis and stress reading as the vertical axis, and plot the internal stress reading sequence as the original stress curve; The original stress curve is processed using the sliding window averaging method. The window length is set to a preset proportion of the total monitoring period, and the arithmetic mean of the stress readings in each window is calculated. Connect the arithmetic mean points of all sliding windows to form a smooth trend line, which is used as the long-term stress baseline component. The residual sequence is obtained by subtracting the long-term stress baseline component values ​​at corresponding times from the original stress curve point by point. The standard deviation of the residual sequence is calculated, and residual points with absolute values ​​exceeding three times the standard deviation are identified as abnormal fluctuation points and removed. The residual sequence after removing abnormal fluctuation points is used as the short-term stress fluctuation component, and the amplitude sign of each fluctuation point is recorded.

Images