Heat transfer component fault diagnosis system based on deep learning

Abstract

The invention discloses a heat transfer component fault diagnosis system based on deep learning. The heat transfer component fault diagnosis system comprises six modules. The model training data management module divides a normal operation period and a function failure period according to design deformation threshold values of different heat transfer parts, collects deformation quantity and operation time and stores the deformation quantity and the operation time as a data set; the life prediction and health assessment module constructs a model by using the data set, and predicts the time when the target part reaches a deformation threshold value; the deformation feature sensing module collects target part data and obtains a primary deformation value; an intelligent maintenance strategy screening module calls a strategy library to screen an adaptive scheme and records the adaptive scheme into a training set; the maintenance effect verification module calculates the secondary deformation quantity after maintenance at an interval of prediction time, and compares the secondary deformation quantity with a threshold value; and the self-learning strategy optimization module brings effective schemes into an optimal selection set according to a comparison result, and ineffective schemes enter a screening set for manual judgment. The system realizes intelligent diagnosis and maintenance optimization of faults of the heat transfer component.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault diagnosis, more particularly, to a heat transfer component fault diagnosis system based on deep learning. BACKGROUND

[0002] The existing heat transfer component fault diagnosis system mostly adopts multi-source data fusion learning.

[0003] Firstly, the data pool is cooperatively constructed. The real-time monitoring pool built by the system covers multiple types of key parameters, including dynamic physical quantities, chemical characteristics and environmental parameters. The historical case library is classified in three dimensions according to the deformation threshold "component type-failure mode-working condition", and a large number of complete records are stored, including the whole life cycle monitoring curve from the commissioning of the component to the failure. Then, the deformation threshold cross-domain data association mining mechanism is established. Through the association algorithm, the system automatically identifies the hidden association between the data, including strong association rules and time sequence association. The strong association rule refers to the high confidence of the occurrence of a certain fault under a certain parameter combination; the time sequence association reflects the time sequence relationship between the parameter change trend and the fault risk improvement. In the data fusion process, the evidence theory is used to handle conflicting data, and the influence of measurement error is reduced through credibility allocation, thereby improving the accuracy of feature mapping. Finally, the deformation threshold edge computing and cloud cooperation architecture is established. In order to solve the problem of data transmission delay in industrial sites, the system completes the preprocessing of real-time data such as outlier elimination and normalization at the edge node, and then only uploads the key feature values to the cloud for fusion analysis with the historical case library. This architecture can improve the data processing efficiency and ensure that the association rules can respond within milliseconds when dealing with sudden working conditions, thereby ensuring their effectiveness.

[0004] 1. Traditional methods are mostly limited to judging the current fault and cannot predict the fault time. This makes it often only passive maintenance after the fault occurs, which can easily cause unplanned shutdown, disrupt production rhythm, reduce production efficiency, and is difficult to ensure production continuity and stability. 2. The traditional method needs to manually identify the failure threshold of each component, which is time-consuming and prone to error.

[0005] 3. The traditional model needs to observe the entire failure period of the component, which is time-consuming and labor-intensive for components with long service life, seriously slowing down the model development efficiency, making the model unable to be applied in time, and difficult to quickly meet the needs of equipment maintenance for the model, affecting the timeliness and effectiveness of maintenance work. SUMMARY

[0006] In order to overcome the above-mentioned defects of the prior art, the present application provides a heat transfer component fault diagnosis system based on deep learning, which solves the problems raised in the above background art through the following scheme.

[0007] A heat transfer component fault diagnosis system based on deep learning, comprising a model training data management module: based on different heat transfer component design deformation thresholds, the heat transfer components are divided into normal operation period and functional failure period according to the deformation amount of the heat transfer components, the deformation amount of a certain number of heat transfer components is calculated and its running time is recorded, and it is transmitted and stored in the data set; A life prediction and health assessment module: using the information in the data set to construct a prediction model, and predicting the time when the target heat transfer component reaches the deformation threshold; A deformation feature perception module: data acquisition is performed on the target heat transfer component to obtain a first deformation value of the heat transfer component; An intelligent maintenance strategy screening module: calling a preset maintenance strategy library, screening a maintenance scheme suitable for the process flow, and entering this scheme into the training set; A maintenance effect verification module: at intervals, the system prediction time is calculated, the deformation amount of the heat transfer component after maintenance is calculated again, a second deformation value is obtained, and the second deformation value is compared with the deformation threshold; A self-learning strategy optimization module: if the second deformation value is less than the deformation threshold, the maintenance scheme is included in the preferred scheme set; if the second deformation value is greater than the deformation threshold, the maintenance is included in the screening set for manual judgment.

[0008] Preferably, the model training data management module reads the current wear amount D w and the running time t w of the heat transfer component of the same model as the target heat transfer component.

[0009] Preferably, the present example selects a weibull model to judge the time when the heat transfer component reaches the deformation threshold, and uses a power law model to verify the prediction result.

[0010] Preferably, a double camera is used to collect images of the surface points of the heat transfer component.

[0011] The present application has the advantages of 1. The present application predicts the failure time, gives the relevant personnel enough time to prepare in advance, can timely purchase spare parts, arrange maintenance personnel, etc., thereby effectively reducing the probability of unplanned shutdown and ensuring the continuity and stability of production.

[0012] 2. The present application sets the threshold value by a standardized method, avoids the standard difference problem caused by manual identification lacking unified standard in the traditional method, greatly simplifies the threshold setting process, and saves a lot of time.

[0013] 3. This invention establishes a database to train the model based on the current wear, without needing to observe the entire failure cycle, which greatly shortens the model training cycle and enables usable models to be obtained more quickly. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the overall structure of the present invention.

[0015] Figure 2 This is a schematic diagram of the system architecture of the present invention.

[0016] Figure 3 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] like Figure 1 The system shown is a deep learning-based fault diagnosis system for heat transfer components. It performs two deformation calculations on the target heat transfer component at intervals of model prediction time. If the second deformation does not reach the predicted value after using the maintenance plan, it proves that the maintenance plan is not a high-quality maintenance plan. If the second deformation reaches the predicted value after using the maintenance plan, it proves that the maintenance plan is not a poor-quality maintenance plan.

[0019] like Figure 2 The system shown is a deep learning-based fault diagnosis system for heat transfer components, including a model training data management module, a life prediction and health assessment module, a deformation feature perception module, an intelligent maintenance strategy screening module, a maintenance effect verification module, and a self-learning strategy optimization module.

[0020] like Figure 3 The system shown is a deep learning-based fault diagnosis system for heat transfer components. It establishes a model using the current wear of heat transfer components of the same model as the target heat transfer component. It calculates the individual theoretical life and the group theoretical life of the same model heat transfer components separately. If the error between the individual theoretical life and the group theoretical life is within a reasonable range, the model is successfully established.

[0021] The model training data management module designs deformation thresholds based on different heat transfer components. Based on the deformation of the heat transfer components, it divides them into normal operation periods and functional failure periods. For a certain number of heat transfer components, it calculates their deformation and records their operating time, transmitting and storing the data in the dataset. Different deformation thresholds are designed for different heat transfer components; the calculation formula is as follows: Deformation threshold , the deformation threshold is m. Wherein, △T is the difference between the highest temperature generated in the production process of the heat transfer component and the temperature at the end of the process, in ℃; α is the material expansion coefficient, the elongation per unit length of the material when it rises by 1K, in ℃ -1 ; L is the material gap diameter of the heat transfer component, in m; E cff is the effective elastic modulus, which represents the actual stiffness of the material at the working temperature, in Pa, measured by high temperature tensile test or referred to the material manual; σ is the allowable stress of the material, which represents the maximum allowable stress (avoiding plastic deformation or fracture) of the material for long-term safe operation at the working temperature, reflecting the material's ability to resist deformation and reflecting the material's thermal deformation sensitivity, in Pa; S is the safety factor, which reflects the safety redundancy of engineering design, and is used to avoid structural failure caused by accidental factors such as material performance fluctuation, working condition deviation, calculation error, etc. According to the importance classification of components, the key path components take 2.0-2.5, and the non-key components take 1.5-1.8; r is the structure correction coefficient, which reflects the influence of the structure form of the heat transfer component on the deformation sensitivity, and the straight pipe takes 1.0, the elbow pipe takes 1.2-1.5, which increases with the decrease of the curvature radius, and the special-shaped components take 1.5-2.0. When the number of samples of the same type of heat transfer component is sufficient, usually ≥30, the multi-sample parallel collection mode is adopted, and the data representativeness is improved through the comparison of the differences between samples. When the number of samples of the same type is insufficient, <10, the collection period of single sample is prolonged, and the key stage collection frequency is encrypted to construct a longitudinal time series data chain.

[0022] The life prediction and health assessment module uses the information in the data set to build a prediction model and predict the time for the target heat transfer component to reach the deformation threshold; the data set is used to establish a prediction model to predict the time for the total deformation of the target heat transfer component to reach the deformation threshold, and the model establishment steps are as follows, S1, modeling target determination and sample data collection: calling the information in the data set, reading the current wear D w and the running time t w of the heat transfer component of the same type as the target heat transfer component; S2, power-law wear model establishment: the power-law wear model is used to describe the objective law of the evolution of the wear D(t) of a certain product with the increase of time t. The formula is D(t) is the wear amount of the heat transfer component at time t; t is the actual running time; k is the wear coefficient, representing the change trend of the wear rate, the smaller k is, the more wear-resistant the material is; m is the wear index, representing the change trend of the wear rate; and γ is the wear incubation period, representing the stage in which the wear amount in the initial running of the equipment is negligible. The parameters k, m and γ in the power-law wear model are calculated. At least three groups of running time and corresponding wear amount samples are collected and recorded as {(D1, t1), (D2, t2),..., (D w , t w )}. Taking the natural logarithm of both sides of the formula, the linear equation is obtained 2 ; the values of D(t) and t are adjusted, and the calculation is repeated until R 2 is maximum, and generally R 2 needs to be greater than or equal to 0.9, at which time γ is the optimal value. The slope of the linear equation is m, and the intercept A = lnk. The values of k and m are calculated by substituting different D(t) and t.

[0023] S3, individual theoretical life calculation: after the parameters k, γ and m are calculated, the power-law model is mathematically transformed to obtain the single-sample theoretical life formula, and the conversion from the wear index to the time index is completed; the formula is ; a set threshold D max is substituted into D0 to obtain the time t0 at which a single heat transfer component reaches the threshold of deformation, and multiple groups of t0 are collected to form a life data set {t a , t b , t c ,..., t n}. S4, group theoretical life calculation: the cumulative distribution function and the reliability function of the weibull distribution are introduced to quantify the statistical distribution characteristics of the group life; the cumulative distribution function formula is , which represents the proportion of samples with a life less than or equal to t in the group; the reliability function is , which represents the proportion of samples with a life greater than t in the group; wherein β is the shape parameter, and η is the scale parameter; the life data set {t a , t b , t c ,..., t n} is arranged in ascending order to obtain an ordered life data set {t1, t2, t3,..., t max}; and the specific values of the parameters γ, β and η are calculated according to the data; the median rank algorithm is used to calculate , i is the order number. By twice logarithmic transformation, the cumulative distribution function is transformed into a linear equation , First logarithmic change , Second logarithmic change Let Let Establish a linear equation Based on the linear regression results corresponding to the optimal gamma: calculate the slope a = beta, the intercept b = - beta * ln eta, find beta, eta.

[0024] S5, model verification and eligibility determination: extract the specific quantile life to calculate the group theoretical life t q , individual theoretical life t0 is compared, and the gap is within a reasonable range. The relative difference between the group theoretical life and the individual theoretical life is calculated, In this example, the relative difference is less than 15%, so the model is considered correct. If the error is greater than 15%, reselect the data in the data set for modeling.

[0025] The deformation feature perception module collects data of the target heat transfer component to obtain a first deformation value of the heat transfer component; the deformation of the target heat transfer component is based on a digital image method, and the surface of the unused heat transfer component needs to be pretreated; First, a discrete feature point system with stable recognition performance needs to be designed. The positions of the points are uniformly distributed macroscopically, but each local area is random. The ideal point should cover 3-5 pixels, and the spacing between points should be at least 2-3 times the diameter of the point to ensure that each subset contains enough unique point features. For high-temperature working conditions, the marker point material needs to be consistent with the component body to avoid marker point falling off due to the difference in thermal expansion coefficient. In a humid environment, the dot matrix area needs to be passivated to improve corrosion resistance and ensure that there is no obvious rust within 1 year.

[0026] Before the experiment, obtain waste or small samples of the same material as the heat transfer component. Perform laser marking tests on them, systematically adjusting laser power, speed, frequency, pulse width, and marking frequency. Laser dot marking is used to matte finish the heat transfer component. First, eliminate the interference of light reflection. The surface of the heat transfer component usually has a mirror reflection, which can cause the dot-marked area in the image to appear overexposed or have light spots, hiding the edge features of the dot mark and affecting the accuracy of subsequent wear measurement in the image. Second, enhance the contrast. The matte surface can highlight the gray difference between the dot mark and the background, making it easier for the image recognition algorithm to accurately extract the position and shape of the mark, providing a reliable image basis for subsequent wear calculation based on mark displacement or deformation. The surface of the heat transfer component that has been matte finished uses laser direct marking technology to create a dot matrix structure on the surface of the heat transfer component, serving as a reference mark for subsequent displacement measurement. Use a camera to capture dot matrix images at different time intervals. Select a high-speed camera with appropriate field of view and resolution. Use a fixed-focus lens to reduce focal plane changes. Select a filter with a center wavelength that matches the wavelength of the illumination light source, allowing only the illumination light to pass through and greatly suppressing the interference of the component's own radiation, thereby improving the image signal-to-noise ratio and reducing image distortion caused by air thermal disturbance. Use a high-precision ceramic calibration board to capture reference images at room temperature and under no load. Capture at least 20 images from multiple angles to ensure that the calibration error is less than 0.02 pixels. Start the production line and apply the heat load according to the predetermined program. Synchronously trigger the dual-camera to capture images at the preset frequency.

[0027] For a certain mark point, calculate the three-dimensional coordinate displacement of the mark point before and after the time period through image matching algorithm.

[0028] Take the surface of the heat transfer component at room temperature and under no load as the reference. The dot matrix image captured at this time is the standard. The three-dimensional coordinate of the mark point is denoted as P0=(X0, Y0, Z0), For the preprocessed laser dot matrix, extract the point information. Use the gray contrast of the matte finish to automatically segment the mark point area through Otsu algorithm. Sub-pixel positioning: perform ellipse fitting on the segmented area to calculate the sub-pixel coordinates of the point. For the left and right camera images at the same time, use binocular disparity and calibration parameters to calculate the three-dimensional coordinates of the mark point in the world coordinate system: , , where B is the binocular baseline distance, i.e. the straight-line distance between the optical centers of the left and right cameras, which is a hardware inherent parameter of the binocular vision system and needs to be obtained through precise measurement, directly affecting the depth calculation accuracy. The longer the baseline, the higher the depth resolution for distant objects. xis the focal length of the left camera in the x direction, in pixels, determined by camera calibration, reflecting the scaling ratio of the lens projecting a three-dimensional space point to a two-dimensional image plane, the greater the focal length, the more obvious the parallax change at the same depth; d is the left and right parallax, that is, the difference between the x coordinates u of the same marker point in the left camera image, in pixels, which is the core basis for binocular vision to obtain depth and the key intermediate variable for three-dimensional coordinate calculation; c is the principal point x coordinate of the left camera, that is, the x coordinate of the intersection of the optical principal axis and the image sensor plane, in pixels, determined by camera calibration, which is the reference point for converting pixel coordinates to world coordinates and is used to eliminate the influence of image center offset on coordinate calculation; v is the y direction pixel coordinate of the marker point in the left camera image, in pixels, also obtained by sub-pixel positioning; c is the principal point y coordinate of the left camera, in pixels, determined by camera calibration, which has a similar effect to c and is used to correct the coordinate reference in the vertical direction; f is the focal length of the left camera in the y direction, in pixels, because the pixels of the camera sensor may not be square, f may be different from f, and they need to be obtained by calibration respectively to ensure the accuracy of the vertical direction coordinate calculation. l is the x coordinate u of the same marker point in the right camera image, in pixels, which is the core basis for binocular vision to obtain depth and the key intermediate variable for three-dimensional coordinate calculation; c is the principal point x coordinate of the left camera, that is, the x coordinate of the intersection of the optical principal axis and the image sensor plane, in pixels, determined by camera calibration, which is the reference point for converting pixel coordinates to world coordinates and is used to eliminate the influence of image center offset on coordinate calculation; v is the y direction pixel coordinate of the marker point in the left camera image, in pixels, also obtained by sub-pixel positioning; c is the principal point y coordinate of the left camera, in pixels, determined by camera calibration, which has a similar effect to c and is used to correct the coordinate reference in the vertical direction; f is the focal length of the left camera in the y direction, in pixels, because the pixels of the camera sensor may not be square, f may be different from f, and they need to be obtained by calibration respectively to ensure the accuracy of the vertical direction coordinate calculation. r x l y x y y x

[0029] The displacement of the point is ΔX = X - X0, ΔY = Y - Y0, and ΔZ = Z - Z0. The total deformation of the heat transfer component is calculated according to the displacement of the point, and the total deformation of the heat transfer component is L = ΔX + ΔY + ΔZ, in meters. The total deformation of the heat transfer component L is substituted into the Weibull model to calculate the prediction time of the heat transfer component deformation value reaching the deformation threshold.

[0030] The intelligent maintenance strategy screening module calls the preset maintenance strategy library, screens out the adaptive maintenance scheme, and enters this scheme into the training set. The strategy library stores maintenance schemes according to three dimensions of process flow type, component characteristics, and running time, divides the process flow into high temperature and high pressure working condition, intermittent running working condition, and corrosive medium working condition, associates metal material, non-metal material, straight pipe, and elbow pipe according to component characteristics, and matches corresponding maintenance means in combination with the three types of situations. The execution scheme with the highest adaptation degree is screened out, and the training set is entered in a fixed format. The fields include: basic information, input data, and execution result: the basic information includes scheme ID, application component ID, execution time, and process flow parameters; the input data includes cumulative running time before maintenance, prediction running time to threshold, and component characteristic data; the execution result is to be supplemented, and the running state data after maintenance and threshold comparison result are filled in by the secondary verification module. ​​​​​​​​​

[0031] The maintenance effect verification module interval system predicts the time, and the secondary deformation variable value is obtained by calculating the deformation variable of the heat transfer component after maintenance, and the secondary deformation variable value is compared with the deformation threshold value; the total deformation variable of the current heat transfer component is calculated in the same way, and the total deformation variable is compared with the deformation threshold value. Calculate According to the life extension ratio value, the maintenance scheme is classified.

[0032] The self-learning strategy optimization module determines whether the maintenance strategy is effective, if the secondary deformation variable value is less than the deformation threshold value, the maintenance scheme is included in the preferred scheme set; if the secondary deformation variable value is greater than the deformation threshold value, the maintenance is included in the screening set for manual judgment. For the scheme with secondary deformation variable exceeding the standard, generate a manual judgment form, including basic information such as component model, maintenance scheme ID and operating condition parameters; abnormal analysis including comparison of deformation variable curves before and after maintenance and normal curves of similar components at the same period, to determine whether the invalid maintenance is caused by sudden change of working condition, whether the maintenance parameters need to be adjusted, and whether it belongs to the defect of the scheme. After the manual judgment result is entered, the system automatically includes the adjusted scheme in the verification set, and the next maintenance test is preferentially arranged. The first choice for daily maintenance is the scheme with life extension ratio less than 1, especially less than 0.5.

[0033] Secondly, the drawings of the disclosed embodiments of the present application only involve the structures involved in the disclosed embodiments of the present application, other structures can refer to the usual design, and in the case of no conflict, the same embodiment and different embodiments of the present application can be combined with each other; Finally, the above only describes the preferred embodiments of the present application, and is not used to limit the present application, any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A deep learning-based heat transfer component failure diagnosis system, characterized by, The model training data management module includes: based on the design deformation threshold of different heat transfer components, the heat transfer components are divided into normal operation period and functional failure period according to the deformation of the heat transfer components; the deformation of a certain number of heat transfer components is calculated and the running time is recorded; the data is transmitted and stored in the data set; The life prediction and health assessment module: using the information in the data set to construct a prediction model, and predicting the time when the target heat transfer component reaches the deformation threshold; The deformation feature sensing module: data acquisition is performed on the target heat transfer component to obtain the first deformation value of the heat transfer component; The intelligent maintenance strategy screening module: calling the preset maintenance strategy library, screening the maintenance scheme suitable for the process flow, and recording the scheme into the training set; The maintenance effect verification module: interval system prediction time, secondary calculation of the deformation of the heat transfer component after maintenance, obtaining the secondary deformation value, and comparing the secondary deformation value with the deformation threshold; The self-learning strategy optimization module: if the secondary deformation value is less than the deformation threshold, the maintenance scheme is included in the optimal scheme set; if the secondary deformation value is greater than the deformation threshold, the maintenance is included in the screening set for manual judgment.

2. The deep learning-based heat transfer component fault diagnosis system of claim 1, wherein, The data set is used to establish a prediction model to predict the time when the total deformation of the target heat transfer component reaches the deformation threshold, and the model establishment steps are as follows, S1, modeling target determination and sample data acquisition: call the information in the data set, read the current wear amount D of the heat transfer component of the same model as the target heat transfer component w and its running time t w ; S2, power law wear model is established: the power law wear model is used for describing the objective law that the wear D(t) of a product increases with time t. The formula is , D(t) is the wear of the heat transfer component at time t; t is the actual running time; k is the wear coefficient, which represents the change trend of the wear rate, the smaller k is, the more wear-resistant the material is; m is the wear index, which represents the change trend of the wear rate; γ is the wear incubation period, which represents the stage that the wear amount in the initial stage of equipment operation can be ignored; S3, individual theoretical life calculation: after calculating the parameters k, γ, m, the power law model is mathematically transformed to obtain a single sample theoretical life formula, and the conversion from wear index to time index is completed; the formula is The deformation threshold D max is set a , t b , t c ,..., t n}; S4, Population Theory Life Calculation: Introduce the cumulative distribution function and reliability function of Weibull distribution to quantify the statistical distribution characteristics of population life; the cumulative distribution function formula is , which represents the sample proportion of life ≤t in the population; the reliability function is: , which represents the sample proportion of life >t in the population; where β is the shape parameter and η is the scale parameter; arrange the life data set {t a , t b , t c ,..., t n} from small to large to obtain the ordered life data set {t1, t2, t3,..., t max}; and calculate the specific values of parameters γ, β, and η according to the data; S5, model verification and qualification judgment: extract a specific quantile life and substitute it into the weibull model to calculate the group theoretical life and individual theoretical life t0, and compare the difference within a reasonable range to determine whether the model is established qualified.

3. The deep learning-based heat transfer component fault diagnosis system of claim 2, wherein, The R(t) represents the percentage of samples that can work normally to the time point, and has no fixed value. The R(t) value is specified according to actual requirements, and the weibull model is used to calculate the time t0 when the target heat transfer component reaches the failure period.

4. The deep learning-based heat transfer component fault diagnosis system of claim 1, wherein, The deformation of the target heat transfer component is based on the digital image method, which needs to pretreat the surface of the unused heat transfer component; first, a discrete feature point system with stable recognition is designed, and a laser direct marking technology is used on the surface of the heat transfer component treated with matte to construct a dot matrix structure as a reference mark for subsequent displacement measurement; the camera is used to collect dot matrix images at different time periods, and for a certain mark point, the three-dimensional coordinate displacement of the mark point before and after the time period is calculated through image matching algorithm.

5. The deep learning-based heat transfer component fault diagnosis system of claim 3, wherein, The displacement of the mark point three-dimensional coordinates is denoted as Δx, Δy, Δz, and the total deformation of the heat transfer component is denoted as , with the unit of m.

6. The deep learning-based heat transfer component fault diagnosis system of claim 1, wherein, Different deformation threshold values are designed for different heat transfer components, and the calculation formula is , with the unit of m.

7. The deep learning-based heat transfer component fault diagnosis system of claim 1, wherein, The total deformation of the heat transfer component is compared with the deformation threshold, and the heat transfer component with total deformation less than the deformation threshold is judged as normal operation period, and the heat transfer component with total deformation greater than the deformation threshold is judged as functional failure period.

8. The deep learning-based heat transfer component fault diagnosis system of claim 1, wherein, According to the process route, different maintenance schemes are set, and the total deformation of the target heat transfer component is calculated again using the image matching algorithm after reaching the target time. If it is lower than the deformation threshold, the maintenance scheme is effective; if it is higher than the deformation threshold, the maintenance scheme is invalid.

Images