A visual-based device health quantification detection method

Abstract

The application discloses a kind of based on visual equipment health quantification detection method, which comprises: video stream is framed and image normalization processing is carried out;Image is divided into multiple sub-blocks, and the static and dynamic characteristics of each sub-block are extracted;The mean vector, covariance matrix and corresponding static and dynamic feature tolerance interval of each sub-block are calculated, and these parameters are updated with time sliding;When new real-time features appear, the deviation of each sub-block is calculated;According to the deviation and the preset threshold, it is determined whether the sub-block is abnormal;The overall abnormal quantification value of the equipment is obtained by the abnormal propagation quantification model, and the comprehensive health index of the equipment is finally calculated.The application has self-adaptive updating capability, can finely detect and locate local abnormalities, and realize the quantitative evaluation of health status, improve the accuracy, sensitivity and practicality of detection.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection technology for industrial equipment, and in particular to a vision-based quantitative detection method for equipment health. Background Technology

[0002] Stable operation of industrial equipment is crucial for ensuring production efficiency and safety. Traditional methods for monitoring and diagnosing equipment health status, such as those relying on physical sensors for vibration, temperature, and pressure, face numerous challenges in practical applications. For example, achieving comprehensive monitoring often requires deploying various types of sensors at multiple points on the equipment. This not only increases hardware procurement costs but also introduces complex wiring, installation, commissioning, and subsequent maintenance challenges, resulting in high overall deployment and operation costs.

[0003] Existing methods based on traditional sensors or some early visual inspection methods have the following specific shortcomings in terms of intelligence and precision:

[0004] Poor adaptability: Many existing methods rely on pre-defined fixed parameters to determine whether equipment is abnormal. However, equipment states often change gradually with accumulated operating time, environmental changes, or normal wear and tear. Fixed parameters are difficult to adapt to such dynamic changes, which may lead to normal state drift being misjudged as abnormal, or insufficient sensitivity to gradually developing early faults. The lack of effective adaptive learning and dynamic adjustment capabilities makes the system poorly adaptable to gradual changes in equipment states.

[0005] Global Feature Ignoring vs. Local Anomaly Ignoring: Some detection methods tend to extract equipment features from a holistic or global perspective for judgment. When local, subtle anomalies occur in the equipment (such as slight corrosion of a single component, micro-cracks, color changes caused by localized overheating, etc.), these local anomaly signals are easily "overwhelmed" or averaged by the overall, dominant normal state data, leading to missed detections. This problem is particularly prominent in large or complex equipment.

[0006] Binary Judgment and Lack of Quantitative Assessment: Many existing methods ultimately provide only a binary judgment result: the equipment is "normal" or "abnormal." This simple judgment cannot quantify the severity of the abnormality or the degree of deterioration in health status. In actual operation and maintenance, managers not only want to know if the equipment is abnormal, but also want to understand the scope and extent of the abnormality, as well as how much the health status has declined, in order to make more accurate maintenance decisions and resource allocation. Summary of the Invention

[0007] This invention primarily addresses the technical problems of existing technologies, such as poor adaptability due to fixed parameters, missed detection of local anomalies due to global misjudgment, and lack of quantitative evaluation by providing only binary judgment. It provides a vision-based quantitative detection method for device health, which has adaptive update capabilities, can perform fine-grained detection on local areas, and can quantitatively evaluate the health status of devices.

[0008] The present invention addresses the aforementioned technical problems primarily through the following technical solution: a vision-based method for quantitative detection of device health, comprising the following steps:

[0009] S1: Generate a sampled image sequence by extracting frames from the video stream under normal conditions at a fixed frequency; the frame extraction frequency can generally be set to 10-30.

[0010] S2: Perform multimodal normalization on the sampled image sequence to obtain a normalized image sequence;

[0011] S3: Divide each image in the normalized image sequence into N. rows ×N cols Each block;

[0012] S4: Extract the static and dynamic features of each sub-block; static and dynamic features are collectively referred to as spatiotemporal features;

[0013] S5: Calculate the mean vector and covariance matrix of each sub-block;

[0014] S6: Calculate the static and dynamic allowable intervals of each sub-block; the sequence used to calculate the mean vector, covariance matrix, static and dynamic allowable intervals of each sub-block is continuously updated over time as long as no anomalies occur. Generally, a normal state sequence of 30-300 minutes can be collected to calculate the above parameters, and the window is slid every 5 minutes.

[0015] S7: The newly acquired sampled image sequence is processed through steps S2-S5 to obtain real-time static features and real-time dynamic features;

[0016] S8: Calculate the deviation of each sub-block: when the real-time static feature f of the s-th sub-block s_t The static allowable interval C of the s-th sub-block s_s Furthermore, the real-time dynamic characteristics of the s-th sub-block are within the dynamic tolerance range C of the s-th sub-block. d_s Within the time frame, the deviation P of the s-th sub-block s =0; otherwise, the deviation P of the s-th sub-block s =[(f s_t -μ) T Σ -1 (f s_t-μ)] / γ, where γ is the normalization coefficient, γ=max(diag(Σ)); μ is the mean vector of the s-th sub-block, and Σ is the covariance matrix of the s-th sub-block;

[0017] S9: Sub-block anomaly determination: When the deviation of a sub-block is greater than or equal to the deviation threshold, the sub-block is considered abnormal; otherwise, the sub-block is considered normal. For core safety equipment, the deviation threshold is 0.5 to maintain high sensitivity, while for general equipment, the deviation threshold is set to 0.7 to reduce the false alarm rate. By setting different deviation thresholds for equipment of different importance, this invention can ensure high detection sensitivity for key equipment while appropriately reducing the possibility of false alarms for general equipment, thus achieving a balance between detection performance and practical application requirements.

[0018] S10: Anomaly Propagation Quantization: If there is only one sub-block anomaly, the anomaly quantization value P is... total =50%; if there are several sub-blocks that are anomalous and the anomalous sub-blocks are consecutive, then the anomalous quantification value P is 50%. total =1-0.5 n , where n is the number of anomalous sub-blocks; if several sub-blocks are anomalous and these anomalous sub-blocks are non-contiguous, then the anomalous quantization value P is... total =max(1-0.5 n1 , 1-0.5 n2 ...), where n1 is the number of abnormal sub-blocks in the continuous region of the first abnormal sub-block, n2 is the number of abnormal sub-blocks in the continuous region of the second abnormal sub-block, and so on; here, continuous means that the sub-blocks are adjacent sub-blocks;

[0019] S11: Calculate the overall health of the equipment using the following formula:

[0020] H index =100%−P total ∈[0%,100%];

[0021] When H index <60%: Trigger Level 1 alarm;

[0022] When H index <30%: Trigger Level 2 Emergency Shutdown.

[0023] When an alarm or shutdown occurs, the location information of the abnormal block is also provided, which enhances the interpretability of the detection results, facilitates quick location of the source of the problem, and guides maintenance work.

[0024] Preferably, step S4 specifically involves:

[0025] The chromaticity statistics of a sub-block are constructed as the static features of the sub-block. The chromaticity statistics f of the s-th sub-block at time t are... s t =[μ R,σ R 2 ,μ G ,σ G 2 ,μ B ,σ B 2 ] T 1≤s≤N rows ×N cols μ R μ G and μ B σ represents the pixel mean of the R, G, and B channels within the s-th sub-block at time t. R 2 σ G 2 and σ B 2 Let R, G, and B be the variances of the R, G, and B channels within the s-th sub-block at time t, respectively. The T in the upper right corner indicates transpose. Then, the chroma change of the sub-block at the same position in adjacent frames is calculated as the dynamic feature of the sub-block using the following formula:

[0026] Δf d_s t =‖f s t -f s t-1 ‖2;

[0027] Δf d_s (t) Let f be the dynamic feature of the s-th sub-block at time t. s t-1 Let be the chromaticity statistics of the s-th sub-block at time t-1.

[0028] Preferably, the mean vector μ and covariance matrix Σ of the s-th sub-block are determined by the following formula:

[0029] ;

[0030] ;

[0031] In the formula, f s f s t One of them, N total f s The number of f s (k) Indicates the k-th f s The "Σ" with subscripts and superscripts on the right side of the equals sign is the summation symbol, and the "Σ" on the left side of the equals sign represents the covariance matrix.

[0032] As a preferred option, the static allowable variation interval C of the s-th sub-block s_sDetermined in the following ways:

[0033] C s_s ={f s |(f s -μ) T Σ -1 (f s -μ)≤χ α 2 (6);

[0034] In the formula, α=0.95, χ α 2 (6) is the critical value of the chi-square distribution, which is usually 12.592. μ is the mean vector and Σ is the covariance matrix.

[0035] As a preferred option, the dynamic tolerance variation range C of the s-th sub-block is... s_s Determined by the following formula:

[0036] C d_s ={Δf d_s |Q 1_s -1.5×IQR≤Δf d_s ≤Q 3_s +1.5×IQR};

[0037] In the formula, Δf d_s For the dynamic characteristics of the s-th sub-block, Q 1_s and Q 3_s Let IQR be the quartile of the dynamic feature of the s-th sub-block; IQR is the interquartile range of the s-th sub-block, IQR = Q. 3_s -Q 1_s .

[0038] Q 1_s and Q 3_s Determined in the following ways:

[0039] All N values ​​in the s-th sub-block all The dynamic features are arranged in ascending order, and the (N)th dynamic feature is... all The dynamic features corresponding to +1) / 4 positions are Q. 1_s , the 3rd × (N) all The dynamic features corresponding to +1) / 4 positions are Q. 3_s If (N) all +1) / 4 or 3×(N) all If +1) / 4 is not an integer, then round up. Besides rounding up, it can also be rounded down or obtained by averaging two adjacent dynamic features.

[0040] Preferably, the multimodal normalization in step S2 includes geometric normalization and photometric normalization. Geometric normalization is achieved by scaling the image to a standard size using bilinear interpolation, and photometric normalization is achieved using the following formula:

[0041] R norm =μ target •R P / μ R ;

[0042] G norm =μ target •G P / μ G ;

[0043] B norm =μ target •B P / μ B ;

[0044] In the formula, μ R μ G and μ B These represent the pixel mean values ​​of the R, G, and B channels of the original image, respectively; and the target brightness mean value μ. target =128; R P G P and B P These are the pixel values ​​of the R, G, and B channels of the original image, respectively. norm G norm and B norm These are the pixel values ​​of the R, G, and B channels of the image after photometric normalization.

[0045] Preferably, each sub-block has a size of M×M pixels, where M∈[3,10], and the edge areas are filled with mirror images to ensure the integrity of the blocks.

[0046] The substantial effects of this invention are: 1. Enhanced dynamic adaptability and lower false alarm rate: This invention can effectively track and adapt to the gradual changes in equipment state caused by normal wear and tear, environmental changes, etc. This avoids false alarms caused by the inability of traditional fixed threshold methods to adapt to state drift, significantly improving the accuracy of detection and adaptability to the actual state of the equipment; 2. Significantly improved sensitivity of local anomaly detection: By using image block processing and performing independent deviation calculation and anomaly judgment for each sub-block, this invention can effectively capture and locate local and subtle anomalies occurring on the equipment. These local anomalies are easily overlooked in traditional global feature analysis methods, while this invention, through refined local monitoring, greatly reduces the missed detection rate of such anomalies, which is of great significance for early fault warning; 3. Refined quantitative assessment and interpretability of health status: This invention not only provides a binary judgment of whether the equipment is abnormal, but also realizes a quantitative assessment of the equipment health status through an anomaly propagation quantitative model and a comprehensive equipment health index. This allows managers to intuitively understand the severity of equipment malfunctions and overall health status; 4. Relying solely on visual data reduces deployment and maintenance costs: This invention primarily relies on visual image data for equipment health detection. Compared to traditional methods that require the installation of multiple physical sensors (such as vibration, temperature, etc.), it can significantly reduce the cost and complexity of sensor procurement, wiring, installation, and subsequent maintenance and calibration. It has obvious deployment advantages, especially for scenarios where existing equipment needs to be modified or where it is inconvenient to install physical sensors. Attached Figure Description

[0047] Figure 1 This is a flowchart of a vision-based quantitative detection method for device health according to the present invention. Detailed Implementation

[0048] The technical solution of the present invention will be further described in detail below through embodiments and in conjunction with the accompanying drawings.

[0049] Example: This example illustrates a vision-based method for quantitative detection of device health, such as... Figure 1 As shown, it includes the following steps:

[0050] S1: Generate a sampled image sequence by sampling frames from the video stream under normal conditions at a fixed frequency; the video stream is captured from a Hikvision DS-2CD2043G0-I camera at 25fps with a frame sampling frequency of 20.

[0051] S2: Perform multimodal normalization on the sampled image sequence to obtain a normalized image sequence;

[0052] S3: Divide each image in the normalized image sequence into N. rows ×N cols Each block;

[0053] S4: Extract the static and dynamic features of each sub-block; static and dynamic features are collectively referred to as spatiotemporal features;

[0054] S5: Calculate the mean vector and covariance matrix of each sub-block;

[0055] S6: Calculate the static and dynamic allowable intervals of each sub-block; the sequence used to calculate the mean vector, covariance matrix, static and dynamic allowable intervals of each sub-block is continuously updated over time as long as no anomalies occur. Generally, a normal state sequence of 30-300 minutes can be collected to calculate the above parameters, and the window is slid every 5 minutes.

[0056] S7: The newly acquired sampled image sequence is processed through steps S2-S5 to obtain real-time static features and real-time dynamic features;

[0057] S8: Calculate the deviation of each sub-block: when the real-time static feature f of the s-th sub-block s_t The static allowable interval C of the s-th sub-block s_s Furthermore, the real-time dynamic characteristics of the s-th sub-block are within the dynamic tolerance range C of the s-th sub-block. d_s Within the time frame, the deviation P of the s-th sub-block s =0; otherwise, the deviation P of the s-th sub-block s =[(f s_t -μ) T Σ -1 (f s_t -μ)] / γ, where γ is the normalization coefficient, γ=max(diag(Σ)); μ is the mean vector of the s-th sub-block, and Σ is the covariance matrix of the s-th sub-block;

[0058] S9: Sub-block anomaly determination: When the deviation of a sub-block is greater than or equal to the deviation threshold, the sub-block is considered abnormal; otherwise, the sub-block is considered normal. For core safety equipment, the deviation threshold is 0.5 to maintain high sensitivity, while for general equipment, the deviation threshold is set to 0.7 to reduce the false alarm rate. By setting different deviation thresholds for equipment of different importance, this invention can ensure high detection sensitivity for key equipment while appropriately reducing the possibility of false alarms for general equipment, thus achieving a balance between detection performance and practical application requirements.

[0059] S10: Anomaly Propagation Quantization: If there is only one sub-block anomaly, the anomaly quantization value P is... total =50%; if there are several sub-blocks that are anomalous and the anomalous sub-blocks are consecutive, then the anomalous quantification value P is 50%. total =1-0.5 n , where n is the number of anomalous sub-blocks; if several sub-blocks are anomalous and these anomalous sub-blocks are non-contiguous, then the anomalous quantization value P is...total =max(1-0.5 n1 , 1-0.5 n2 ...), where n1 is the number of abnormal sub-blocks in the continuous region of the first abnormal sub-block, n2 is the number of abnormal sub-blocks in the continuous region of the second abnormal sub-block, and so on; here, continuous means that the sub-blocks are adjacent sub-blocks;

[0060] S11: Calculate the overall health of the equipment using the following formula:

[0061] H index =100%−P total ∈[0%,100%];

[0062] When H index <60%: Trigger Level 1 alarm;

[0063] When H index <30%: Trigger Level 2 Emergency Shutdown.

[0064] When an alarm or shutdown occurs, the location information of the abnormal block is also provided, which enhances the interpretability of the detection results, facilitates quick location of the source of the problem, and guides maintenance work.

[0065] Step S4 specifically involves:

[0066] The chromaticity statistics of a sub-block are constructed as the static features of the sub-block. The chromaticity statistics f of the s-th sub-block at time t are... s t =[μ R ,σ R 2 ,μ G ,σ G 2 ,μ B ,σ B 2 ] T 1≤s≤N rows ×N cols μ R μ G and μ B σ represents the pixel mean of the R, G, and B channels within the s-th sub-block at time t. R 2 σ G 2 and σ B 2 Let R, G, and B be the variances of the R, G, and B channels within the s-th sub-block at time t, respectively. The T in the upper right corner indicates transpose. Then, the chroma change of the sub-block at the same position in adjacent frames is calculated as the dynamic feature of the sub-block using the following formula:

[0067] Δf d_s t =‖fs t -f s t-1 ‖2;

[0068] Δf d_s (t) Let f be the dynamic feature of the s-th sub-block at time t. s t-1 Let be the chromaticity statistics of the s-th sub-block at time t-1.

[0069] The mean vector μ and covariance matrix Σ of the s-th sub-block are determined by the following formula:

[0070] ;

[0071] ;

[0072] In the formula, f s f s t One of the elements, N total f s The number of f s (k) Indicates the k-th f s .

[0073] The static allowable variation interval C of the s-th sub-block s_s Determined in the following ways:

[0074] C s_s ={f s |(f s -μ) T Σ -1 (f s -μ)≤χ α 2 (6);

[0075] In the formula, α=0.95, χ α 2 (6) is the critical value of the chi-square distribution, which is usually 12.592. μ is the mean vector and Σ is the covariance matrix.

[0076] The dynamic tolerance range C of the s-th sub-block s_s Determined by the following formula:

[0077] C d_s ={Δf d_s |Q 1_s -1.5×IQR≤Δf d_s ≤Q 3_s +1.5×IQR};

[0078] In the formula, Δfd_s For the dynamic characteristics of the s-th sub-block, Q 1_s and Q 3_s Let IQR be the quartile of the dynamic feature of the s-th sub-block; let IQR be the interquartile range of the s-th sub-block, IQR = Q. 3_s -Q 1_s .

[0079] The multimodal normalization in step S2 includes geometric normalization and photometric normalization. Geometric normalization is achieved by scaling the image to a standard size using bilinear interpolation. Photometric normalization is implemented using the following formula:

[0080] R norm =μ target •R P / μ R ;

[0081] G norm =μ target •G P / μ G ;

[0082] B norm =μ target •B P / μ B ;

[0083] In the formula, μ R μ G and μ B These represent the pixel mean values ​​of the R, G, and B channels of the original image, respectively; and the target brightness mean value μ. target =128; R P G P and B P These are the pixel values ​​of the R, G, and B channels of the original image, respectively. norm G norm and B norm These are the pixel values ​​of the R, G, and B channels of the image after photometric normalization.

[0084] Each sub-block is M×M pixels in size, M∈[3,10], and the edge areas are filled with mirror images to ensure the integrity of the blocks.

[0085] The effect of this embodiment is as follows:

[0086] Quantitative accuracy: Experiments show that the anomaly probability error rate is ≤ 5.2% (compared to manual detection results).

[0087] Response speed: The delay from image acquisition to output Hindex is ≤ 0.8 seconds.

[0088] Interpretability: Provides an anomaly location map, marking the locations of high-probability sub-blocks.

[0089] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

[0090] Although this paper uses terms such as sampled image sequence, dynamic features, and covariance matrix extensively, the possibility of using other terms is not excluded. These terms are used merely for the convenience of describing and explaining the essence of this invention; interpreting them as any additional limitation would contradict the spirit of this invention.

Claims

1. A vision-based quantitative detection method for device health, characterized in that, Includes the following steps: S1: Generate a sampled image sequence by extracting frames from a video stream under normal conditions at a fixed frequency; S2: Perform multimodal normalization on the sampled image sequence to obtain a normalized image sequence; multimodal normalization includes geometric normalization and photometric normalization; S3: Divide each image in the normalized image sequence into N. rows ×N cols Each block; S4: Extract the static and dynamic features of each sub-block. The static feature of the s-th sub-block at time t is denoted as f. s t The dynamic characteristics of the s-th sub-block at time t are denoted as Δf. d_s t ; S5: Calculate the mean vector μ and covariance matrix ∑ of each sub-block in each image in the normalized image sequence using the static features of each sub-block. S6: Calculate the static and dynamic allowable intervals for each sub-block; S7: The newly acquired sampled image sequence is processed through steps S2-S5 to obtain real-time static features and real-time dynamic features; S8: Calculate the deviation of each sub-block: when the real-time static feature f of the s-th sub-block s_t The static allowable interval C of the s-th sub-block s_s Furthermore, the real-time dynamic characteristics of the s-th sub-block are within the dynamic tolerance range C of the s-th sub-block. d_s Within the time frame, the deviation P of the s-th sub-block s =0; otherwise, the deviation P of the s-th sub-block s =[(f s_t -μ) T Σ -1 (f s_t -μ)] / γ, where γ is the normalization coefficient, γ=max(diag(Σ)); μ is the mean vector of the s-th sub-block, and Σ is the covariance matrix of the s-th sub-block; The static allowable variation interval C of the s-th sub-block s_s Determined in the following ways: C s_s ={f s |(f s -m) T S -1 (f s -μ)≤χ α 2 (6)}; In the formula, α=0.95, χ α 2 (6) is the critical value of the chi-square distribution, μ is the mean vector of the s-th sub-block, and Σ is the covariance matrix of the s-th sub-block; f s It is the static feature of the s-th sub-block; The dynamic tolerance range C of the s-th sub-block s_s Determined by the following formula: C d_s ={Δf d_s |Q 1_s -1.5×IQR≤Δf d_s ≤Q 3_s +1.5×IQR}; In the formula, Δf d_s For the dynamic characteristics of the s-th sub-block, Q 1_s and Q 3_s Let IQR be the quartile of the dynamic feature of the s-th sub-block; IQR is the interquartile range of the s-th sub-block, IQR = Q. 3_s -Q 1_s ; S9: Sub-block anomaly determination: When the deviation of a sub-block is greater than or equal to the deviation threshold, the sub-block is considered to be an anomaly; Otherwise, the sub-block is considered normal; S10: Anomaly Propagation Quantization: If there is only one sub-block anomaly, the anomaly quantization value P is... total =50%; if there are several sub-blocks that are anomalous and the anomalous sub-blocks are consecutive, then the anomalous quantification value P is 50%. total =1-0.5 n , where n is the number of anomalous sub-blocks; if several sub-blocks are anomalous and these anomalous sub-blocks are non-contiguous, then the anomalous quantization value P is... total =max(1-0.5 n1 , 1-0.5 n2 ...), where n1 is the number of abnormal sub-blocks in the first continuous region of the abnormal sub-block, n2 is the number of abnormal sub-blocks in the second continuous region of the abnormal sub-block, and so on; S11: Calculate the overall health of the equipment using the following formula: H index =100%−P total ∈[0%,100%]; When H index <60%: Trigger Level 1 alarm; When H index <30%: Trigger Level 2 Emergency Shutdown.

2. The vision-based quantitative detection method for device health according to claim 1, characterized in that, Step S4 specifically involves: The chromaticity statistics of a sub-block are constructed as the static features of the sub-block. The chromaticity statistics f of the s-th sub-block at time t are... s t =[μ R ,σ R 2 ,μ G ,σ G 2 ,μ B ,σ B 2 ] T 1≤s≤N rows ×N cols μ R μ G and μ B σ represents the pixel mean of the R, G, and B channels within the s-th sub-block at time t. R 2 σ G 2 and σ B 2 Let R, G, and B be the variances of the R, G, and B channels within the s-th sub-block at time t, respectively. The T in the upper right corner indicates transpose. Then, the chroma change of the sub-blocks at the same position in adjacent frames is calculated as the dynamic feature of the sub-block using the following formula: Δ f d_s t =‖f s t - f s t-1 ‖2; Δf d_s t Let f be the dynamic feature of the s-th sub-block at time t. s t-1 Let be the chromaticity statistics of the s-th sub-block at time t-1.

3. The vision-based quantitative detection method for device health according to claim 2, characterized in that, The mean vector μ and covariance matrix Σ of the s-th sub-block are determined by the following formula: ； ； In the formula, f s (k) Indicates the k-th f s N total f s The number of.

4. The vision-based quantitative detection method for device health according to claim 1, characterized in that, The geometric normalization is achieved by scaling the image to a standard size using bilinear interpolation, and the photometric normalization is implemented using the following formula: R norm =m target •R P / m R ; G norm =m target •G P / m G ; B norm =m target •B P / m B ; In the formula, μ R μ G and μ B These represent the pixel mean values ​​of the R, G, and B channels of the original image, respectively; and the target brightness mean value μ. target =128; R P G P and B P These are the pixel values ​​of the R, G, and B channels of the original image, respectively. norm G norm and B norm These are the pixel values ​​of the R, G, and B channels of the image after photometric normalization.

5. The vision-based quantitative detection method for device health according to claim 1, characterized in that, Each sub-block is M×M pixels in size, M∈[3,10], and the edge areas are filled with mirror images to ensure the integrity of the blocks.

Images