Online real-time detection and early warning method for gear grinding burn based on thermal imaging

Abstract

The invention provides a gear grinding burn on-line real-time detection and early warning method based on thermal imaging, which comprises the following steps of: continuously acquiring a temperature field of a grinding contact area by adopting an infrared thermal imager, performing non-uniformity correction and process parameter synchronization on a thermal image, generating a thermal image flow with a time scale, extracting temperature and gradient characteristics through space grid division, and performing on-line real-time detection and early warning on gear grinding burns. A space-time temperature matrix and a thermal evolution vector field are constructed, a dynamic graph model is utilized to describe a thermal diffusion relation, time sequence spectrum features extracted through short-time Fourier transform are combined, the space-time temperature matrix and the thermal evolution vector field are input into a graph convolution recursive network, regional temperature probability distribution is output, and burn risk thermodynamic diagram generation and multi-level early warning are achieved. And grinding parameters are dynamically adjusted based on the node space aggregation state. According to the method, the collaborative modeling mechanism of the space-time temperature matrix and the dynamic graph structure is constructed to carry out probabilistic prediction on the temperature state of each area in multiple time steps in the future, burn omen can be recognized in a high-precision and real-time mode, risks can be actively restrained, and the gear machining quality and process safety are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring and thermal imaging time-series analysis of grinding processes, and in particular to an online real-time detection and early warning method for gear grinding burns based on thermal imaging. Background Technology

[0002] Intelligent monitoring and early warning of gear burns during grinding has always been a key technological requirement in the high-end manufacturing and equipment industries. Infrared thermal imaging technology, as a mainstream non-contact, real-time method for acquiring the temperature field distribution in the grinding area, is widely used in online monitoring of gear processing. Existing gear grinding burn detection systems mostly employ single-frame thermal image static analysis technology. This involves thresholding or morphologically processing the temperature distribution of each frame of the thermal image to directly identify areas within the grinding zone where temperature anomalies occur. Mainstream methods generally include using the maximum surface temperature or temperature rise rate of the workpiece as a criterion, combined with traditional edge detection and region growing algorithms, to mark potential burn locations, and utilizing prior knowledge such as material critical temperatures and process parameter experience to achieve alarms and simple feedback.

[0003] Recent research has attempted to introduce deep learning methods into the field of burn detection, such as automatic thermal image segmentation and classification based on convolutional neural networks (CNNs), or automatic classification of abnormal patterns through feature clustering. These technologies have effectively improved spatial resolution and recognition accuracy. However, these methods generally remain within the scope of single-time-frame analysis. For complex grinding thermal evolution processes, they can only reflect the current thermal distribution state and are unable to capture the dynamic evolution path of temperature anomalies, as well as the physical evolution laws before burns occur. Publicly available technical information shows that existing thermal imaging burn detection systems mostly focus on static anomaly identification and result judgment, lacking the ability to model and predict dynamic trends in thermal evolution, resulting in delayed early warning responses and difficulty in achieving early intervention for burn risks.

[0004] Regarding the temporal evolution trend of burn risk during gear grinding, current technologies lack a systematic solution for dynamic modeling and evolution prediction of heat migration. Currently, no publicly available technology combines temporal spectral features, dynamic graph structures, and complex spatiotemporal models to jointly extrapolate burn trends. Most solutions rely on static features of single-frame images or empirical thresholds for judgment, while some employ hotspot tracking or serial data analysis. However, the following major technical bottlenecks remain: Single-frame thermal image analysis cannot effectively reflect the continuous temporal evolution of thermal anomalies and lacks the ability to extract precursor signals and extrapolate risk trends in the burn process.

[0005] Existing methods have failed to establish a spatial dynamic propagation model of heat migration, and cannot quantify the expansion behavior of burn risk in multiple regions, making it difficult to predict local anomalies as large-area burns in a timely manner.

[0006] Static criteria-based methods are not robust enough to environmental disturbances, process changes, and material fluctuations. Early warnings often lag behind actual burn occurrences, making it difficult to achieve intelligent closed-loop control of the production process.

[0007] The lack of a time-series spectrum feature extraction mechanism that combines frequency domain analysis and has consistency with physical processes makes it difficult to capture the low-frequency heating-dominant mode of burn formation, thus limiting the ability to identify and predict trends in the early stages.

[0008] Therefore, there is an urgent need for a method that integrates dynamic temporal features of thermal imaging, intelligent modeling of thermal migration processes, and multi-level early warning and response mechanisms. Summary of the Invention

[0009] This invention provides an online real-time detection and early warning method for gear grinding burns based on thermal imaging to solve the technical problems existing in the background art.

[0010] The technical solution of this invention is: an online real-time detection and early warning method for gear grinding burns based on thermal imaging, comprising the following steps: S1: Acquire thermal image sequences and process parameters of the grinding area during gear grinding to obtain a thermal image stream; S2: Preprocess the thermal image stream to construct the spatiotemporal temperature matrix; S3: Based on the spatiotemporal temperature matrix, calculate the temperature change rate and gradient direction consistency of each sub-region between adjacent time steps to generate a thermal evolution vector field per unit time. S4: Using each sub-region as a graph node, establish an initial topological connection based on their spatial adjacency relationship, and dynamically adjust the edge weights according to the temperature change synergy between adjacent nodes in the thermal evolution vector field to construct a dynamic graph structure that evolves over time, in order to simulate the propagation path of thermal anomalies. S5: Perform a short-time Fourier transform on the historical temperature sequence of each graph node to extract its energy distribution characteristics in different frequency bands, identify the heating mode that is dominated by low frequency bands and continuously enhanced, and generate a time-series spectral feature vector. S6: Predict the temperature state probability distribution of each node in multiple future time steps by combining the temporal spectrum feature vector with the dynamic graph structure input graph convolutional recursive network. S7: Determine whether the predicted temperature of each node exceeds the critical value of material phase transformation and whether the confidence level of two consecutive frames is greater than 85%. If the condition is met, generate a first-level warning signal. S8: If three or more adjacent nodes are detected to enter a high-risk state at the same time, the first-level warning will be upgraded to a second-level warning, and the linkage control mechanism will be triggered to adjust the grinding wheel speed or feed rate.

[0011] Furthermore, in step S2, the preprocessing is as follows: each frame in the thermal image stream is spatially gridded to divide the grinding contact area into multiple regular sub-regions, and the average temperature and temperature gradient features of each sub-region are extracted to construct a spatiotemporal temperature matrix.

[0012] Furthermore, in step S6, the graph convolutional recurrent network employs multiple graph convolutional layers, combined with gated recurrent units, and outputs the probability distribution of the temperature state for multiple future time steps for each node using Softmax.

[0013] Furthermore, in step S6, in the graph convolutional layer, a multi-layer graph convolution operation is performed based on the weighted adjacency matrix and node spectral feature vectors of the dynamic graph structure. The spatial aggregation mechanism of the graph convolutional recurrent network is used to fuse the thermal state information of adjacent nodes and generate a high-dimensional node embedding representation containing the local heat conduction context in order to capture the non-uniform heat diffusion behavior in the grinding contact area. The node embedding sequence output by the graph convolutional layer is input into the recursive unit in chronological order. The long-term dependency of the node state is modeled through the hidden state propagation mechanism. The global evolution law of thermal anomalies from local initiation to regional expansion is learned, and the hidden state sequence of future multiple time steps is output. Based on the future hidden states output by the recursive units, the fully connected decoding layer maps them to the probability distribution of the temperature state of each node at future times, generating prediction results with confidence estimates.

[0014] Furthermore, in step S7, a burn risk heat map is generated based on the temperature state probability distribution.

[0015] The beneficial technical effects of this invention are as follows: 1. This invention effectively overcomes the problems of lag and high false alarm rate caused by relying solely on instantaneous temperature thresholds in traditional gear grinding burn detection by constructing a collaborative modeling mechanism of spatiotemporal temperature matrix and dynamic graph structure. This invention divides the grinding area into spatial sub-grids and generates a thermal evolution vector field based on the temperature change rate and gradient direction consistency between adjacent time steps, dynamically characterizing the thermal conduction correlation between local areas, and then establishing a weighted graph topology structure that evolves over time. This can accurately identify the potential diffusion path of non-steady-state heat accumulation behavior, significantly improving the spatial positioning capability and evolution trend perception accuracy of early burn signs. 2. This invention introduces a joint learning framework of temporal spectrum feature extraction and graph convolutional recurrent network GCRN, which greatly enhances the foresight and robustness of the early warning system; 3. This invention extracts low-frequency dominant energy patterns as precursor features of burns by performing short-time Fourier transform on the historical temperature sequences of each node, effectively filtering out transient noise interference and highlighting the physical nature of the continuous heating process. After fusing the spectral features with the dynamic graph structure, the model is input into the GCRN model. The graph convolutional layer is used to capture the spatial dependence of local heat conduction, and the long-term evolution law is modeled by recursive units, thereby realizing the probabilistic prediction of the temperature state of each region in multiple future time steps. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the method of an embodiment of the present invention; Figure 2 This is a schematic diagram of the process for establishing a thermal evolution vector field in an embodiment of the present invention; Figure 3 This is a schematic diagram of the process for constructing a dynamic graph structure in an embodiment of the present invention. Detailed Implementation

[0017] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0018] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0019] like Figure 1 As shown in the figure, this embodiment provides an online real-time detection and early warning method for gear grinding burns based on thermal imaging, which specifically includes the following steps: S1: During gear grinding, the thermal image sequence of the grinding area is continuously acquired by infrared thermal imaging equipment, and the timestamp of each frame and the corresponding process parameters are recorded simultaneously to form a time-stamped thermal image stream. S2: Spatial gridding is performed on each frame in the thermal image stream to divide the grinding contact area into multiple regular sub-regions, and the average temperature and temperature gradient features of each sub-region are extracted to construct a spatiotemporal temperature matrix as the basic data structure for subsequent time series analysis. S3: Based on the spatiotemporal temperature matrix, calculate the temperature change rate and gradient direction consistency of each sub-region between adjacent time steps, and generate a thermal evolution vector field per unit time to characterize the dynamic migration trend of heat in the grinding region. S4: Using each sub-region as a graph node, establish an initial topological connection based on their spatial adjacency relationship, and dynamically adjust the edge weights according to the temperature change synergy between adjacent nodes in the thermal evolution vector field to construct a dynamic graph structure that evolves over time, in order to simulate the propagation path of thermal anomalies. S5: Perform a short-time Fourier transform on the historical temperature sequence of each graph node to extract its energy distribution characteristics in different frequency bands, identify the heating mode that is dominated by low frequency bands and continuously enhanced, generate a time-series spectral feature vector and use it as an indicator of burn precursors;

[0020] S6: Input the time-series spectrum feature vector and the dynamic graph structure into the graph convolutional recursive network. The graph convolutional layer uses the dynamic graph structure to capture local heat conduction behavior, while the recursive unit learns the global evolution law based on the historical state and jointly outputs the temperature state probability distribution of each node in multiple future time steps. S7: Generate a burn risk heat map based on the temperature state probability distribution, and determine whether the predicted temperature of each node exceeds the critical value of material phase transformation and whether the confidence level of two consecutive frames is greater than 85%. If the condition is met, generate a first-level warning signal. S8: If three or more adjacent nodes are detected to enter a high-risk state at the same time, the first-level warning will be upgraded to a second-level warning, and the linkage control mechanism will be triggered to adjust the grinding wheel speed or feed rate, so as to realize dynamic intervention of process parameters and suppression of burn risk.

[0021] In this embodiment, in step S1, during the gear grinding process, an infrared thermal imaging device continuously acquires a sequence of thermal images of the grinding area, and synchronously records the timestamp of each frame and the corresponding process parameters to form a time-stamped thermal image stream, including the following steps: S1.1: Based on an industrial-grade infrared thermal imager, the original thermal radiation signal of the gear grinding contact area is obtained. Its non-contact temperature measurement characteristics are used to scan the temperature field of the grinding interface in the whole area, generating a single-frame thermal image with spatial resolution as the basic unit for temperature information acquisition. S1.2: Perform non-uniformity correction and environmental radiation compensation processing on the single-frame thermal image, and use the blackbody reference source calibration curve to linearly map the pixel-level temperature values ​​to eliminate measurement deviations caused by lens obstruction, background thermal interference and sensor drift, and output the radiation-calibrated temperature distribution matrix. S1.3: Based on the PLC system, the process parameters corresponding to the current frame are synchronously collected, including the grinding wheel speed, workpiece feed speed and coolant flow rate, and the parameters are time-aligned with the completed radiometrically calibrated thermal image to generate an information association group containing multi-dimensional process context, ensuring the consistency between the thermal image stream and the processing status. S1.4: Add a precise timestamp to each frame of processed thermal image, use a high-precision clock module to record the acquisition time, construct a time-ordered thermal image sequence, and encapsulate all information into a unified data structure to form a time-stamped thermal image stream to support subsequent time series analysis; S1.5: The time-stamped thermal image stream is transmitted to the edge computing node for cache management. The most recent N frames of data are organized using a sliding window mechanism to generate dynamic data batches that can be used for real-time processing, providing a stable and low-latency data input source for subsequent gridding and spatiotemporal matrix construction.

[0022] In this embodiment, step S2 involves spatially meshing each frame in the thermal image stream, dividing the grinding contact area into multiple regular sub-regions, and extracting the mean temperature and temperature gradient features of each sub-region to construct a spatiotemporal temperature matrix as the basic data structure for subsequent time-series analysis. This includes the following steps: S2.1: Acquire a frame of infrared thermal image from the time-stamped thermal image stream, and determine the region of interest (ROI) of the grinding contact area in the image based on the spatial mapping relationship between the image pixel coordinate system and the grinding area, so as to limit the spatial range of subsequent mesh division; S2.2: Perform regular spatial grid division on the region of interest, uniformly divide it into M×N rectangular sub-regions, and generate a set of spatial grid cells with a fixed topology. Each grid cell corresponds to a spatial location identifier, which is used to establish the mapping relationship between subsequent graph nodes and physical regions. Furthermore, an image-based two-level index mapping method is used to assign a unique spatial location identifier to each rectangular cell, where the identifier is determined by... Combinatorial encoding ensures that node positions can be mapped one-to-one in subsequent spatiotemporal matrix and graph structure mapping.

[0023] Furthermore, by establishing a two-dimensional array structure to store the pixel coordinate range corresponding to the grid cell, the spatial correspondence between the cell and the physical grinding contact area is realized, and the center pixel coordinates are added to each cell as the basic field for constructing the adjacency matrix by inputting the geometric position of the graph node into S4.

[0024] Furthermore, by using a grid cell numbering matrix and a physical coordinate mapping table, each numbered cell is mapped to the actual physical region of the gear grinding contact area, thus realizing the transformation from thermal image spatial decomposition to physical domain discretization. Through regularized spatial grid partitioning, the ROI determined in the previous step is transformed into a set of spatial grid cells with a fixed topology, achieving a standardized mapping of the physical region in the graph structure.

[0025] S2.3: For the temperature value of each pixel in each grid cell, the arithmetic mean method is used to calculate its average temperature, and a scalar sequence of cell temperature representing the local thermal state is generated as the basic cell characteristic parameter reflecting the heat distribution in the grinding area. S2.4: Based on the temperature difference operation between adjacent pixels, perform gradient operator convolution operation on each grid cell to calculate its horizontal and vertical temperature gradient components, and then synthesize the gradient magnitude and direction angle to generate a cell temperature gradient feature vector describing the heat migration trend. S2.5: Arrange the mean temperature and temperature gradient features corresponding to each grid cell in time series to construct a three-dimensional tensor spatiotemporal temperature matrix, where the dimensions are spatial location and time step, respectively. The feature channels (mean, gradient magnitude, gradient direction) serve as the basic input data structure for subsequent thermal evolution vector field generation and dynamic graph modeling.

[0026] In this embodiment, in step S3, based on the spatiotemporal temperature matrix, the temperature change rate and gradient direction consistency of each sub-region between adjacent time steps are calculated to generate a thermal evolution vector field per unit time, which is used to characterize the dynamic migration trend of heat within the grinding region, such as... Figure 2 As shown, it includes the following steps: S3.1: Obtain the spatiotemporal temperature matrix output in S2, where each row represents a spatial sub-region and each column represents a time step. The elements of the spatiotemporal temperature matrix are the average temperature of the corresponding sub-region at the corresponding time. Based on the spatiotemporal temperature matrix, perform the difference operation between adjacent time steps on each spatial sub-region to calculate its temperature change rate within a unit time interval, and generate a temperature change rate sequence to quantify the local temperature rise rate. The time series differencing algorithm is used to perform temperature value difference operations between adjacent columns in each row of the temperature sequence to obtain the temperature change. The sequence reflects the temperature dynamics of the sub-region over time.

[0027] S3.2: Based on the temperature gradient feature sequence of each sub-region extracted in S2, calculate the cosine similarity of the gradient direction between adjacent time steps to obtain the gradient direction consistency coefficient; this coefficient reflects the continuity of the spatial distribution pattern of temperature during the propagation of the thermal front and serves as the basis for judging whether heat migration has directional diffusion characteristics. Furthermore, the gradient direction cosine similarity of each pair of spatial sub-regions between adjacent time steps is calculated using the following formula. :

[0028] in, For time steps The corresponding gradient direction unit vector, For time steps The corresponding gradient direction unit vector, This represents the vector dot product operation. This represents the magnitude of the vector.

[0029] S3.3: The temperature change rate obtained in S3.1 and the gradient direction consistency coefficient obtained in S3.2 are weighted and fused to construct a composite thermal evolution intensity index; the weight coefficient is dynamically adjusted according to the grinding process state, and the contribution ratio of enhancing gradient consistency under high feed rate conditions is used to suppress pseudo migration signals under noise interference.

[0030] Furthermore, the composite thermal evolution intensity index is calculated using a weighted linear superposition formula:

[0031] in, As a composite thermal evolution intensity index, and , where are weighting coefficients, representing the contribution ratios of temperature rise rate and directional consistency to the strength evaluation, respectively; , These are the normalized rate of temperature change and the gradient direction consistency coefficient, respectively.

[0032] S3.4: Taking each sub-region as the center, and combining the results of the polarity of its temperature change rate and the consistency of the gradient direction, determine the direction and amplitude of the local heat flow vector; use the vector interpolation algorithm to generate a continuous thermal evolution vector field on the spatial grid, which represents the instantaneous migration path and accumulation trend of heat in the grinding contact area; S3.5: Perform spatiotemporal smoothing on the generated thermal evolution vector field, and use a sliding time window to perform a weighted average operation on the vector field of multiple consecutive frames to suppress vector jitter caused by measurement noise or transient disturbances; output a stable thermal evolution vector field sequence as the input basis for constructing dynamic graph edge weights in S4.

[0033] In this embodiment, in step S4, each sub-region is used as a graph node, and an initial topological connection is established based on their spatial adjacency. The edge weights are then dynamically adjusted according to the temperature change synergy between adjacent nodes in the thermal evolution vector field to construct a dynamic graph structure that evolves over time, used to simulate the propagation path of thermal anomalies.Figure 3 As shown, it includes the following steps: S4.1: Based on the regular sub-regions of the grinding contact area divided in S2, the center coordinates of each sub-region and its spatial geometric adjacency relationship are used as input to construct an initial undirected graph topology, where the sub-regions correspond to graph nodes, and unweighted edges are established between adjacent sub-regions to generate an initial adjacency matrix, so as to provide a spatial connection basis for the dynamic graph structure. S4.2: Obtain the thermal evolution vector field data generated in S3, extract the temperature change rate and gradient direction consistency value of each adjacent sub-region pair under adjacent time steps as a measure of temperature change synergy, calculate the synergy coefficient between each pair of adjacent nodes, and generate an initial edge weight set to quantify the degree of synchronization of heat migration. S4.3: Based on the aforementioned synergy coefficient, a normalized exponential weighting function is used to perform a nonlinear mapping on the initial edge weights to generate dynamic edge weights. High synergy corresponds to high connection strength, while low synergy inhibits edge propagation ability, thereby optimizing the adjacency matrix and forming a weighted graph structure that reflects the consistency of thermal response. S4.4: Align the weighted adjacency matrix with the graph node sequence according to time steps, and periodically recalculate the edge weights in conjunction with the real-time updated thermal evolution vector field to realize the temporal iterative update of the adjacency matrix and construct a time-dependent dynamic graph structure to characterize the time-varying characteristics of the thermal anomaly propagation path. S4.5: Output the dynamic graph structure corresponding to each time step, including the updated weighted adjacency matrix and node feature set, as the topological input for graph convolution operations in the graph convolutional recurrent network, supporting the spatial aggregation of local heat conduction patterns and the learning of evolutionary laws across time steps.

[0034] In this embodiment, step S5 involves performing a short-time Fourier transform on the historical temperature sequence of each graph node to extract its energy distribution characteristics in different frequency bands, identifying a heating mode dominated by low frequencies and continuously increasing thereafter, generating a time-series spectral feature vector and using it as an indicator of burn precursors, including the following steps: S5.1: Obtain the historical temperature time series of the sub-regions corresponding to each graph node, wherein the historical temperature time series is a one-dimensional time series signal composed of the average temperature of the same spatial location in the spatiotemporal temperature matrix at continuous time steps, and is used as input data for time series spectrum analysis. S5.2: Based on the historical temperature time series, a sliding time window mechanism is used to perform segmentation processing to generate multiple overlapping time segment sequences. Each time segment has a fixed length and covers no less than three sampling periods to ensure the stability of frequency domain analysis and the identifiability of local time-varying characteristics. S5.3: Apply the short-time Fourier transform algorithm to each time segment sequence to calculate its complex spectrum output at the preset frequency resolution, and further obtain the power spectral density value at each frequency point to form a preliminary time-frequency energy distribution matrix, reflecting the degree of energy concentration of temperature changes in different frequency bands; S5.4: Based on the time-frequency energy distribution matrix, the low-frequency dominant energy ratio index is obtained by integration in the low-frequency range (0.1–2 Hz), and dynamic threshold comparison is performed by combining the trend change rate between adjacent time segments to identify the low-frequency energy growth pattern with continuous enhancement characteristics as a candidate feature for potential burn precursors. The power spectral density values ​​of each time segment are numerically integrated within the low-frequency range using a frequency band energy integration algorithm to obtain the total energy value of the low-frequency band for the corresponding time segment.

[0035] S5.5: The frequency band energy characteristics, change slope and start time information corresponding to the identified low-frequency continuous enhancement mode are vectorized and encoded to generate a time-series spectrum feature vector of uniform dimension. This vector is then output to the next processing stage for coupling modeling with the dynamic graph structure to achieve a physical consistency representation of the burn evolution path.

[0036] In this embodiment, in step S6, the time-series spectral feature vector and the dynamic graph structure are jointly input into the graph convolutional recurrent network. The graph convolutional layer uses the dynamic graph structure to capture local heat conduction behavior, while the recurrent unit learns the global evolution law based on historical states and jointly outputs the probability distribution of the temperature state of each node in multiple future time steps, including the following steps: S6.1: Based on the dynamic graph structure generated in the previous steps, each sub-region serves as a graph node, and the edge weights are dynamically determined by the synergy of temperature changes between adjacent nodes. This constructs a spatiotemporally adaptive graph topology input to characterize the potential paths of heat propagation within the grinding region and provides structured adjacency relationships for graph convolution operations. S6.2: Obtain the time-series spectral feature vector extracted in S5, that is, by performing short-time Fourier transform on the historical temperature sequence of each graph node, the energy distribution features of different frequency bands are obtained, with a focus on retaining the features of the continuously enhanced heating mode in the low frequency band, which are then used as node attributes to be input into the graph convolutional recurrent network to identify the frequency domain evolution characteristics of burn precursors. S6.3: In the graph convolutional layer, multi-layer graph convolution operation is performed based on the weighted adjacency matrix and node spectral feature vectors of the dynamic graph structure. The spatial aggregation mechanism of the graph neural network is used to fuse the thermal state information of adjacent nodes and generate a high-dimensional node embedding representation containing the local heat conduction context to capture the non-uniform heat diffusion behavior in the grinding contact area. S6.4: Input the node embedding sequence output by the graph convolutional layer into the recursive unit in chronological order. Model the long-term dependency of node states through the hidden state propagation mechanism, learn the global evolution law of thermal anomalies from local emergence to regional expansion, and output the hidden state sequence for multiple future time steps. S6.5: Based on the future hidden state output by the recursive unit, it is mapped to the probability distribution of the temperature state of each node at future time through a fully connected decoding layer, generating a prediction result with confidence estimation, which serves as the quantitative basis for subsequent burn risk heat map generation and multi-level early warning judgment.

[0037] Furthermore, the cross-entropy calculation method is used to measure the error between the Softmax-normalized probability distribution and the known training labels, as shown in the following formula:

[0038] in, For nodes The actual category label, For this node in the future The predicted probability of a time step; Cross-entropy loss; Furthermore, the weights of the fully connected layer and recursive units are jointly updated through the backpropagation algorithm, and the optimal decoder parameter set that can be directly used in the testing phase is generated.

[0039] In this embodiment, in step S7, a burn risk heat map is generated based on the temperature state probability distribution, and it is determined whether the predicted temperature at each node exceeds the material phase transition critical value and whether the confidence level is greater than 85% for two consecutive frames. If the condition is met, a first-level warning signal is generated, including the following steps: S7.1: Obtain the probability distribution data of temperature status of each graph node in the future multiple time steps output by the graph convolutional recurrent network, calculate the expected temperature value of each node in the key time window based on the probability distribution sequence, and generate a spatiotemporally aligned preliminary assessment matrix of burn risk as the basic data structure of the burn risk heat map. S7.2: Compare the preliminary burn risk assessment matrix with the phase transformation critical temperature threshold of the gear material element by element, mark all graph nodes whose expected temperature values ​​exceed the phase transformation critical value, and generate a high temperature risk identification map to characterize the spatial distribution of potential burn areas. S7.3: Perform confidence analysis on the probability distribution of temperature status of each marked node in the high temperature risk identification map for two consecutive historical frames. Use the probability vector output by softmax to extract the confidence level of its largest category and determine whether the node has reached or exceeded 85% high confidence in the prediction of two consecutive time steps, so as to filter out accidental misjudgments. S7.4: Based on the high temperature risk identification map and the high confidence judgment result, execute logic and operations to generate the final burn risk heat map, in which only nodes that simultaneously meet the conditions of 'predicted temperature exceeding the limit' and 'two-frame high confidence' are retained, and their values ​​are mapped to standardized risk level intensity, realizing the transformation from probability output to interpretable risk visualization; S7.5: Scan all active risk nodes in the final burn risk heat map. If at least one node meets the dual judgment conditions, generate a first-level warning signal and encapsulate the signal into a standardized communication message containing the trigger time, location coordinates, predicted temperature and confidence information for the host computer system to read and respond to.

[0040] In this embodiment, in step S8, if three or more adjacent nodes are detected to simultaneously enter a high-risk state, the first-level warning is upgraded to a second-level warning, and a linkage control mechanism is triggered to adjust the grinding wheel speed or feed rate, thereby achieving dynamic intervention of process parameters and suppression of burn risk. Specifically, the steps include the following: S8.1: Obtain the probability distribution of temperature status of each node in multiple future time steps output by the graph convolutional recurrent network, and generate a burn risk heat map based on the probability distribution. Each sub-region is mapped with a color gradient to the probability value of its predicted temperature exceeding the material phase transition critical value, forming a spatially interpretable risk visualization data structure, which serves as the basic input for subsequent multi-level early warning judgment. S8.2: Based on the burn risk heat map, perform spatial connectivity analysis, use the four-neighbor or eight-neighbor connectivity criterion to cluster and segment high-risk nodes with a probability value greater than 85%, extract the geometric center, area size and boundary contour of each connected region, and generate a set of high-risk region clusters to quantify the spatial aggregation degree of risk. S8.3: Calculate the number of consecutive high-risk nodes contained in each high-risk area cluster, and determine whether there is a connected region consisting of at least three adjacent nodes; if so, confirm that the cluster meets the triggering conditions for a level 2 warning, and generate a regional risk level identifier as the basis for the judgment of the warning escalation logic. S8.4: Based on the regional risk level identifier that meets the conditions for a level 2 early warning, the linkage control protocol is activated, a level 2 early warning signal is sent to the production line PLC control system, and the preset process adjustment strategy table is called through the industrial bus interface to match the optimal combination of grinding wheel speed and feed rate adjustment parameters under the current working conditions. S8.5: The PID control mechanism is implemented to achieve closed-loop control of the grinding wheel drive motor and feed servo system. Based on the matching combination of process adjustment parameters, the grinding wheel linear speed or feed rate is dynamically reduced, thereby reducing the grinding heat input per unit time, realizing the active suppression of the burn evolution trend, and feeding back the adjusted process parameters to the monitoring platform to complete the control closed loop.

[0041] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0042] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0043] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for online real-time detection and early warning of gear grinding burns based on thermal imaging, characterized in that, Includes the following steps: S1: Acquire thermal image sequences and process parameters of the grinding area during gear grinding to obtain a thermal image stream; S2: Preprocess the thermal image stream to construct the spatiotemporal temperature matrix; S3: Based on the spatiotemporal temperature matrix, calculate the temperature change rate and gradient direction consistency of each sub-region between adjacent time steps to generate a thermal evolution vector field per unit time. S4: Using each sub-region as a graph node, establish an initial topological connection based on their spatial adjacency relationship, and dynamically adjust the edge weights according to the temperature change synergy between adjacent nodes in the thermal evolution vector field to construct a dynamic graph structure that evolves over time, in order to simulate the propagation path of thermal anomalies. S5: Perform a short-time Fourier transform on the historical temperature sequence of each graph node to extract its energy distribution characteristics in different frequency bands, identify the heating mode that is dominated by low frequency bands and continuously enhanced, and generate a time-series spectral feature vector. S6: Predict the temperature state probability distribution of each node in multiple future time steps by combining the temporal spectrum feature vector with the dynamic graph structure input graph convolutional recursive network. S7: Determine whether the predicted temperature of each node exceeds the critical value of material phase transformation and whether the confidence level of two consecutive frames is greater than 85%. If the condition is met, generate a first-level warning signal. S8: If three or more adjacent nodes are detected to enter a high-risk state at the same time, the first-level warning will be upgraded to a second-level warning, and the linkage control mechanism will be triggered to adjust the grinding wheel speed or feed rate.

2. The online real-time detection and early warning method for gear grinding burns based on thermal imaging according to claim 1, characterized in that, In step S1, the original thermal radiation signal of the gear grinding contact area is obtained based on an industrial-grade infrared thermal imager. Its non-contact temperature measurement characteristics are used to scan the temperature field of the grinding interface in the whole area, generating a single-frame thermal image with spatial resolution. Non-uniformity correction and environmental radiation compensation are performed on a single frame thermal image. The pixel-level temperature values ​​are linearly mapped using a blackbody reference source calibration curve to obtain a radiation-calibrated temperature distribution matrix.

3. The online real-time detection and early warning method for gear grinding burns based on thermal imaging according to claim 1, characterized in that, In step S2, the preprocessing is as follows: each frame in the thermal image stream is spatially gridded, the grinding contact area is divided into multiple regular sub-regions, and the average temperature and temperature gradient features of each sub-region are extracted to construct a spatiotemporal temperature matrix.

4. The online real-time detection and early warning method for gear grinding burns based on thermal imaging according to claim 3, characterized in that, In step S2, for the temperature value of each pixel in each grid cell, the arithmetic mean method is used to calculate its average temperature, and a scalar sequence of cell temperature representing the local thermal state is generated. Based on the temperature difference calculation between adjacent pixels, a gradient operator convolution operation is performed inside each grid cell to calculate its horizontal and vertical temperature gradient components, synthesize the gradient magnitude and direction angle, and generate a cell temperature gradient feature vector describing the heat migration trend. The mean temperature and temperature gradient features corresponding to each grid cell are arranged in time series to construct a spatiotemporal temperature matrix.

5. The online real-time detection and early warning method for gear grinding burns based on thermal imaging according to claim 4, characterized in that, In step S3, each row of the spatiotemporal temperature matrix is ​​taken as a spatial sub-region; the difference operation between adjacent time steps is performed on each spatial sub-region to calculate its temperature change rate within a unit time interval, and a temperature change rate sequence is generated. Based on the temperature gradient feature sequence of each spatial sub-region, the cosine similarity of the gradient direction between adjacent time steps is calculated to obtain the gradient direction consistency coefficient. A composite thermal evolution intensity index is constructed by weighting and fusing the temperature change rate with the gradient direction consistency coefficient. Centered on each spatial sub-region, and combining the polarity of its temperature change rate and the consistency of its gradient direction, the direction and magnitude of the local heat flow vector are determined; a continuous thermal evolution vector field is generated on the spatial grid using a vector interpolation algorithm.

6. The online real-time detection and early warning method for gear grinding burns based on thermal imaging according to claim 1, characterized in that, In step S4, the center coordinates of each regular sub-region and its spatial geometric adjacency relationship are used as input to construct an initial undirected graph topology, where the sub-regions correspond to graph nodes, unweighted edges are established between adjacent sub-regions, and an initial adjacency matrix is ​​generated. The temperature change rate and gradient direction consistency value of adjacent spatial sub-regions are used to calculate the coordination coefficient between each pair of adjacent nodes, and an initial edge weight set is generated. A normalized exponential weighting function is used to perform a nonlinear mapping on the initial edge weights to generate dynamic edge weights; the adjacency matrix is ​​optimized through the dynamic edge weights to form a weighted graph structure that reflects the consistency of thermal response; By aligning the weighted adjacency matrix with the graph node sequence according to time steps and periodically recalculating edge weights using a real-time updated thermal evolution vector field, the adjacency matrix is ​​updated iteratively over time, thus constructing a time-dependent dynamic graph structure.

7. The online real-time detection and early warning method for gear grinding burns based on thermal imaging according to claim 6, characterized in that, In step S5, a sliding time window mechanism is used to segment the historical temperature time series into multiple overlapping time segment sequences; The complex spectrum output of each time segment sequence at a preset frequency resolution is calculated using the short-time Fourier transform algorithm, and the power spectral density value at each frequency point is obtained to form a time-frequency energy distribution matrix; the low-frequency dominant energy proportion index is obtained by integrating within 0.1-2Hz of the time-frequency energy distribution matrix; and the low-frequency energy growth mode is identified. The frequency band energy characteristics, change slope, and start time information corresponding to the low-frequency continuous enhancement mode are vectorized and encoded to generate a time-series spectrum feature vector.

8. The online real-time detection and early warning method for gear grinding burns based on thermal imaging according to claim 7, characterized in that, In step S6, the graph convolutional recurrent network uses multiple graph convolutional layers, combined with gated recurrent units, and outputs the probability distribution of temperature state for multiple future time steps for each node through Softmax.

9. The online real-time detection and early warning method for gear grinding burns based on thermal imaging according to claim 8, characterized in that, In step S6, in the graph convolutional layer, multi-layer graph convolution operation is performed based on the weighted adjacency matrix and node spectral feature vectors of the dynamic graph structure. The spatial aggregation mechanism of the graph convolutional recurrent network is used to fuse the thermal state information of adjacent nodes and generate a high-dimensional node embedding representation containing the local heat conduction context in order to capture the non-uniform heat diffusion behavior in the grinding contact area. The node embedding sequence output by the graph convolutional layer is input into the recursive unit in chronological order. The long-term dependency of the node state is modeled through the hidden state propagation mechanism. The global evolution law of thermal anomalies from local initiation to regional expansion is learned, and the hidden state sequence of future multiple time steps is output. Based on the future hidden states output by the recursive units, the fully connected decoding layer maps them to the probability distribution of the temperature state of each node at future times, generating prediction results with confidence estimates.

10. The online real-time detection and early warning method for gear grinding burns based on thermal imaging according to claim 1, characterized in that, In step S7, a burn risk heat map is generated based on the temperature state probability distribution.

Images