Induction furnace coil temperature field visual monitoring method based on thermal imaging

Abstract

The invention relates to the technical field of image data processing, in particular to an induction furnace coil temperature field visual monitoring method based on thermal imaging, and the method comprises the steps: obtaining the three-dimensional temperature field data of an induction furnace coil, and extracting high-temperature voxels; an adaptive weight is constructed for each high-temperature voxel, and the weight is composed of two parts: 1, the local linearity of the voxel is determined by calculating the projection density of the voxel in multiple directions; secondly, obtaining a colinearity score by calculating the consistency of the candidate voxel and other candidate voxels in the neighborhood of the candidate voxel in the local main direction; and performing weighted Hough transform on the candidate voxels to vote by taking the product of the two as an adaptive weight. Through the intelligent weighting mechanism, real linear fault signals can be effectively amplified, noise interference is suppressed, and the detection accuracy and robustness of early discontinuous linear high-temperature faults are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image data processing, and in particular to an induction furnace coil temperature field visualization monitoring method based on thermal imaging. BACKGROUND

[0002] The induction furnace is a key equipment in modern metallurgical industry. The induction coil operates in harsh conditions such as strong electromagnetic interference, and is prone to faults such as inter-turn short circuit. Using infrared thermal imaging technology combined with three-dimensional reconstruction to obtain and analyze the real-time three-dimensional temperature field of the coil is an effective technical path to realize early warning of faults.

[0003] In fault diagnosis practice, key faults such as inter-turn short circuit often show linear or band-shaped high-temperature regions distributed along the inter-turn gap in the three-dimensional temperature field. Therefore, how to identify and locate these linear high-temperature regions from complex three-dimensional temperature data has become a key problem to break through the bottleneck of induction furnace coil monitoring technology.

[0004] In the prior art, the Hough transform algorithm is usually used to vote the high-temperature voxels in the three-dimensional temperature field to the parameter space, and detect the peak value in the parameter space to identify the potential straight line path. However, the traditional Hough transform has inherent defects when applied to this scenario: it uses equal weight voting for all high-temperature voxels. However, in actual industrial environments, the fault heat zone is not an ideal straight line, and is often slightly curved due to the inherent curvature of the coil, discontinuously distributed due to early faults, or has a certain width due to heat diffusion. When faced with these imperfect linear structures, the voting of the traditional Hough transform will diverge to multiple adjacent units in the parameter space, and cannot form a sharp and reliable peak value, resulting in high false alarm rate or missed detection of the algorithm for real faults, which is difficult to meet the monitoring needs of industrial sites. SUMMARY

[0005] To solve the technical problem that the traditional Hough transform uses equal voting for high-temperature voxels, which cannot adapt to the non-ideal shape of the induction furnace coil fault heat zone, and thus leads to high false alarm rate or missed detection, the present application provides an induction furnace coil temperature field visualization monitoring method based on thermal imaging, which comprises the following steps: Acquire three-dimensional temperature field data of an induction furnace coil, extract high-temperature voxels higher than a preset temperature threshold from the three-dimensional temperature field data to obtain a high-temperature voxel set; for any voxel in the high-temperature voxel set, calculate projection densities of the voxel in multiple preset candidate directions through the distribution of voxels in the neighborhood of the voxel, determine local linearity of the voxel based on the projection densities; screen voxels with local linearity greater than a set threshold from the high-temperature voxel set to obtain candidate voxels, and take a candidate direction with the maximum projection density of any candidate voxel as the local main direction of the candidate voxel; calculate collinearity scores of each candidate voxel, the collinearity score being used to represent consistency of the candidate voxel with each candidate voxel in the neighborhood of the candidate voxel in the local main direction; determine an adaptive weight of the candidate voxel according to the product of the local linearity and the collinearity score of the candidate voxel; perform weighted Hough transform on the candidate voxels, and vote using the adaptive weight; determine a linear high-temperature region according to the voting accumulation result, and perform visual monitoring.

[0006] The application realizes intelligent management of the voting process by constructing an adaptive weight for each high-temperature voxel. Firstly, the local linearity of each voxel is calculated to quantify the linear feature strength of the voxel itself. Secondly, candidate voxels are obtained by screening according to the set threshold, and the consistency of the candidate voxel with other candidate voxels in the neighborhood of the candidate voxel in the local main direction is evaluated by calculating the collinearity score of the candidate voxel. Finally, the adaptive weight is obtained by multiplying the local linearity and the collinearity score. This mechanism enables the candidate voxels with strong linear features and good linear continuation relationship with surrounding voxels to obtain much higher voting weight than isolated noise points or non-linear high-temperature region voxels in the weighted Hough transform. The dual effects of signal enhancement and noise suppression can significantly strengthen the convergence effect of the real linear high-temperature region in the Hough parameter space, and improve the detection accuracy of early, discontinuous or certain width linear high-temperature faults.

[0007] Preferably, the projection density satisfies the relationship: ; Wherein, is the projection density of the i th voxel in the j th candidate direction; is the coordinate vector of the i th voxel, is the coordinate vector of the j th voxel in the neighborhood of the i th voxel; is the unit vector of the j th candidate direction; is the coordinate vector of each voxel in the neighborhood of the i th voxel to the i th voxel; is the coordinate vector of each voxel in the neighborhood of the i th voxel to the i th voxel; is the coordinate vector of each voxel in the neighborhood of the i th voxel to the i th voxel; is the coordinate vector of each voxel in the neighborhood of the i th voxel to the i th voxel; is the coordinate vector of each voxel in the neighborhood of the i th voxel to the i th voxel; is the coordinate vector of each voxel in the neighborhood of the i th voxel to the i th voxel; is the coordinate vector of each voxel in the neighborhood of the i th voxel to the i th voxel; is the coordinate vector of each voxel in the neighborhood of the i th voxel to the i th voxel; a standard deviation of the shortest distances of the candidate directions; is the first total number of voxels within the neighborhood of the voxel; is a natural exponential function; is a module length symbol; is a first preset infinitesimal value; is a dot product operation.

[0008] The present application weights the perpendicular distance from each voxel within the neighborhood of the first voxel to the axis of the candidate direction: the closer the voxel to the first voxel, the greater the contribution, and the farther away, the contribution gradually decreases. This way can not only evaluate the tightness of the neighborhood voxels along a certain direction, but also be insensitive to the position of noise points, and can provide a more stable and reliable numerical basis for subsequent local linearity calculation.

[0009] Preferably, the local linearity of the voxel is determined based on the projection density, including: calculating the ratio of the mean value to the maximum value of the projection density of the voxel in a plurality of preset candidate directions; and recording the difference between 1 and the ratio as the local linearity of the voxel.

[0010] The present application uses the ratio of the maximum value to the mean value of the projection density to measure the local linearity of the voxel, which can more effectively reflect the prominence of a direction. When a voxel in a linear structure, its projection density in the main direction will be much larger than the average level of other directions, so as to get a local linearity score close to 1. This way can effectively distinguish the candidate voxels with clear directionality and the voxels in irregular high-temperature blocks, and provides a reliable basis for screening high-quality candidate voxels.

[0011] Preferably, the calculation of the collinearity score of each candidate voxel includes: recording any of the candidate voxels as a target voxel, and pairing the target voxel with all candidate voxels in its neighborhood one by one to obtain a plurality of voxel pairs; combining the angle relationship between the local principal directions of the two candidate voxels in any voxel pair, and the vector sum of the coordinate vectors of the local principal directions of the two candidate voxels, and the angle relationship between the spatial displacement direction of the two candidate voxels, to calculate the vector consensus score between the two candidate voxels in the voxel pair; calculating the vector consensus score between the two candidate voxels in all the voxel pairs one by one, and determining the collinearity score of the target voxel based on the weighted average of the vector consensus scores; wherein the weight of the weighted average is the local linearity of each candidate voxel in the neighborhood of the target voxel.

[0012] This invention uses the local linearity of other candidate voxels in the neighborhood as the weight of the weighted average. If a target voxel is aligned with a neighboring voxel that is already very linear, its collinearity score will be significantly improved. This mechanism allows voxels distributed along the real fault chain to corroborate and reinforce each other, thereby widening the weight gap between the real signal and isolated noise points, making the final adaptive weight more discriminative.

[0013] Preferably, the vector consensus score satisfies the following relation: ; in, It is the first The candidate voxel and its neighborhood... Vector consensus score among candidate voxels; It is the first Local principal directions of candidate voxels It is the first Within the neighborhood of the candidate voxel Local principal directions of candidate voxels; From the first The candidate voxel points to the first one in its neighborhood. A unit vector of candidate voxels; It is the modulus symbol; It is the second preset minute value; It is a dot product operation.

[0014] This invention multiplies the consistency of the local principal direction of two candidate voxels with the consistency of the average principal direction and the spatial displacement direction, two key geometric constraints. This dual-constraint verification mechanism can determine whether two voxels belong to the same continuous straight line, effectively avoiding misjudgment of two parallel but non-collinear line segments in space, and improving the accuracy of identifying continuous fault patterns.

[0015] Preferably, the acquisition of the plurality of preset candidate directions includes: generating uniformly distributed candidate directions by recursively subdividing a regular icosahedron based on a unit sphere.

[0016] Preferably, the plurality of preset candidate directions includes at least 64 directions.

[0017] Preferably, performing a weighted Hough transform on the candidate voxels includes: initializing a Hough parameter space accumulator, obtaining the voting surface corresponding to each candidate voxel in the Hough parameter space accumulator for each candidate voxel, and adding the adaptive weight corresponding to the candidate voxel to all cells on the voting surface.

[0018] Preferably, the neighborhood is a three-dimensional spherical neighborhood.

[0019] Preferably, the three-dimensional temperature field data of the induction furnace coil is acquired by using at least one infrared thermal imager to collect multi-angle two-dimensional temperature images of the surface of the induction furnace coil, and three-dimensional reconstruction is performed on the multi-angle two-dimensional temperature images of the surface to generate the three-dimensional temperature field data.

[0020] The present application has the following advantages: The present application realizes the detection of linear high-temperature faults of an induction furnace coil through multi-dimensional technical design. In the local linearity calculation link, the present application first weights the perpendicular distance of each voxel in the voxel neighborhood to the candidate direction axis, and the closer the voxel is to the axis, the greater the contribution, and the farther the distance, the contribution gradually decreases. This method can evaluate the tightness of the neighborhood voxels along a certain direction and is not sensitive to the position of noise points, providing a stable and reliable numerical basis for subsequent calculations. On this basis, the ratio of the maximum value to the average value of the projection density is used to measure the local linearity of the voxel, which can more effectively reflect the direction prominence. The voxels in the linear structure have a main direction projection density much greater than the average level of other directions, and the local linearity score is close to 1, thereby distinguishing candidate voxels with clear directionality from irregular high-temperature block voxels and providing a reliable basis for high-quality candidate voxel screening. In the collinearity score calculation link, the local linearity of other candidate voxels in the neighborhood is used as the weighted average weight. If the target voxel is consistent with the direction of the neighborhood voxels with strong linear features, the collinearity score will be significantly improved. This mechanism enables the voxels distributed along the real fault chain to confirm and enhance each other, greatly widening the weight gap between real signals and isolated noise points. Finally, the present application multiplies the local linearity and the collinearity score to construct an adaptive weight for each high-temperature voxel, realizing intelligent management of the weighted Hough transform voting process. This dual action of signal enhancement and noise suppression can strengthen the convergence effect of real linear high-temperature regions in the Hough parameter space, effectively improving the detection accuracy of early, discontinuous or certain width linear high-temperature faults. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 A flowchart of the induction furnace coil temperature field visualization monitoring method based on thermal imaging provided by the embodiment of the present application is shown in the figure. Figure 2 A schematic diagram of the perpendicular distance of the vector of the target voxel pointing to any voxel in the neighborhood to the axis of any candidate direction is shown in the figure. DETAILED DESCRIPTION

[0022] The embodiment of the present application provides an induction furnace coil temperature field visualization monitoring method based on thermal imaging, as shown in the figure. Figure 1 The method comprises steps S100-S400: Step S100, acquire the three-dimensional temperature field data of the induction furnace coil, extract high-temperature voxels higher than a preset temperature threshold from the three-dimensional temperature field data, and obtain a high-temperature voxel set.

[0023] It should be noted that the abnormal temperature rise caused by turn-to-turn short circuit and other faults during the operation of the induction furnace coil is a key representation for judging the health status thereof, and therefore, this step needs to preliminarily screen out high-temperature regions with potential fault indication significance from the collected temperature data, so as to provide a high-quality data set for subsequent analysis.

[0024] Specifically, first, the three-dimensional temperature field data of the induction furnace coil is acquired. As an optional embodiment, at least one infrared thermal imager is used to collect multi-angle two-dimensional temperature images of the surface of the induction furnace coil from different angles; the occluded regions covering the surface of the coil are collected from different angles to avoid missing local temperature information from a single perspective and ensure the integrity of three-dimensional reconstruction. Second, a three-dimensional reconstruction algorithm is used to perform three-dimensional reconstruction on the surface multi-angle two-dimensional temperature images, thereby generating discretized three-dimensional temperature field data. The three-dimensional reconstruction algorithm is a prior art and will not be described in detail here. The three-dimensional temperature field data can be represented as a three-dimensional voxel matrix, wherein each voxel records its spatial coordinates and the corresponding temperature value .

[0025] Then, a preset temperature threshold is set, the entire three-dimensional temperature field data is traversed, voxels with a temperature value higher than the preset temperature threshold are extracted, and a high-temperature voxel set is formed.

[0026] Regarding the setting of the preset temperature threshold, it can be set as the average temperature of the coil surface in the stable operation state of the induction furnace plus three times the standard deviation. This setting can effectively filter out temperature fluctuations and sensor noise during normal operation of the device, ensure that the extracted high-temperature voxels have high statistical significance, thereby reducing the computational load of subsequent processing and reducing false positives. The implementer can also set it according to the requirements.

[0027] At this point, the high-temperature voxel set is obtained.

[0028] Step S200, for any voxel in the high-temperature voxel set, the projection density of the voxel in multiple preset candidate directions is calculated through the distribution of voxels in the neighborhood of the voxel, and the local linearity of the voxel is determined based on the projection density.

[0029] It should be noted that this invention uses a geometric projection statistical method to analyze the geometric distribution of each voxel within its three-dimensional spherical neighborhood to determine the linearity of its local structure. This method projects voxels within the neighborhood into multiple preset directions and statistically analyzes the voxel aggregation density in each direction; the greater the density difference, the higher the linearity. The aim is to distinguish, at the microscopic level, whether a voxel tends to form part of a linear structure or is merely a disordered scattered or clustered noise point, which is a crucial prerequisite for subsequent identification of linear fault patterns.

[0030] Specifically, firstly, multiple preset candidate directions are generated, uniformly distributed on a unit sphere. As a preferred approach, this can be achieved by recursively subdividing all the triangular faces of an icosahedron. Since the icosahedron is the closest regular polyhedron to a sphere, recursively subdividing its triangular faces and normalizing its vertices allows the generated directions to be approximately uniformly distributed on the unit sphere. Each subdivision divides a triangle into four smaller triangles and normalizes its vertices to the unit sphere. This method can generate direction vectors that are uniformly distributed in space. In this embodiment, the number of preset candidate directions is preferably 64 or more, for example, 128. When the number of directions is less than 64, the direction resolution is too low, which may lead to inaccurate quantization of local linearity; while too many directions increase computational costs. Therefore, choosing 128 directions is a good balance between accuracy and efficiency.

[0031] Regarding the radius of the three-dimensional spherical neighborhood, it can be set according to the typical fault scale of the induction furnace coil. In this embodiment, it is preferably 3 to 5 times the voxel side length, and the implementer can set it according to the requirements.

[0032] Secondly, for any voxel in the high-temperature voxel set, calculate the projection density of the voxel along each of the aforementioned preset candidate directions within its three-dimensional spherical neighborhood. The projection density satisfies the following relationship: ; in, It is the first Individual elements in the first Projection density in each candidate direction; It is the first The coordinate vector of an individual element. It is the first Within the neighborhood of the individual element Individual element coordinate vector; It is the first Unit vectors of candidate directions; It is the first The coordinate vectors of each voxel in the neighborhood of a voxel to the first voxel The standard deviation of the shortest distance for each candidate direction; It is the first The total number of voxels in the neighborhood of a voxel; It is the natural exponential function; It is the modulus symbol; This is the first preset micro value, used to prevent the denominator from being 0. It can be set to 0.01, or can be set according to needs. It is a dot product operation.

[0033] In this relation, It is the first Individual elements point to the first element in their neighborhood. Vector of individual elements; It is the first Individual elements point to the first element in their neighborhood. The vector of individual elements, up to the th The square of the perpendicular distance of the axis of the candidate direction, when the th candidate direction is within the neighborhood The closer the individual is to the first The smaller this distance value, the stronger the exponential function becomes, representing the axis of each candidate direction. The closer the output value is to 1, the better the total projected density. The greater the contribution, the higher the projected density along that linear direction when the neighborhood voxels of a voxel are arranged linearly in space. A significant peak will appear.

[0034] like Figure 2 As shown in the figure, the first... Within the neighborhood of the individual element Individual elements to the first The perpendicular distances between the axes of the candidate directions are illustrated in a two-dimensional plane for visual clarity. In the figure, It is the origin of the coordinate system. It is the first The coordinates of the individual elements It is the first Within the neighborhood of the individual element The coordinates of the individual elements It is the first The unit vector of the candidate direction is also the th Axis of candidate directions, yes To the The distance between the axes of each candidate direction.

[0035] Then, calculate the first one one by one. The projected density of an individual element across all candidate directions yields a sequence of projected density values. ,in, It represents the total number of candidate directions.

[0036] It should be noted that if a voxel is located on a linear structure, there will be a peak value in the projection density sequence of the voxel, which is much larger than other values; on the contrary, if the voxel is located in a group or scattered area, the projection density of the voxel in each direction will be relatively close, and the peak value will not be prominent. Therefore, the linear feature can be measured by comparing the difference between the mean value and the maximum value of the projection density.

[0037] According to the above logic, the local linearity of the voxel satisfies the relationship: ; wherein, is the local linearity of the i-th voxel; is the total number of preset candidate directions; is the projection density of the i-th voxel in the j-th candidate direction; is the projection density of the i-th voxel in the j-th candidate direction; is the maximum value in the projection density sequence, which is also the maximum projection density of the i-th voxel in all candidate directions. In the relationship, the denominator is the maximum value in the projection density sequence, and the numerator is the mean value of the projection density sequence. If the maximum value is much higher than the mean value, the ratio tends to 0, and the local linearity tends to 1, indicating that the voxel has a strong linear feature; on the contrary, if the peak value is close to the mean value, the ratio tends to 1, and the local linearity tends to 0.

[0038] At this point, the local linearity of each voxel is obtained.

[0039]

[0040] Step S300, from the set of high-temperature voxels, voxels with a local linearity greater than a set threshold are screened to obtain candidate voxels, and the candidate direction with the maximum projection density of any candidate voxel is taken as its local main direction; the collinearity score of each candidate voxel is calculated.

[0041] It should be noted that the local linearity of a voxel can only determine whether a single voxel is in a local linear structure, but cannot guarantee that all voxels in a local linear structure belong to the same linear fault. Therefore, this step aims to introduce a collinearity score to evaluate the consistency of voxels in direction and spatial position, that is, whether they are collinear, so as to distinguish the macroscopically continuous linear structure from the discrete and irrelevant local linear fragments.

[0042] ​​​Specifically, first, from the high-temperature voxel set, the voxel with local linearity greater than a set threshold is screened to obtain a candidate voxel set; for each candidate voxel in the candidate voxel set, the candidate direction corresponding to the candidate voxel when the projection density reaches the maximum value is determined as the local main direction of the candidate voxel, and the local main direction is a unit vector.

[0043] As a preferred embodiment, the set threshold can be set to 0.6, which can effectively filter out most of the voxels in the cluster or random noise region, focus the analysis on voxels with higher linearity, and improve the efficiency and accuracy of subsequent calculations. The implementer can set it according to the needs.

[0044] Secondly, the collinearity score of each candidate voxel is calculated, which is used to represent the consistency of the candidate voxel and each candidate voxel in its neighborhood in the local main direction. To calculate the collinearity score, the vector consensus score also needs to be calculated. The shape of the neighborhood and its radius are consistent with step S200.

[0045] Specifically, the first candidate voxel to be calculated is denoted as the target voxel, the target voxel and other candidate voxels in its three-dimensional spherical neighborhood are paired one by one to obtain a plurality of voxel pairs, then the vector consensus score between the two candidate voxels in any voxel pair is calculated, and the vector consensus score satisfies the following relationship: ; is the vector consensus score between the first candidate voxel and the first candidate voxel in its neighborhood; is the local main direction of the first candidate voxel, is the local main direction of the first candidate voxel in the neighborhood of the first candidate voxel; is the unit vector from the first candidate voxel to the first candidate voxel in its neighborhood; is the modulus symbol; is a second predetermined small value, which is used to prevent the denominator of the second term from being 0, which can be set to 0.01, or can be set according to the needs; is the dot product operation.

[0046] In this relationship, is the dot product of the local main direction of the first candidate voxel and the local main direction of the first candidate voxel in its neighborhood, which is used to measure whether the main direction vectors of the two candidate voxels are consistent,​ the value range of the first part is controlled in interval. The second term in represents the vector sum of the coordinate vectors of the local principal directions of the two candidate voxels, i.e. their average principal direction, is the unit vector from the first candidate voxel to the first candidate voxel in its neighborhood, i.e. their spatial displacement direction, and the second term calculates the consistency of the average principal direction and the spatial displacement direction between the two voxels. Only when the local principal directions of the two voxels are not only parallel to each other, but also their connecting direction is consistent with the principal direction, a higher score will be obtained, which describes the geometric nature of collinearity.

[0047] Then, by weightedly averaging the vector consensus scores of the candidate voxels in the neighborhood of the target voxel, the collinearity score of the target candidate voxel is obtained, which satisfies the relationship: ; wherein, is the collinearity score of the first candidate voxel; is the local linearity of the first voxel in the neighborhood of the first candidate voxel; is the total number of candidate voxels in the neighborhood of the first candidate voxel; is the vector consensus score between the first candidate voxel and the first candidate voxel in its neighborhood; is a third preset infinitesimal value for preventing the denominator from being 0, which can be set as 0.01 or can be set according to requirements.

[0048] This weighting manner ensures that the neighbor voxels with more obvious linear features have greater influence when judging collinearity.

[0049] At this point, the collinearity score of each candidate voxel is obtained.

[0050] Step S400, determining the adaptive weight of each candidate voxel according to the product of its local linearity and collinearity score; performing weighted Hough transform on the candidate voxels, using the adaptive weight to vote; determining the linear high-temperature region according to the voting accumulation result and performing visual monitoring.

[0051] It should be noted that the traditional Hough transform treats all feature points equally, resulting in poor robustness to imperfect and noisy data. This invention integrates local linearity and collinearity scores into an adaptive weight, and uses this weight to guide the voting process of the Hough transform. This allows the algorithm to focus on voxels with high geometric consistency formed by real faults, ultimately achieving robust and accurate detection of imperfect linear structures.

[0052] Specifically, firstly, for any candidate voxel in the candidate voxel set, its adaptive weight is determined based on the product of its local linearity and collinearity score. This adaptive weight satisfies the following relationship: ; in, It is the first Adaptive weights for each candidate voxel It is the first Local linearity of candidate voxels; It is the first The collinearity score of each candidate voxel.

[0053] In this relation, Indicates the first Are the candidate voxels on a straight line? Indicates the first Candidate voxels must be on the same straight line as each candidate voxel in their neighborhood. Only candidate voxels that are both on the same straight line and on the same straight line as each candidate voxel in their neighborhood can be given a higher weight and are considered to be more credible.

[0054] Next, a weighted Hough transform is performed on the candidate voxels.

[0055] Specifically, a four-dimensional Hough parameter space accumulator is initialized to represent a three-dimensional straight line. A straight line in three-dimensional space can be described by four parameters: the direction vector and the perpendicular distance from a point on the line to the origin. Therefore, the Hough parameter space is four-dimensional. Then, each candidate voxel is traversed to obtain its corresponding voting surface in the Hough parameter space accumulator. Unlike the traditional Hough transform, which adds a fixed value to all cells on the surface, this invention adds an adaptive weight corresponding to the candidate voxel to all cells on the voting surface without modifying other operations of the Hough transform, which will not be elaborated here.

[0056] Finally, after all the candidate voxels complete the weighted voting, search the peak point with the maximum accumulated value in the Hough parameter space accumulator. Due to the use of adaptive weights, the peak value corresponding to the real fault mode will be significantly enhanced, while the pseudo-peak caused by environmental interference or random noise will be effectively suppressed. The parameters corresponding to the detected peak value are the best fitting straight line parameters of the identified linear high temperature region. Reverse project the straight line parameters back onto the three-dimensional temperature field model, highlight or render the linear high temperature region with special colors, and the intuitive and clear visualization monitoring of the suspected inter-turn short circuit fault position, direction and range can be realized.

[0057] The above are preferred embodiments of the present application, which do not limit the protection scope of the present application, and therefore: any equivalent changes made in the structure, shape, principle of the present application should be covered within the protection scope of the present application.

Claims

1. A method for visualizing monitoring of temperature field of induction furnace coil based on thermography, characterized in that, The method comprises the steps of: obtaining three-dimensional temperature field data of an induction furnace coil, extracting high-temperature voxels higher than a preset temperature threshold from the three-dimensional temperature field data to obtain a high-temperature voxel set; for any voxel in the high-temperature voxel set, calculating the projection density of the voxel in multiple preset candidate directions through the distribution of voxels in the neighborhood of the voxel, determining the local linearity of the voxel based on the projection density; from the high-temperature voxel set, screening voxels with local linearity greater than a set threshold to obtain candidate voxels, and taking the candidate direction with the maximum projection density of any candidate voxel as the local main direction of the candidate voxel; calculating the collinearity score of each candidate voxel, the collinearity score being used to represent the consistency of the candidate voxel and each candidate voxel in the neighborhood of the candidate voxel in the local main direction; determining the adaptive weight of the candidate voxel according to the product of the local linearity and the collinearity score of the candidate voxel; performing weighted Hough transform on the candidate voxel, using the adaptive weight to vote; determining a linear high-temperature region according to the voting accumulation result and performing visual monitoring.

2. The thermographic-based induction furnace coil temperature field visualisation monitoring method as claimed in claim 1, wherein, The projection density satisfies the relationship: ； wherein, is the projection density of the i-th voxel in the j-th candidate direction; is the coordinate vector of the i-th voxel, is the coordinate vector of the j-th voxel within the neighborhood of the i-th voxel, is the unit vector of the j-th candidate direction; is the standard deviation of the shortest distance from the coordinate vectors of the voxels within the neighborhood of the i-th voxel to the j-th candidate direction; is the total number of voxels within the neighborhood of the i-th voxel; is the natural exponential function; is the first preset infinitesimal value; is the dot product operation.​​​​​​​​​​ 3. The thermographic-based induction furnace coil temperature field visualization monitoring method of claim 1, wherein, The determination of the local linearity of the voxel based on the projection density comprises: calculating the ratio of the mean value to the maximum value of the projection density of the voxel in multiple preset candidate directions; taking the difference between 1 and the ratio as the local linearity of the voxel.

4. The thermographic-based induction furnace coil temperature field visualization monitoring method of claim 1, wherein, The calculation of the collinearity score of each candidate voxel comprises: taking any candidate voxel as a target voxel, pairing the target voxel with all candidate voxels in the neighborhood of the target voxel one by one to obtain a plurality of voxel pairs; combining the included angle relationship between the local main directions of the two candidate voxels in any voxel pair and the vector sum of the coordinate vectors of the local main directions of the two candidate voxels, and the included angle relationship between the spatial displacement direction of the two candidate voxels, to calculate the vector consensus score between the two candidate voxels in the voxel pair; calculating the vector consensus score between the two candidate voxels in all voxel pairs one by one, and determining the collinearity score of the target voxel based on the weighted average of the vector consensus scores; wherein the weight of the weighted average is the local linearity of each candidate voxel in the neighborhood of the target voxel.

5. The thermographic-based induction furnace coil temperature field visualisation monitoring method as claimed in claim 4, wherein, The vector consensus score satisfies the relationship: ； wherein, is a vector consensus score between the i-th candidate voxel and the j-th candidate voxel within its neighborhood; is a local principal direction of the i-th candidate voxel, is a local principal direction of the j-th candidate voxel within the neighborhood of the i-th candidate voxel; is a unit vector pointing from the i-th candidate voxel to the j-th candidate voxel within its neighborhood; is a modulus symbol; is a second preset infinitesimal value; is a dot product operation.​​​​​​​ 6. The thermographic-based induction furnace coil temperature field visualization monitoring method of claim 1, wherein, The acquisition of the multiple preset candidate directions comprises: based on a unit sphere, generating uniformly distributed candidate directions by recursively subdividing a regular icosahedron.

7. The thermographic-based induction furnace coil temperature field visualisation monitoring method as claimed in claim 6, wherein, The multiple preset candidate directions contain at least 64 directions.

8. The thermographic-based induction furnace coil temperature field visualization monitoring method of claim 1, wherein, The weighted Hough transform on the candidate voxels comprises: initializing a Hough parameter space accumulator, for each candidate voxel, obtaining its corresponding voting surface in the Hough parameter space accumulator, and increasing the adaptive weight corresponding to the candidate voxel for all cells on the voting surface.

9. The thermographic-based induction furnace coil temperature field visualization monitoring method of claim 1, wherein, The neighborhood is a three-dimensional spherical neighborhood.

10. The thermographic-based induction furnace coil temperature field visualization monitoring method of claim 1, wherein, The acquisition of the three-dimensional temperature field data of the induction furnace coil comprises: using at least one infrared thermal imager to collect surface multi-angle two-dimensional temperature images of the induction furnace coil; performing three-dimensional reconstruction on the surface multi-angle two-dimensional temperature images to generate the three-dimensional temperature field data.

Images