Intelligent mold box material monitoring method based on machine vision

Abstract

The invention relates to the technical field of image processing, and discloses a mold box material intelligent monitoring method based on machine vision. The method comprises the following steps: acquiring mold box data through a multi-dimensional visual acquisition network, and generating a material and environment data set; extracting a material parameter matrix by adopting an improved Gaussian mixture algorithm; fusing the data by using an LSTM-Attention network to generate a feature vector; a heating strategy is constructed based on the material characteristics, and a temperature control instruction is formed; and establishing a scheduling model according to the motion characteristics and the temperature control instruction, and outputting a track scheme and maintenance parameters. The technical problem that accurate recognition and intelligent control of the material state cannot be achieved in an existing mold box material monitoring method is solved, comprehensive real-time monitoring, accurate temperature control and automatic intelligent scheduling of the material in the mold box are achieved by fusing machine vision, multi-sensor data and the deep learning technology, and the material monitoring efficiency is improved. The material maintenance quality and the production efficiency are improved.

Description

Technical Field

[0001] The present application relates to the field of image processing technology, and in particular to a method for intelligently monitoring mold box materials based on machine vision. Background Art

[0002] Mold boxes are commonly used material processing or maintenance equipment in industrial production, and are widely used in the fields of building materials, food processing, chemical synthesis, etc. Traditional mold box material monitoring mainly relies on manual inspections, fixed-point temperature sensors and simple timing control systems. Operators need to judge the material status based on experience and manually adjust the temperature and maintenance parameters. With the development of automation technology, some intelligent control systems have been introduced into mold box management, such as constant temperature systems based on PID control, automation equipment controlled by preset programs, and networked equipment with remote monitoring functions. These technologies have improved the automation level and management efficiency of mold boxes to a certain extent.

[0003] However, the existing technology still has obvious shortcomings in the monitoring and control of mold box materials. First, traditional temperature sensors can only provide point monitoring data in limited locations and cannot fully reflect the internal state and temperature distribution of the material; second, the existing system is difficult to realize intelligent identification and prediction of the material development process, and often adopts fixed parameter control, lacking the ability to adaptively adjust the real-time state of the material; third, the mobile scheduling of the mold box mainly relies on predetermined routes or manual intervention, and fails to form a closed-loop optimization with the material state monitoring and temperature control system; fourth, there is a lack of coordination mechanism between multiple mold boxes, which is inefficient in resource utilization and energy consumption. These problems lead to large fluctuations in material quality, serious energy waste, and low production efficiency, especially when processing high-value or environmentally sensitive materials. Summary of the invention

[0004] The present application provides a method for intelligent monitoring of mold box materials based on machine vision, which is used to solve the technical problem that the existing mold box material monitoring methods cannot achieve accurate identification and intelligent control of material status. By integrating machine vision, multi-sensor data and deep learning technology, comprehensive real-time monitoring, precise temperature control and automated intelligent scheduling of materials in the mold box can be achieved, thereby improving material maintenance quality and production efficiency.

[0005] In a first aspect, the present application provides a mold box material intelligent monitoring method based on machine vision, and the mold box material intelligent monitoring method based on machine vision includes: The image of the material in the mold box and the surrounding environment is collected through a multi-dimensional visual acquisition network to generate a material image dataset and an environmental obstacle image dataset; Based on the material image data set and combined with the temperature sensor data, the improved Gaussian mixture background modeling algorithm is used to extract features to obtain the material consistency parameter matrix, hardness parameter matrix and development state parameter matrix; The material consistency parameter matrix, hardness parameter matrix and development state parameter matrix are input into the improved LSTM-Attention network for multimodal data fusion. At the same time, the environmental obstacle image dataset is combined to generate the material development feature vector and the mold box motion feature vector. According to the material development feature vector, an intelligent heating control strategy is constructed to adaptively adjust the insulation system and auxiliary heating system of the mold box to form a temperature control instruction set; An intelligent scheduling model is established based on the mold box motion feature vector and temperature control instruction set, and the mold box motion trajectory planning scheme and material maintenance parameters are output to realize unmanned intelligent scheduling of the mold box.

[0006] In the technical solution provided in the present application, the machine vision-based mold box material intelligent monitoring method uses a multi-dimensional visual acquisition network to perform all-round image acquisition of the material in the mold box and the surrounding environment, overcomes the limitations of traditional single-point temperature sensor monitoring, realizes real-time monitoring of the full surface state of the material, and significantly improves the spatial coverage and accuracy of the monitoring data. At the same time, the generated material image data set and environmental obstacle image data set provide a comprehensive data basis for subsequent intelligent analysis; combined with temperature sensor data, an improved Gaussian mixture background modeling algorithm is used for feature extraction. The algorithm optimizes the adaptability of the traditional Gaussian model in temperature fluctuation scenarios and effectively suppresses environmental noise interference. The extracted material consistency parameter matrix, hardness parameter matrix and development state parameter matrix comprehensively quantify the physical properties and development degree of the material, and realize multi-dimensional and accurate characterization of the material state; these parameter matrices are input into the improved LSTM-Attention network for multimodal data fusion. The network significantly improves the depth of understanding and prediction accuracy of the material state by introducing attention mechanism and time series analysis capabilities, and can capture subtle changes in material development trends and key moments. The generated material development feature vectors and mold box motion feature vectors realize high-dimensional characterization of material state and equipment movement; based on material The intelligent heating control strategy constructed by the development feature vector realizes the precise control of the temperature of different areas in the mold box by adaptively adjusting the control parameters of the insulation system and the auxiliary heating system, greatly improving the temperature uniformity and energy utilization efficiency, and the formed temperature control instruction set is highly targeted and adaptive; the intelligent scheduling model established based on the mold box motion feature vector and the temperature control instruction set effectively integrates the material state information and environmental constraints, and the output mold box motion trajectory planning scheme and material maintenance parameters realize the full automation and optimization of the maintenance process, reducing the need for manual intervention, shortening the material development cycle, and improving the consistency of material quality; especially In the application of artificial intelligence algorithms, the improved LSTM-Attention network shows significant advantages in processing time series data and multimodal information fusion. Its feature attention layer calculates the importance weights of the material state evolution sequence, effectively identifying the key nodes and turning points in the material development process. The multimodal fusion layer realizes the deep integration of material time series information and environmental space information, solving the problem of heterogeneous data fusion that is difficult to handle with traditional methods. At the same time, through the feature decoupling mechanism of the feature decoding layer, the fused feature matrix is ​​decomposed into two independent dimensions of material development and mold box movement, laying a data foundation for subsequent precise control and intelligent scheduling. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0008] Figure 1 This is a schematic diagram of an embodiment of a method for intelligently monitoring mold box materials based on machine vision in an embodiment of the present application; Figure 2 This is a schematic diagram of the process of capturing images of materials in a mold box and the surrounding environment through a multi-dimensional visual acquisition network in an embodiment of the present application. DETAILED DESCRIPTION

[0009] The embodiment of the present application provides a method for intelligent monitoring of mold box materials based on machine vision.

[0010] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiment of the present application, an embodiment of the mold box material intelligent monitoring method based on machine vision includes: Step S101, collecting images of materials in the mold box and the surrounding environment through a multi-dimensional visual collection network to generate a material image data set and an environmental obstacle image data set; Step S102: Based on the material image data set and in combination with the temperature sensor data, an improved Gaussian mixture background modeling algorithm is used to perform feature extraction to obtain a material consistency parameter matrix, a hardness parameter matrix, and a development state parameter matrix; Step S103, inputting the material consistency parameter matrix, hardness parameter matrix and development state parameter matrix into the improved LSTM-Attention network for multimodal data fusion, and combining the environmental obstacle image data set to generate the material development feature vector and the mold box motion feature vector; Step S104: construct an intelligent heating control strategy according to the material development feature vector, adaptively adjust the insulation system and auxiliary heating system of the mold box, and form a temperature control instruction set; Step S105: Establish an intelligent scheduling model based on the mold box motion feature vector and the temperature control instruction set, output the mold box motion trajectory planning scheme and material maintenance parameters, and realize unmanned intelligent scheduling of the mold box.

[0011] It is understandable that the execution subject of the present application can be a mold box material intelligent monitoring system based on machine vision, or a terminal or a server, which is not limited here. The present application embodiment is described by taking the server as the execution subject as an example.

[0012] Specifically, the multi-dimensional visual acquisition network is used to collect images of the materials in the mold box and the surrounding environment to generate material image datasets and environmental obstacle image datasets. The multi-dimensional visual acquisition network is a stereoscopic imaging system composed of multiple industrial cameras, including a top-view camera, a side wall transverse camera, and an oblique camera. The top-view camera is installed above the mold box and is equipped with an infrared fill light system to capture the surface morphology and height information of the material; the side wall transverse camera is fixed on the side wall of the mold box and uses an anti-fog lens to mainly observe the edge contour and consistency changes of the material; the oblique camera observes the overall flow state of the material from a specific angle. During the acquisition process, the depth camera scans the internal space of the mold box, and the acquired point cloud data is calibrated through spatial registration to separate the material area from the non-material area to generate a material space distribution map. Through the material space distribution map, the material surface contour point set is extracted to establish the material surface height field data. At the same time, the environmental monitoring camera scans the surrounding environment of the mold box, identifies obstacles that may affect the movement of the mold box, and generates an environmental obstacle image dataset. These two datasets together constitute the basic data for subsequent analysis. According to the material image data set, combined with the temperature sensor data, the improved Gaussian mixture background modeling algorithm is used for feature extraction to obtain the material consistency parameter matrix, hardness parameter matrix and development state parameter matrix. The improved Gaussian mixture background modeling algorithm is an image foreground / background separation method based on statistical learning, which is optimized for the characteristics of aerated concrete materials. The algorithm performs multi-scale segmentation on the material image data set, extracts the material surface texture feature point set, and then combines the temperature sensor data for thermal mapping to obtain the material temperature distribution matrix. By separating the background layer of the material temperature distribution matrix, a material surface state feature map is generated, based on which a material consistency evaluation function is constructed, and the material consistency parameter matrix is ​​calculated. At the same time, the material surface state feature map is gradient calculated, and the material surface hardness evaluation criteria are established to obtain the hardness parameter matrix. The material temperature distribution matrix and the material surface state feature map are feature combined, and data mapping is performed according to the material development standard to obtain the development state parameter matrix. These three parameter matrices fully describe the current physical state of the material and provide a data basis for subsequent control decisions.

[0013] The material consistency parameter matrix, hardness parameter matrix and development state parameter matrix are input into the improved LSTM-Attention network for multimodal data fusion, and the environmental obstacle image dataset is combined to generate the material development feature vector and the mold box motion feature vector. The improved LSTM-Attention network is a deep learning architecture that combines the long short-term memory network and the attention mechanism, and is designed for processing time series data and multi-source heterogeneous data. The processing process performs time series alignment and calibration on the material parameter matrix to obtain the material feature time series data, and then inputs it into the network's time series encoding layer for time step decomposition to generate the material state evolution sequence. The feature attention layer calculates the importance weight of the material state evolution sequence and extracts the key time series features of the material, while the spatial encoding layer processes the environmental obstacle image dataset to generate the environmental constraint feature map. The multimodal fusion layer fuses these features and generates the material development feature vector and the mold box motion feature vector through the feature decoding layer.

[0014] According to the material development feature vector, an intelligent heating control strategy is constructed to adaptively adjust the insulation system and auxiliary heating system of the mold box to form a temperature control instruction set. This step analyzes the heating demand based on the material development feature vector, generates a heating area distribution map, and intuitively displays the heating demand of different areas in the mold box. The temperature field is mapped on the heating area distribution map to obtain the target temperature control point sequence, which provides a basis for the power allocation of the insulation system. According to the target temperature control point sequence, the insulation unit of the insulation system is allocated power, the insulation power curve is constructed, and the auxiliary heating power compensation table is established through temperature gradient calculation. The insulation power curve and the auxiliary heating power compensation table are combined to generate a heating control matrix. According to the matrix, the control parameters of the insulation system and the auxiliary heating system are dynamically adjusted to form a temperature control instruction set. An intelligent scheduling model is established based on the mold box motion feature vector and the temperature control instruction set, and the mold box motion trajectory planning scheme and material maintenance parameters are output. This step decomposes the mold box motion feature vector into a position, velocity and acceleration sequence, and establishes a three-dimensional thermal distribution map in combination with the temperature control instruction set, and obtains the temperature field constraint conditions through time and space slicing. Combine motion sequences into dynamic motion state space and map temperature constraints to form a comprehensive constraint space. Construct trajectory cost function based on comprehensive constraint space, and generate candidate trajectory set that meets temperature and obstacle avoidance requirements through path sampling optimization. Extract key path nodes from candidate trajectories, establish maintenance time window sequence, and output mold box motion trajectory planning scheme and material maintenance parameters through multi-step processing to realize unmanned intelligent scheduling of mold boxes.

[0015] In the embodiment of the present application, the machine vision-based mold box material intelligent monitoring method uses a multi-dimensional visual acquisition network to perform all-round image acquisition of the material in the mold box and the surrounding environment, overcomes the limitations of traditional single-point temperature sensor monitoring, realizes real-time monitoring of the full surface state of the material, and significantly improves the spatial coverage and accuracy of the monitoring data. At the same time, the generated material image data set and environmental obstacle image data set provide a comprehensive data basis for subsequent intelligent analysis; combined with temperature sensor data, an improved Gaussian mixture background modeling algorithm is used for feature extraction. The algorithm optimizes the adaptability of the traditional Gaussian model in temperature fluctuation scenarios and effectively suppresses environmental noise interference. The extracted material consistency parameter matrix, hardness parameter matrix and development state parameter matrix comprehensively quantify the physical properties and development degree of the material, and realize multi-dimensional and accurate characterization of the material state; these parameter matrices are input into the improved LSTM-Attention network for multimodal data fusion. The network significantly improves the depth of understanding and prediction accuracy of the material state by introducing attention mechanism and time series analysis capabilities, and can capture subtle changes in material development trends and key moments. The generated material development feature vector and mold box motion feature vector realize high-dimensional characterization of material state and equipment movement; based on material development The intelligent heating control strategy constructed by the feature vector realizes the precise control of the temperature of different areas in the mold box by adaptively adjusting the control parameters of the insulation system and the auxiliary heating system, greatly improving the temperature uniformity and energy utilization efficiency, and the formed temperature control instruction set is highly targeted and adaptive; the intelligent scheduling model established based on the mold box motion feature vector and the temperature control instruction set effectively integrates the material state information and environmental constraints, and the output mold box motion trajectory planning scheme and material maintenance parameters realize the full automation and optimization of the maintenance process, reducing the need for manual intervention, shortening the material development cycle, and improving the consistency of material quality; especially in the application of artificial intelligence algorithms, the improved LSTM-Attention network shows significant advantages in processing time series data and multimodal information fusion. Its feature attention layer calculates the importance weight of the material state evolution sequence, effectively identifying the key nodes and turning points in the material development process, and the multimodal fusion layer realizes the deep integration of material time series information and environmental space information, solving the problem of heterogeneous data fusion that is difficult to handle with traditional methods. At the same time, through the feature decoupling mechanism of the feature decoding layer, the fused feature matrix is ​​decomposed into two independent dimensions of material development and mold box movement, laying a data foundation for subsequent precise control and intelligent scheduling.

[0016] In a specific embodiment, the process of executing step S101 may specifically include the following steps: (1) Use a depth camera to scan the internal space of the mold box, organize the acquired point cloud data according to the spatial coordinate sequence, and obtain three-dimensional point cloud sequence data; (2) Perform spatial registration and calibration on the 3D point cloud sequence data, separate the material area from the non-material area, and generate a material spatial distribution map; (3) Through the material spatial distribution map, extract the material surface contour point set and establish the material surface height field data; (4) Fuse and map the material surface height field data with the time series image to generate a material dynamic change sequence; (5) Combined with the temperature field data collected by the infrared sensor, the thermal distribution of the material dynamic change sequence is mapped to obtain the material temperature distribution characteristics; (6) The material temperature distribution characteristics and the environmental scanning data are divided into regions to generate a material image dataset and an environmental obstacle image dataset, respectively.

[0017] Specifically, Figure 2 As shown, it is a schematic diagram of the process of collecting images of materials in the mold box and the surrounding environment through a multi-dimensional visual acquisition network in an embodiment of the present application. A depth camera is used to scan the internal space of the mold box. The depth camera obtains three-dimensional information through structured light or time-of-flight principle, and the acquired point cloud data is sorted according to the spatial coordinate sequence. Point cloud data is a data set composed of a large number of spatial three-dimensional coordinate points, and each point has spatial position information and possible color information. These scattered point cloud data are sorted and organized according to the spatial position relationship to form an ordered three-dimensional point cloud sequence data.

[0018] Spatial registration and calibration of three-dimensional point cloud sequence data is the process of unifying point cloud data collected from multiple perspectives into the same coordinate system. This process uses the iterative closest point (ICP) algorithm for point cloud registration. By finding corresponding point pairs and minimizing the error between them, the point clouds of different perspectives can be accurately spliced. After the registration is completed, the material area is separated from the non-material area by a segmentation method based on region growing. This method selects seed points and then gradually expands the area based on spatial continuity and color similarity to generate a material spatial distribution map, so that the material area and the non-material area are clearly distinguished in space. The material surface contour point set is extracted from the material spatial distribution map using edge detection and surface fitting technology. Edge detection uses the density change and normal vector change of points in three-dimensional space to identify the material boundary and extract the surface contour point set. The contour point set describes the external boundary and internal structural characteristics of the material surface. Based on the contour point set, the material surface height field data is established by B-spline surface fitting technology. The height field is a two-dimensional array that records the height value of each position point on the material surface, forming a digital representation of the material surface.

[0019] Through the feature point matching algorithm, the spatial correspondence between the height field data and the two-dimensional image is established. At the same time, combined with the timestamp information, the height field data at different times are arranged in chronological order to form a dynamic sequence of material surface changes over time. This spatiotemporal fusion method not only retains the geometric features of the material surface, but also contains the dynamic information of the material changes over time, providing a comprehensive data basis for subsequent material state analysis. Combined with the temperature field data collected by the infrared sensor, thermal distribution mapping of the dynamic change sequence of the material is a process of fusing temperature information with geometric information. This step uses the following thermal distribution mapping formula:

[0020] in, Represents the mapped temperature value of the coordinate point (x, y) at time t; N is the number of temperature sampling points; is the weight coefficient of the i-th temperature point; is the measured temperature value of the i-th temperature point; is the spatial attenuation factor, controlling the temperature influence range; Represents the point (x, y) and the temperature sampling point ( , ) between ; It is a temperature correction function related to the height field H(x,y,t), which reflects the influence of the material surface height on the temperature distribution. Through this formula, the discrete temperature sampling point data is converted into a continuous temperature distribution field, and combined with the material dynamic change sequence, the material temperature distribution characteristics are obtained.

[0021] The temperature distribution characteristics of materials and environmental scanning data are divided into regions using a clustering method based on temperature and spatial characteristics. This method comprehensively considers spatial position, temperature value and texture characteristics, and uses an improved K-means clustering algorithm to classify data points, and classifies points with similar characteristics into the same category. By setting reasonable temperature thresholds and spatial distance thresholds, the data is divided into material areas and environmental obstacle areas, and material image datasets and environmental obstacle image datasets are generated respectively.

[0022] For example, the depth camera scans the internal space of the mold box at a frequency of 30 frames per second, and each frame contains about 100,000 point cloud data points. After coordinate transformation and noise filtering, these point cloud data form an orderly three-dimensional point cloud sequence. The ICP algorithm is used to align the point clouds collected from different angles, and the alignment error is controlled within 0.5mm to achieve accurate splicing of the point clouds. The segmentation method based on region growing separates the concrete material area from the non-material area such as the mold box wall to generate a material space distribution map. About 10,000 surface contour points are extracted from the material space distribution map, and the material surface height field data with a resolution of 500×500 is established through B-spline surface fitting. The height field data is matched with the time-series image collected at 5 frames per second for feature points, and the time-space correspondence is established to generate a material dynamic change sequence. The infrared sensor is arranged in a 16×16 array to collect the temperature field data in the mold box with a sampling frequency of 1Hz. The temperature data is mapped to the material dynamic change sequence through the thermal distribution mapping formula to obtain the material temperature distribution characteristics. The temperature range is between 20℃ and 80℃, with an accuracy of ±0.5℃. Finally, the improved K-means clustering algorithm is used to combine the spatial position and temperature characteristics to divide the data into material area and environmental obstacle area, and generate material image datasets and environmental obstacle image datasets respectively.

[0023] In a specific embodiment, the process of executing step S102 may specifically include the following steps: (1) Perform multi-scale image segmentation on the material image dataset and extract the material surface texture feature point set; (2) Combine the temperature sensing data to perform thermal mapping on the material surface texture feature point set to obtain the material temperature distribution matrix; (3) The background layer of the material temperature distribution matrix is ​​separated by improving the Gaussian mixture background modeling algorithm to generate a material surface state characteristic map; (4) Based on the material surface state characteristic diagram, the material consistency evaluation function is constructed and the material consistency parameter matrix is ​​calculated; (5) Perform gradient calculation on the material surface state characteristic diagram, establish the material surface hardness evaluation criteria, and obtain the hardness parameter matrix; (6) The material temperature distribution matrix and the material surface state characteristic diagram are combined, and data mapping is performed according to the material development standard to obtain the material consistency parameter matrix, hardness parameter matrix and development state parameter matrix.

[0024] Specifically, multiscale image segmentation refers to the segmentation of images in different scale spaces, and the original image is converted into an image sequence of different resolutions through pyramid decomposition. A Gaussian blur filter is applied to the original image, and then a downsampling operation is performed to generate images of different resolutions. In each scale space, the superpixel-based SLIC (Simple Linear Iterative Clustering) algorithm is used for image segmentation. The SLIC algorithm groups image pixels according to color similarity and spatial proximity to form superpixel regions with similar characteristics. Then, the information of the segmentation results of each scale is fused to extract the texture feature point set of the material surface. The texture feature point set contains the local structural features of the material surface, such as edges, corners, texture gradients, etc. These feature points are important indicators for characterizing the surface state of the material. When combining temperature sensor data to perform thermal mapping on the texture feature point set of the material surface, it is necessary to establish a spatial correspondence between the temperature sensor data and the texture feature point set. The temperature sensor data is mapped to the material surface spatial coordinate system through coordinate transformation, and then the discrete temperature data points are interpolated and extended using the bilinear interpolation algorithm to generate a continuous temperature field distribution. For each point in the texture feature point set, the corresponding temperature value is queried in the temperature field according to its spatial position, and the temperature value is associated as the attribute information of the point to form a texture feature point set with temperature attributes. These feature points with temperature information are rearranged and organized according to spatial distribution to construct a material temperature distribution matrix, which contains both the texture structure information of the material surface and the thermal distribution information.

[0025] The key to accurately identifying the material state is to separate the background layer of the material temperature distribution matrix by improving the Gaussian mixture background modeling algorithm. The standard Gaussian mixture model (GMM) is a probability density estimation method that assumes that the data is composed of multiple Gaussian distributions. In this method, the improved Gaussian mixture background modeling algorithm introduces time factors and temperature weight factors on the basis of the standard GMM, making the model more suitable for material state monitoring scenarios. The algorithm establishes multiple Gaussian distribution models for each point in the temperature distribution matrix, and each Gaussian distribution describes the temperature and texture feature distribution of the point under different states. The parameters of the Gaussian mixture model, including mean, variance and weight, are iteratively optimized by the expectation maximization (EM) algorithm. Then, according to the weights and variances of each Gaussian distribution, the points in the temperature distribution matrix are classified into foreground (active material area) and background (stable material area), the background layer is separated, and the material surface state feature map is generated. The feature map clearly shows the activity level and state change area of ​​the material surface.

[0026] Based on the material surface state characteristic diagram, the material consistency evaluation function is constructed and the material consistency parameter matrix is ​​calculated. This process uses the following consistency evaluation function: ; in, Represents the material consistency parameter at the coordinate point (i, j); is the value of the material surface state characteristic diagram at point (i, j); is the neighborhood window radius; is the weight coefficient within the window, satisfying ; is the Laplace operator value of S(i,j), which represents the second-order rate of change of the surface state of the material; is the average value of S; Yes Surface fluctuation value at ; is the standard deviation of V; is the temperature value at point (i, j); is the temperature influence function; , , is the weight coefficient, which controls the influence of neighborhood state, surface curvature and temperature fluctuation on consistency evaluation, m is the horizontal coordinate offset of coordinate point (i, j), and n is the vertical coordinate offset of coordinate point (i, j). By applying this function to each point of the material surface state characteristic diagram, the material consistency parameter matrix is ​​calculated, which quantifies the consistency state of the material at different positions.

[0027] The gradient calculation is performed on the material surface state characteristic map, and the material surface hardness evaluation criteria are established. In the process of obtaining the hardness parameter matrix, the Sobel operator is applied to the material surface state characteristic map to calculate the gradient. The Sobel operator is a commonly used edge detection operator. By applying horizontal and vertical convolution kernels to the image, the gradient size and direction of the image at each point are calculated. For each point in the characteristic map, its horizontal gradient and vertical gradient are calculated, and then the total gradient size is synthesized. According to the gradient information and temperature information, a hardness evaluation criterion is constructed. The criterion takes into account the gradient change, temperature distribution and time factors of the material surface to form a hardness parameter calculation model. The model is applied to each point of the characteristic map to calculate the hardness parameter matrix, which quantifies the hardness state of the material surface at different positions. It is an important part of the comprehensive evaluation of the material state to combine the material temperature distribution matrix with the material surface state characteristic map, map the data according to the material development standard, and obtain the material consistency parameter matrix, hardness parameter matrix and development state parameter matrix. The feature combination adopts a multi-channel fusion method, and the temperature distribution matrix and the state characteristic map are combined as different feature channels. Then, according to the predefined material development standard, a data mapping rule is established to convert the combined features into development state parameters. The material development standard is a reference standard pre-set according to the material type and application scenario, which includes the consistency, hardness and state characteristic indicators of the material at different development stages. By comparing and mapping the combined characteristics with the development standard, a material development state parameter matrix is ​​generated, which comprehensively reflects the development state of the material in spatial distribution.

[0028] Taking the monitoring of cement concrete materials in the mold box as an example, when performing multi-scale image segmentation on the collected material image dataset with a resolution of 1280×960, a 4-level image pyramid is constructed with resolutions of 1280×960, 640×480, 320×240, and 160×120. The SLIC algorithm is used for superpixel segmentation at each scale, and the number of superpixels is set to 1000, 500, 250, and 125, respectively. The segmentation results are subjected to feature extraction to obtain about 5000 surface texture feature points. The temperature sensor array is arranged around the mold box to collect temperature data at a frequency of 5Hz, and a temperature field with a resolution of 100×100 is generated by bilinear interpolation. The temperature field data is spatially mapped with the texture feature point set to generate the material temperature distribution matrix. The improved Gaussian mixture background modeling algorithm is used to analyze the temperature distribution matrix, in which the mixed Gaussian model uses three Gaussian components with initial weights of 0.6, 0.3, and 0.1, respectively. The EM algorithm was iterated 20 times to optimize the Gaussian model parameters, achieve background layer separation, and generate the material surface state characteristic map. Based on the characteristic map, the consistency evaluation function was constructed, the window radius w was set to 3, the weight coefficients λ1, λ2, and λ3 were set to 0.5, 0.3, and 0.2, respectively, and a 100×100 consistency parameter matrix was calculated. The 3×3 Sobel operator was applied to the characteristic map to calculate the gradient, and the hardness evaluation criteria were established in combination with the temperature information to generate the hardness parameter matrix. Finally, the temperature distribution matrix and the state characteristic map were combined, and data mapping was performed according to the predefined cement concrete development standard to obtain the material development state parameter matrix. Through these parameter matrices, the accurate monitoring and evaluation of the concrete material state in the mold box was achieved.

[0029] In a specific embodiment, the process of executing step S103 may specifically include the following steps: (1) Perform time series alignment and calibration on the material consistency parameter matrix, hardness parameter matrix and development state parameter matrix to obtain material characteristic time series data; (2) Input the material feature time series data into the time series encoding layer of the improved LSTM-Attention network for time step decomposition to generate a material state evolution sequence; (3) The feature attention layer of the improved LSTM-Attention network is used to calculate the importance weight of the material state evolution sequence and extract the key time series features of the material; (4) Input the environmental obstacle image dataset into the spatial encoding layer of the improved LSTM-Attention network to generate an environmental constraint feature map; (5) The multimodal fusion layer of the improved LSTM-Attention network performs feature mapping on the key timing features of the materials and the environmental constraint feature graph to obtain a fusion feature matrix; (6) The fused feature matrix is ​​input into the feature decoding layer of the improved LSTM-Attention network for feature decoupling to generate the material development feature vector and the mold box movement feature vector.

[0030] Specifically, time series alignment calibration is to correspond multiple parameter matrices collected at different time points to a unified time axis to ensure the consistency of data in the time dimension. A timestamp is added to each parameter matrix to record the exact time of data collection. Next, a reference time axis is selected, usually the time axis of the parameter matrix with the highest sampling frequency. For parameter matrices with lower sampling frequency, the data values ​​at the intermediate moments are calculated by linear interpolation or spline interpolation methods, and all parameter matrices are resampled to the reference time axis. Then, time window division is performed to divide the continuous time series into windows of fixed length, each of which contains a certain number of time steps. Finally, the aligned consistency parameter matrix, hardness parameter matrix, and development state parameter matrix are combined in chronological order to form a time series data set containing multiple material characteristics, namely, material feature time series data. Inputting material feature time series data into the time series encoding layer of the improved LSTM-Attention network for time step decomposition is a key step in deep feature extraction of time series data. LSTM (Long Short-Term Memory Network) is a special recurrent neural network that can effectively handle long-term dependencies in long sequence data. The temporal coding layer is composed of multiple layers of LSTM units, and each layer of LSTM units receives the output of the previous layer as input. In the improved LSTM-Attention network, the standard LSTM unit is improved and a time step adaptive mechanism is introduced, which can automatically adjust the size of the time step according to the data characteristics. The temporal coding layer divides the material feature time series data according to the preset initial time step, and then adjusts the step size of each time period through the adaptive mechanism, using a smaller step size for fine encoding of important time periods and a larger step size for general encoding of time periods with insignificant changes. After multi-layer LSTM processing, the temporal coding layer outputs the feature representation of the material state at different time scales to form a material state evolution sequence.

[0031] Calculating the importance weight of the material state evolution sequence through the feature attention layer of the improved LSTM-Attention network is an important step in extracting key information. The feature attention layer uses the attention mechanism to automatically learn the importance of each time step and feature dimension in the sequence. The attention mechanism calculates the query vector (Query), key vector (Key) and value vector (Value). The query vector represents the information that needs to be paid attention to at present, the key vector represents the characteristics of the reference information, and the value vector represents the content of the reference information. The attention weight score of each time step is obtained by calculating the similarity between the query vector and the key vector. Then these scores are normalized to obtain the attention weight. Finally, the attention weight is weighted and summed with the value vector to obtain the context vector, which concentrates the most important information in the sequence. In this method, the feature attention layer calculates the attention weight for each time step in the material state evolution sequence, identifies the most critical time series features for material state judgment, and extracts the key time series features of the material.

[0032] Inputting the environmental obstacle image dataset into the spatial encoding layer of the improved LSTM-Attention network is an important step in processing environmental information. The spatial encoding layer adopts a convolutional neural network (CNN) structure, which can effectively extract spatial features from image data. The spatial encoding layer applies multi-layer convolution operations to the environmental obstacle image dataset, and each convolution layer includes convolution, batch normalization, and activation function. The convolution operation extracts features from the image by sliding windows, batch normalization accelerates network convergence, and the activation function introduces nonlinear transformation. After multi-layer convolution, the pooling operation is applied to reduce the size of the feature map, reduce the amount of calculation, and extract higher-level features. Finally, the feature map is converted into a feature vector through a fully connected layer to generate an environmental constraint feature map, which contains the spatial distribution, shape characteristics, and potential influencing factors of environmental obstacles. The multimodal fusion layer of the improved LSTM-Attention network performs feature mapping on the key time series features of materials and the environmental constraint feature map, which is the key to realizing the fusion of different types of data. The multimodal fusion layer adopts a cross-attention mechanism to establish the association relationship between different modal data. The cross-attention mechanism uses the features of one modality as the query vector and the features of the other modality as the key vector and value vector, calculates the attention weights between them, and realizes information interaction between modalities. In this method, the multimodal fusion layer projects the key time series features of the material and the environmental constraint feature map into the same feature space, and then calculates the attention weight matrix between them. The features are weighted and fused through the attention weight matrix to generate a fused feature matrix containing material status information and environmental constraint information. This fused feature matrix not only retains the change characteristics of the material in the time dimension, but also includes the impact of environmental factors on the material status.

[0033] Inputting the fused feature matrix into the feature decoding layer of the improved LSTM-Attention network for feature decoupling is the last step to generate the feature vector required for a specific application. The feature decoding layer adopts an autoencoder structure, which can decompose the mixed features into different semantic components. The autoencoder consists of two parts: the encoder and the decoder. The encoder compresses the input data into a potential representation, and the decoder reconstructs the potential representation into the original data. In the process of feature decoupling, orthogonal constraints are introduced to ensure that the decoupled features are independent of each other. The feature decoding layer applies a fully connected layer to the fused feature matrix for dimensionality reduction to obtain a potential representation. Then, through two independent decoding branches, one branch focuses on extracting material development-related features, and the other branch focuses on extracting mold box motion-related features. Finally, the outputs of the two branches are further processed to generate material development feature vectors and mold box motion feature vectors. These two feature vectors contain material development state information and mold box motion state information, respectively, providing a data basis for subsequent intelligent heating control and motion scheduling.

[0034] Taking the monitoring of concrete mold box materials as an example, during the time series alignment calibration process, the consistency parameter matrix is ​​collected at a frequency of 1Hz, the hardness parameter matrix is ​​collected at a frequency of 0.5Hz, and the development state parameter matrix is ​​collected at a frequency of 0.2Hz. Based on the time axis of the consistency parameter matrix, cubic spline interpolation is performed on the hardness parameter matrix and the development state parameter matrix, and the three parameter matrices are resampled to a unified 1Hz time axis. The resampled data is divided into a time window of 30 seconds, and each window contains 30 time steps. The data of the three parameter matrices are combined to form material feature time series data with a dimension of [30, 3, 100, 100], where 30 represents the number of time steps, 3 represents the number of parameter types, and 100×100 represents the spatial resolution of each parameter matrix. The material feature time series data is input into the time series encoding layer composed of three layers of LSTM units, the LSTM hidden layer dimension is set to 128, and the initial time step is set to 1 second. Through the time step adaptation mechanism, a small step of 0.5 seconds is used for the period when the material state changes significantly, and a large step of 2 seconds is used for the period when the change is slow, to generate the material state evolution sequence. The feature attention layer adopts a multi-head attention mechanism with the number of heads set to 8. The attention weights are calculated for the material state evolution sequence to extract the key time series features of the material. In parallel, the spatial encoding layer applies five layers of convolution operations to the environmental obstacle image dataset, with the convolution kernel sizes of 7×7, 5×5, 3×3, 3×3, and 3×3, and the number of channels of 32, 64, 128, 256, and 512, respectively, to generate the environmental constraint feature map. The multimodal fusion layer realizes the fusion of the key time series features of the material and the environmental constraint feature map through a 6-head cross attention mechanism, and generates a fusion feature matrix with a dimension of [512, 64]. The feature decoding layer generates a material development feature vector with a dimension of 256 and a mold box motion feature vector with a dimension of 128 through two independent fully connected branches, respectively, to provide input data for the intelligent heating control strategy and scheduling model.

[0035] In a specific embodiment, the process of executing step S104 may specifically include the following steps: (1) Perform heating demand analysis based on material development feature vectors and generate a heating area distribution map; (2) Map the temperature field of the heating area distribution map to obtain the target temperature control point sequence; (3) Allocate power to the insulation units of the insulation system according to the target temperature control point sequence and construct an insulation power curve; (4) Calculate the temperature gradient based on the insulation power curve and establish an auxiliary heating power compensation table; (5) Combine the insulation power curve and the auxiliary heating power compensation table to generate a heating control matrix; (6) Dynamically adjust the control parameters of the insulation system and the auxiliary heating system according to the heating control matrix to form a temperature control instruction set.

[0036] Specifically, the material development feature vector is high-dimensional data extracted by the LSTM-Attention network, which contains key information about the material development state. The heating demand analysis uses principal component analysis to reduce the dimension of the high-dimensional feature vector and retain the main information. Then, the K-means clustering algorithm is applied to classify the regions with similar development states into the same category, and the heating demand index of each region is calculated based on the clustering results. These indexes reflect the degree of heat demand at different locations, and are mapped to the three-dimensional space to form a heating area distribution map, which intuitively displays the heating demand intensity at each location inside the mold box. Temperature field mapping of the heating area distribution map is the process of determining the specific temperature target. The corresponding relationship between the heating demand and the target temperature is established, and the heating demand index is converted into a specific temperature value to form a target temperature distribution map. Then, the non-uniform grid division method is used to select key control points, select dense points in areas with large temperature gradients, and select sparse points in areas with uniform temperature. Each control point is assigned a target temperature value and a control weight, and is sorted according to spatial position and importance to form a target temperature control point sequence.

[0037] A heat transfer model of target temperature and insulation power is established, taking into account factors such as heat conduction, heat convection, and heat radiation. Then, a power allocation optimization problem is constructed, with the goal of minimizing the difference between the actual temperature and the target temperature while taking into account energy minimization and temperature uniformity. A genetic algorithm is used to solve the optimization problem, obtain the optimal power configuration for each insulation unit, and arrange them in chronological order to form an insulation power curve. Calculating the temperature gradient based on the insulation power curve and establishing an auxiliary heating power compensation table are the keys to solving the limitations of the insulation system. By calculating the gradient and rate of change of the predicted temperature field generated by the insulation power curve, the areas and time points that are difficult for the insulation system to meet are identified. For these areas and time points, the required auxiliary heating power compensation value is calculated based on the temperature gradient, the difference between the temperature and the target value, and the temperature change rate to form an auxiliary heating power compensation table.

[0038] Combine the insulation power curve with the auxiliary heating power compensation table to generate a complete heating control strategy. Align the two by spatial position and time to ensure consistent data dimensions. Then, according to the material development status and environmental conditions, dynamically adjust the weight ratio of the insulation system and the auxiliary heating system, increase the insulation weight in the stable stage, and increase the auxiliary heating weight in the rapid adjustment stage. The heating control matrix is ​​obtained through weighted combination, which contains the power configuration of each heating unit at different time points. Dynamic adjustment of the control parameters of the insulation system and the auxiliary heating system according to the heating control matrix is ​​the last step to achieve precise temperature control. Extract the power configuration from the heating control matrix and convert it into specific control parameters such as current, voltage or duty cycle. Combine the temperature sensor feedback, calculate the temperature deviation in real time, and use the PID control method to fine-tune the parameters. At the same time, consider delay compensation, predict future temperature trends, and adjust parameters in advance. Generate a temperature control instruction set, which contains the commands, execution timing and priority information of each control unit.

[0039] Taking the temperature control of the concrete mold box as an example, after the material development feature vector is reduced in dimension, the K-means algorithm divides the space into 5 heating areas and generates a heating area distribution map. The temperature field mapping converts the heating demand into the target temperature, and selects 32 temperature control points to form a control sequence. The power is allocated to the 16 insulation units to generate an insulation power curve. The temperature gradient and change rate are calculated, and an auxiliary heating power compensation table is established. The weight ratio of the two systems is dynamically adjusted to generate a heating control matrix. The control parameters are adjusted through the PID control method to form a temperature control instruction set to achieve a control accuracy of ±1°C.

[0040] In a specific embodiment, the process of executing step S105 may specifically include the following steps: (1) Decompose the mold box motion feature vector according to the time sequence to generate the mold box position change sequence, velocity change sequence and acceleration change sequence; (2) Establish a three-dimensional thermal distribution map based on the temperature control instruction set, slice the thermal distribution map in time and space, and obtain the temperature field constraint conditions at each time point; (3) Combine the position change sequence, velocity change sequence, and acceleration change sequence to construct a dynamic motion state space, and map the temperature field constraints into the motion state space; (4) Perform collision detection on the motion state space, generate a safe obstacle avoidance area, and form a comprehensive constraint space in combination with the temperature field constraint conditions; (5) Construct a trajectory cost function based on the comprehensive constraint space, optimize the path sampling of the cost function, and generate a set of candidate trajectories that meet the temperature requirements and obstacle avoidance requirements; (6) Extract key path nodes from the candidate trajectory set, establish a maintenance time window sequence, match the node information with the maintenance requirements, and output the mold box motion trajectory planning scheme and material maintenance parameters.

[0041] Specifically, the mold box motion feature vector is a high-dimensional feature vector output by the improved LSTM-Attention network, which contains the key information of the dynamic motion state of the mold box. The dynamic time warping (DTW) algorithm is used in the time series decomposition process. The algorithm finds the best matching path between two time series by comparing the similarity of the two time series. The motion feature vector is sliced ​​in the time dimension, and the continuous feature vector is divided into time series according to the timestamp. Then the time series is parsed by the recursive neural network decoder to extract the position, velocity and acceleration information of the mold box at different times. The position change sequence records the coordinate change trajectory of the mold box in three-dimensional space, the velocity change sequence represents the motion speed of the mold box in all directions, and the acceleration change sequence reflects the rate of change of the motion state of the mold box. These three sequences together constitute a complete description of the motion state of the mold box. Establishing a three-dimensional thermal distribution map based on the temperature control instruction set is the key to realizing temperature field visualization and constraint analysis. The temperature control instruction set is a series of temperature control parameters generated by the intelligent heating control strategy, including target temperature at different positions, heating power and other information. The discrete temperature control points in the temperature control instruction set are extended to the entire three-dimensional space through the trilinear interpolation method to generate a continuous temperature field function. Then, the Markov random field model is used to perform probability modeling of the temperature field, taking into account the laws of heat conduction and material properties, and predicting the temperature changes at each point over time. Based on the prediction results, a three-dimensional thermal distribution map is constructed, which represents the temperature values ​​of each point in space in the form of a three-dimensional grid. The thermal distribution map is sliced ​​in time and space, that is, two-dimensional or three-dimensional slices of the spatial temperature distribution are intercepted at different time points to analyze the temporal and spatial variation characteristics of the temperature field. The temperature threshold surface, temperature gradient surface, and temperature change rate surface are extracted from the slices to form the temperature field constraints at each time point. These constraints define the temperature requirements that need to be met during the movement of the mold box.

[0042] Combining the position change sequence, velocity change sequence and acceleration change sequence to construct the dynamic motion state space is an important step to achieve a comprehensive characterization of the mold box state. The dynamic motion state space is a high-dimensional phase space, and each point represents the complete motion state of the mold box at a specific moment. The combination process adopts the state space model, aligns the three sequences in the time dimension and then performs feature fusion. Each state point contains information such as position coordinates, velocity vectors and acceleration vectors. Then the temperature field constraints are mapped to the motion state space. The mapping process uses coordinate transformation and feature projection methods to represent the temperature constraints as constraint areas or boundary conditions in the state space. By establishing the position-temperature correspondence, velocity-temperature gradient correspondence and acceleration-temperature change rate correspondence, the coupling of temperature field constraints and motion states is realized, forming a comprehensive state space containing kinematic information and temperature constraints.

[0043] Collision detection uses a hierarchical bounding box algorithm to represent the mold box and environmental obstacles with a series of nested bounding boxes. By comparing the positional relationship between the bounding boxes, pairs of objects that are unlikely to collide are quickly eliminated to improve detection efficiency. Accurate collision detection is performed on pairs of objects that may collide, and the minimum distance and potential collision points are calculated. Based on the collision detection results, the artificial potential field method is used to construct a safe obstacle avoidance area. This method regards obstacles as repulsive sources and target positions as gravitational sources. By setting appropriate potential field functions, dangerous areas that the mold box should avoid and areas that can be safely passed are generated. The safe obstacle avoidance area is combined with the temperature field constraints, and a multi-objective constraint fusion algorithm is used to generate a comprehensive constraint space that comprehensively considers safety and temperature requirements. This space defines the boundary conditions for the motion planning of the mold box. Constructing a trajectory cost function based on the comprehensive constraint space is the core of trajectory planning optimization. The trajectory cost function quantifies the quality of the trajectory, usually considering factors such as path length, smoothness, energy consumption, and constraint satisfaction. The cost function is constructed using a weighted sum method, which assigns different weights to each indicator according to its importance, and comprehensively evaluates the trajectory quality. When optimizing the path sampling of the cost function, the Rapid Exploration Random Tree (RRT*) algorithm is used. This algorithm gradually builds a tree structure connecting the start and end points by randomly sampling nodes in the state space. During the growth of the tree, the pros and cons of the new nodes are evaluated according to the cost function, and the nodes with low costs are expanded first, and the path is optimized. Through multiple iterations of sampling and optimization, a series of candidate trajectories that meet the temperature requirements and obstacle avoidance requirements are generated. Each trajectory represents a feasible motion plan for the mold box from the start to the end.

[0044] Critical path nodes are locations with special significance in the trajectory, such as inflection points, dwell points, or points where temperature conditions change. The extraction process uses curvature analysis and density clustering methods to calculate the curvature values ​​of each point on the trajectory and identify points where the curvature changes significantly as inflection points; the dwell time analysis is used to identify points where the mold box needs to stay to perform specific operations; combined with temperature condition analysis, points where the temperature environment changes significantly are identified. The extracted key nodes are arranged in chronological order to form a critical path node sequence. The time windows are divided based on the key node sequence, and each window corresponds to the dwell or movement stage of the mold box in a specific area. The node information is matched with the material maintenance requirements, and the temperature conditions at each node position are analyzed to see whether they meet the material maintenance requirements, and the appropriate dwell time and maintenance parameters are calculated. According to the matching results, the trajectory parameters, such as speed curve, dwell time, etc., are adjusted, and the optimized mold box motion trajectory planning scheme and material maintenance parameters are output.

[0045] Taking the intelligent scheduling of concrete mold boxes as an example, the 256-dimensional mold box motion feature vector obtained from the LSTM-Attention network is decomposed into time series by the DTW algorithm and the recursive neural network decoder to generate a position change sequence, a velocity change sequence, and an acceleration change sequence with 10 sampling points per second. The sequence length is 600, covering a 60-second motion cycle. Based on the 32 temperature control points in the temperature control instruction set, a three-dimensional thermal distribution map with a resolution of 50×50×10 is constructed by trilinear interpolation, and the temperature range is between 20℃ and 80℃. The thermal distribution map is sliced ​​and sampled at 10 time points to obtain the temperature field constraints at each time point, including the curing area with a temperature not lower than 45℃ and the uniformity requirement that the temperature gradient does not exceed 5℃ / m. The position, velocity, and acceleration sequences are combined into a 9-dimensional state vector to construct a dynamic motion state space, and the temperature field constraints are mapped to the conditional region in the state space. The hierarchical bounding box algorithm is used to perform collision detection on the five fixed obstacles and three moving obstacles in the workspace to generate a safe obstacle avoidance area, which is combined with the temperature field constraints to form a comprehensive constraint space. A comprehensive cost function was constructed with weights of 0.3 for path length, 0.2 for smoothness, 0.2 for energy consumption, and 0.3 for temperature constraint. The RRT* algorithm was used for 2000 iterations of sampling to generate 15 candidate trajectories that met the requirements. Eight key path nodes were extracted from the optimal trajectory, and seven curing time windows were established based on the curvature threshold of 0.5 and the residence time analysis, with the window duration ranging from 30 seconds to 180 seconds. The temperature conditions at the node positions were matched with the concrete curing requirements, and the residence time and heating power of each node were calculated. The mold box motion trajectory planning scheme and material curing parameter table containing detailed time arrangements and temperature control parameters were output.

[0046] In a specific embodiment, the process of extracting key path nodes from the candidate trajectory set, establishing a maintenance time window sequence, and matching node information with maintenance requirements may specifically include the following steps: (1) Perform path fitting analysis on the candidate trajectory set and extract the key path node sequence through the curvature extreme points; (2) Perform time segmentation processing according to the critical path node sequence and construct the initial maintenance time window sequence; (3) Perform temperature field matching on the initial curing time window sequence to obtain the curing temperature range for each time window; (4) Calculate the dwell time of the key path node sequence based on the maintenance temperature range and generate a node dwell schedule; (5) Match the node stay schedule with the material maintenance requirements in terms of temperature and time, and construct a maintenance control parameter table; (6) Optimize the trajectory of the key path node sequence according to the maintenance control parameter table, and output the mold box motion trajectory planning scheme and material maintenance parameters.

[0047] Specifically, the candidate trajectory set is a series of feasible paths that meet the temperature and obstacle avoidance requirements generated by the trajectory planning algorithm. The path fitting analysis uses a spline curve fitting algorithm to convert discrete trajectory points into a continuous and smooth curve representation. Then the curvature value of each point on the curve is calculated. Curvature is a mathematical quantity that describes the degree of curvature of the curve. The larger the curvature value, the sharper the turn. The curvature function is differentially analyzed to identify the maximum and minimum points of the curvature, which usually correspond to the key turning positions in the trajectory. Considering the speed change and acceleration change at the same time, candidate nodes are added at positions where the speed is close to zero or the acceleration changes significantly. All candidate nodes are screened and merged, and overly dense nodes are removed. Representative key nodes are retained to form a critical path node sequence. Time segmentation processing based on the critical path node sequence is a key step in establishing the maintenance time window. Time segmentation processing calculates the distance between adjacent nodes, considers the motion speed characteristics of the mold box, and converts the distance into an estimated driving time. For nodes that need to stop for maintenance operations, the initial stay time is estimated based on the material state and maintenance requirements. The driving time and the stay time are combined to form a complete timeline. Then, the timeline is segmented to distinguish between the movement phase and the stay phase, and a time label is assigned to each phase. Considering the continuity requirements of material maintenance, the short segments are merged to ensure that each maintenance window has enough time. At the same time, the maintenance efficiency and energy consumption are balanced to avoid too frequent start and stop operations. Through these processes, the initial maintenance time window sequence is constructed, and each window contains information such as the start time, end time, and operation type.

[0048] Temperature field matching for the initial curing time window sequence is an important step in determining the curing temperature requirements. Temperature field matching collects temperature data within each time window, which comes from real-time monitoring by temperature sensors or prediction by thermal models. The collected temperature data constitutes a window temperature sampling point set, which records the temperature values ​​at different positions and time points within the window. Then the temperature gradient distribution is calculated to analyze the temperature variation trend in space and time. The temperature variation trend graph is divided according to the time window boundary to obtain the temperature curve of each window. Fluctuation analysis is performed on these temperature curves, and statistical indicators such as the maximum, minimum, average and standard deviation of the temperature are calculated to construct a temperature fluctuation range table. According to the material's sensitivity to temperature and the allowable fluctuation range, the appropriate temperature range for each time window, i.e., the curing temperature range, is determined. Calculating the residence time of the key path node sequence based on the curing temperature range is the key to determining the residence time of the mold box at each node. The residence time calculation takes into account the relationship between material type, development stage and curing temperature. Different materials require different curing times at different temperatures. A corresponding model between temperature and curing time is established. The higher the temperature, the faster the curing rate is usually, but it is necessary to avoid excessive temperature causing material damage. For each key node, calculate the residence time required to achieve the expected curing effect based on the curing temperature range and current state of the material at its location. Consider the cumulative effect of material curing and add up the curing effects of several consecutive nodes to avoid repeated curing. At the same time, take into account energy efficiency and overall time constraints to balance the allocation of residence time for each node. Generate a node residence schedule to record in detail the start time, end time and duration of each key node.

[0049] Temperature-time matching of node dwell schedules with material curing requirements is an important step in optimizing curing control. Temperature-time matching normalizes the time data in the dwell schedule and maps time values ​​of different scales to a unified interval for easy comparison with standard curing requirements. Then, a temperature-time correspondence table is established based on the material type and curing standard, which defines the time required to achieve a specific curing effect at different temperatures. The normalized time is compared with the standard correspondence to evaluate whether the curing parameters of each node meet the requirements. For nodes that do not meet the requirements, the temperature setting or dwell time is adjusted to ensure that the curing effect meets the standard. At the same time, parameter boundary checks are performed to ensure that all parameters are within the range allowed by the equipment. Through multiple rounds of optimization and adjustment, a curing control parameter table is constructed, which contains specific control parameters such as temperature setting, dwell time, and heating power for each node. Trajectory optimization of the key path node sequence based on the curing control parameter table is a key step in generating a solution. Trajectory optimization adjusts the time attributes of key nodes according to the dwell time requirements in the curing control parameter table. Then, based on the modified time constraints, recalculate the motion speed and acceleration curves between nodes to ensure smooth and coherent motion and meet the equipment capacity limitations. For areas that require precise temperature control, adjust the trajectory path so that the mold box stays in a position with better temperature conditions as much as possible. At the same time, consider energy efficiency and optimize the motion path to reduce unnecessary energy consumption. Combined with safety obstacle avoidance requirements, ensure that the optimized trajectory still meets the collision avoidance conditions. After multi-objective optimization calculations, output the mold box motion trajectory planning scheme and material maintenance parameters, including detailed time-position trajectory data, speed curve, stop node table, and maintenance parameter configuration for each node.

[0050] Taking the cement concrete mold box as an example, 15 candidate trajectories were subjected to curvature analysis, and points with curvature values ​​greater than 0.5 were identified as candidate nodes. Candidate stay points were added at locations where the speed was less than 0.1 m / s, and 8 key path nodes were obtained through comprehensive screening. The travel time between nodes was calculated based on the maximum equipment speed of 0.5 m / s, and the maintenance stay time was preliminarily estimated. A time series containing 4 main maintenance windows was constructed, with window lengths ranging from 30 to 60 minutes. Data was collected through temperature sensors, and combined with thermal model analysis, the temperature range of each window was determined: Window 1 was low-temperature initial maintenance of 25-35°C, Window 2 was standard maintenance of 45-55°C, Window 3 was high-temperature rapid maintenance of 60-70°C, and Window 4 was slow cooling maintenance of 40-50°C. Based on these temperature ranges, the stay time of each node was calculated: the node in Window 1 stayed for 45 minutes, the node in Window 2 stayed for 30 minutes, the node in Window 3 stayed for 20 minutes, and the node in Window 4 stayed for 40 minutes. The residence time is compared with the standard curing requirements of concrete, and the parameters are adjusted to ensure the curing effect, and a curing control table including temperature, time and heating power is generated. Finally, the trajectory is optimized according to the control table, the speed curve and the dwell position are adjusted, and a complete motion trajectory and curing plan with a total duration of 205 minutes is output to meet the concrete quality requirements while optimizing energy use.

[0051] In a specific embodiment, the process of performing the time segmentation processing step according to the critical path node sequence may specifically include the following steps: (1) Calculate the distance of the key path node sequence and generate a distance distribution table between nodes; (2) Divide the time window based on the distance distribution table between nodes to form an initial time segment sequence; (3) Evaluate the material status of each segment in the initial time segment sequence to obtain a material maintenance requirement table; (4) Adjust the window boundaries of the initial time segment sequence in combination with the material maintenance demand table to construct a time window boundary set; (5) Spatially map the time window boundary set with the key path node sequence to generate a node time attribute matrix; (6) Align the time windows according to the node time attribute matrix and construct the initial maintenance time window sequence.

[0052] Specifically, the critical path node sequence consists of important nodes extracted from the candidate trajectory set, including turning points, stop points and maintenance points during the movement of the mold box. The distance calculation uses the Euclidean distance formula to calculate the straight-line distance between adjacent nodes in three-dimensional space. For each pair of adjacent nodes, the coordinate differences in the three directions of x, y and z are calculated, and then the square root of the sum of squares is calculated to obtain the spatial straight-line distance. In addition to the spatial distance, the time interval between nodes is also calculated to record the difference in the expected time of node arrival. The spatial distance and time interval are combined to form an inter-node distance distribution table, which records the spatial distance, time interval and moving speed between each pair of adjacent nodes in the trajectory, providing a basis for subsequent time window division. Time window division based on the inter-node distance distribution table is the process of dividing a continuous trajectory into discrete time periods. Time window division uses an adaptive segmentation algorithm, which determines the segmentation point according to the change of the distance and moving speed between nodes. Set segmentation rules, such as distance threshold, speed change threshold or time interval threshold. When the distance between adjacent nodes exceeds the threshold, the speed change exceeds the threshold or the time interval exceeds the threshold, a window boundary is set between the two nodes. Considering the continuity requirement of material curing, the window length should not be too short, so the minimum window length constraint is introduced to ensure that each window contains enough curing time. According to these rules, the node sequence is divided to form an initial time segment sequence, and each segment corresponds to a period of time when the mold box is in a specific area or a specific state.

[0053] Material state assessment is based on material development feature vectors and environmental parameters to analyze the development state of materials in each time segment. Material feature data of the corresponding spatial area of ​​each segment is extracted, including parameters such as temperature, consistency, and hardness. Then, according to the material type and maintenance standard, a material state assessment model is established. The model considers the gap between the current state of the material and the target state and calculates the maintenance urgency index. Different types of materials have different maintenance requirements. For example, concrete needs to maintain appropriate humidity and temperature, and resin materials need to control the curing rate. According to the evaluation results, a material maintenance requirement table is generated, which records the maintenance type, maintenance intensity, and maintenance duration of the material in each time segment. Adjusting the window boundary of the initial time segment sequence in combination with the material maintenance requirement table is an important part of optimizing the maintenance time arrangement. The window boundary adjustment adopts a constrained optimization method to meet the material maintenance requirements as the goal, while considering time efficiency and energy consumption. The maintenance duration requirements in the material maintenance requirement table are analyzed to determine the shortest and longest maintenance time for each segment. Then adjust the boundaries of the initial time segment sequence, merge adjacent segments with similar maintenance requirements, and split segments with significant changes in maintenance requirements. For segments with high maintenance intensity, appropriately extend the window time; for segments with high maintenance urgency, adjust the window position to ensure timely maintenance. After multiple rounds of iterative optimization, a time window boundary set that meets the material maintenance requirements is constructed, and each boundary point marks the start or end time of the time window.

[0054] Spatial mapping uses an interpolation algorithm to map time points to specific locations on the spatial trajectory. Establish a corresponding relationship between time and path nodes to determine the spatial position of the mold box at each window boundary moment. For the case where the time window boundary falls between two key nodes, use linear interpolation or spline interpolation methods to calculate the corresponding spatial position. At the same time, consider the motion speed and acceleration constraints of the mold box to ensure that the mapping results conform to physical laws. Through spatial mapping, assign time attributes to each key path node, including the time window to which the node belongs, the relative position in the window, and the residence time, etc., to form a node time attribute matrix. Time window alignment is to convert discrete time attributes into a continuous time window sequence to ensure seamless connection between windows and meet maintenance requirements. Arrange all window boundaries in chronological order to form a complete time axis. Then map each key path node to a specific position on the time axis to determine the time window where the node is located. For the case where it is necessary to stay at the node for maintenance, calculate the start time and end time of the stay and mark it on the time axis. According to the time attributes of the node and the material maintenance requirements, adjust the start and end time of the window to ensure the continuity and effectiveness of the maintenance process. Construct an initial maintenance time window sequence, which contains a series of time windows, each of which defines the start time, end time, operation type and operation parameters of a specific maintenance operation.

[0055] Taking the maintenance of concrete mold boxes as an example, the eight key path nodes extracted from the optimal trajectory are calculated by Euclidean distance to obtain the distance between nodes. The distance ranges from 0.5 meters to 2.5 meters, and the time interval is between 30 seconds and 3 minutes. The time window is divided based on the distance threshold of 0.8 meters and the time interval threshold of 45 seconds to form 5 initial time segments. The material state is evaluated for each segment, and the curing requirement table is obtained considering the hardening degree, temperature and humidity of the concrete. The table records that segment 1 requires 30 minutes of standard temperature curing, segment 3 requires 45 minutes of high temperature curing, and other segments require conventional curing. The window boundary is adjusted according to the maintenance requirements, the original segment 2 and segment 3 are merged into the new segment 2, the high temperature curing time is extended to 60 minutes, and a boundary set containing 4 time windows is constructed. The window boundary time points are mapped to the spatial trajectory to obtain the spatial coordinates of the boundary points, forming a node time attribute matrix, which records the window to which each node belongs and the residence time. Finally, time windows are aligned to construct an initial maintenance time window sequence consisting of four time windows with window lengths of 30 minutes, 60 minutes, 45 minutes, and 40 minutes, respectively. The total maintenance time is 175 minutes.

[0056] In a specific embodiment, the process of performing the temperature field matching step for the initial maintenance time window sequence may specifically include the following steps: (1) Collect temperature data for each time window in the initial maintenance time window sequence to generate a window temperature sampling point set; (2) Calculate the temperature gradient distribution based on the window temperature sampling point set to form a temperature change trend graph; (3) Segment the temperature change trend graph according to the time window boundary to obtain a set of segmented temperature curves; (4) Perform temperature fluctuation analysis on the segmented temperature curve set and construct a temperature fluctuation range table; (5) Divide the segmented temperature curve set into intervals based on the temperature fluctuation range table to generate a temperature interval mapping table; (6) The temperature of each time window is matched through the temperature range mapping table to obtain the maintenance temperature range of each time window.

[0057] Specifically, temperature data is collected for each time window in the initial curing time window sequence to generate a window temperature sampling point set. Temperature data collection uses the temperature sensor array and environmental monitoring equipment arranged in the mold box to record temperature values ​​at a fixed frequency (usually 5-10 seconds / time) within each window time range. The collected data includes the measurement point location coordinates, timestamps and temperature values ​​to form a structured temperature sampling point set. For each time window, hundreds of temperature sampling points are usually collected to ensure the temporal and spatial coverage and representativeness of the data. The temperature gradient distribution is calculated based on the window temperature sampling point set to form a temperature change trend graph. The temperature gradient refers to the rate of change of temperature in space or time, reflecting the direction and intensity of heat transfer. The calculation process sorts the temperature sampling points in time series, and then calculates the ratio of the temperature difference between adjacent sampling points to the time difference to obtain the temperature gradient in the time dimension; the temperature gradient between adjacent points in space is calculated in the same way. The calculation results are visualized to generate a temperature change trend graph to intuitively display the temperature change characteristics over time and space. The trend chart takes the form of a heat map or curve chart, with the horizontal axis representing time, the vertical axis representing temperature or spatial position, and the color or curve height representing the temperature value or gradient value.

[0058] The temperature change trend graph is segmented according to the time window boundary to obtain a segmented temperature curve set. The time window boundary refers to the start and end time of each maintenance stage. The continuous temperature trend graph is segmented into multiple independent temperature curve segments according to these boundary points. The segmentation process retains the temperature value at the boundary point to ensure the continuity of the segmented data. The segmented temperature curve set contains multiple independent temperature-time curves, each of which corresponds to a specific maintenance stage and reflects the temperature change characteristics of the stage. Each segmented curve still retains the original sampling point information. Temperature fluctuation analysis is performed on the segmented temperature curve set to construct a temperature fluctuation range table. Temperature fluctuation analysis refers to quantitatively evaluating the stability and fluctuation characteristics of temperature changes to provide a basis for setting maintenance parameters. The analysis method includes calculating the statistical indicators (such as maximum value, minimum value, average value, standard deviation, etc.) and time series characteristics (such as fluctuation period, rise / fall rate, etc.) of each segment. The temperature fluctuation range table records the fluctuation characteristic data of each segment, including indicators such as average temperature, temperature range, fluctuation amplitude and fluctuation frequency. This tabular representation is convenient for subsequent processing and comparative analysis.

[0059] Combined with the temperature fluctuation range table, the segmented temperature curve set is divided into intervals to generate a temperature interval mapping table. Interval division is to divide the temperature range into multiple sub-intervals, which is used to finely distinguish the curing characteristics under different temperature conditions. The division method is based on the temperature sensitivity and curing requirements of the material, and divides the temperature range into multiple meaningful intervals (such as low temperature zone, suitable zone, high temperature zone, etc.). For each segmented curve, according to its temperature distribution characteristics, determine the temperature interval to which it belongs and the distribution ratio in each interval. The temperature interval mapping table records the mapping relationship between each time segment and the temperature interval, including information such as segment ID, corresponding temperature interval, interval range and suitability score. Through the temperature interval mapping table, the temperature of each time window is matched to obtain the curing temperature interval of each time window. Temperature matching is to determine the most suitable curing temperature range for each time window based on the material curing requirements and actual temperature conditions. The matching process analyzes the current development stage and curing requirements of the material to determine the ideal temperature conditions; then compares it with the actual temperature interval to evaluate the suitability; comprehensively considers the feasibility of temperature control and determines the curing temperature interval. The maintenance temperature range includes three key parameters: minimum temperature, maximum temperature and target temperature, providing precise guidance for subsequent temperature control and scheduling decisions.

[0060] In an application example of an aerated concrete production line, temperature data processing was performed on 8 initial curing time windows. The temperature data was collected using 12 temperature sensors distributed in the mold box. About 1,700 temperature data points were collected in 4 hours at a frequency of once every 10 seconds. After calculating the temperature gradient, it was found that the temperature of the third window (initial setting stage) showed a characteristic of first rapid rise and then stability, and the gradient value gradually decreased from the initial 0.2℃ / minute to 0.05℃ / minute; while the temperature of the fifth window (strength development stage) remained relatively stable, and the gradient value remained below 0.03℃ / minute. According to the boundary points of the curing stage, the temperature curve was divided into 8 segments, each corresponding to a time window. The temperature fluctuation analysis showed that the temperature range of the third window was 52-57℃, with an average value of 54.5℃ and a standard deviation of 1.2℃; the temperature range of the fifth window was 48-52℃, with an average value of 50.3℃ and a standard deviation of 0.8℃. Based on the material maintenance requirements, the temperature range is divided into four ranges: low temperature range (<45℃), initial setting suitable range (53-57℃), strength development suitable range (48-52℃) and high temperature range (>60℃). The temperature range mapping table shows that the third window is mainly distributed in the initial setting suitable range, accounting for about 85%; the fifth window is completely located in the strength development suitable range, accounting for 100%.

[0061] The determined curing temperature ranges are: the temperature range of the 3rd window (initial setting stage) is 53-57°C, with a target temperature of 55°C; the temperature range of the 5th window (strength development stage) is 48-52°C, with a target temperature of 50°C. These precise temperature ranges provide a basis for the intelligent scheduling of the mold box, guiding the mold box to move to the appropriate temperature environment at each curing stage, ensuring the best development conditions for aerated concrete materials, and improving the consistency and stability of product quality.

[0062] In a specific embodiment, the process of performing the temperature-time matching step of the node stay schedule and the material maintenance requirements may specifically include the following steps: (1) Normalize the time data in the node stay schedule to generate a standardized time series; (2) Establish a temperature-time correspondence table based on material maintenance requirements to form a maintenance standard matrix; (3) Map the standardized time series to the maintenance standard matrix to obtain the initial control parameter set; (4) Perform boundary constraint checks on the initial control parameter set and construct a parameter feasible domain table; (5) Optimize and screen the data in the parameter feasible domain table to generate a parameter optimization sequence; (6) Integrate and adjust the control parameters based on the parameter optimization sequence and construct a maintenance control parameter table.

[0063] Specifically, the time data in the node residence time table is normalized to generate a standardized time series. Normalization is the process of converting time data of different dimensions and ranges into a unified standard interval (usually 0-1) to facilitate subsequent comparison and matching. The processing method adopts maximum and minimum value normalization to map the original time value to the standard interval according to its relative position in the entire sequence. For the residence time of each node, its normalized value relative to the shortest and longest residence time is calculated to form a standardized time series. This processing eliminates the influence of the absolute value of time, highlights the relative relationship of the residence time of each node, and provides a unified measurement basis for subsequent matching with the maintenance standard. Based on the material maintenance requirements, a temperature-time correspondence table is established to form a maintenance standard matrix. The maintenance standard matrix is ​​a two-dimensional data structure that describes the temperature and time requirements of materials at different development stages. The construction process clarifies the key development stages of the material (such as the initial setting period, strength development period, and stable maturity period, etc.), and then determines the optimal temperature interval and required time interval for each stage based on the process requirements and experimental data. The rows of the curing standard matrix represent different development stages, the columns represent temperature and time parameters, and the cell values ​​represent the suitability or recommended values ​​of the parameters. This matrix converts discrete process requirements into a systematic set of parameters, providing a standard reference for subsequent matching.

[0064] The standardized time series is mapped to the maintenance standard matrix to obtain the initial control parameter set. Data mapping is the process of matching the actual residence time with the standard maintenance requirements. The mapping method searches for the corresponding temperature and other maintenance parameters from the maintenance standard matrix based on the development stage and standardized time value of the node. For each node, the matching temperature value and control parameter are determined according to its current development stage and standardized residence time. The initial control parameter set contains the target temperature, temperature range, humidity requirements and other maintenance control parameters of each node, providing a preliminary plan for subsequent adjustments. Boundary constraint checks are performed on the initial control parameter set to construct a parameter feasible domain table. Boundary constraint checks are the process of verifying whether the control parameters meet physical constraints and process limitations. The inspection content includes whether the temperature range exceeds the equipment capacity, whether the parameter changes between adjacent nodes are too large, and whether the total maintenance time meets the requirements. The parameters that do not meet the constraints are adjusted to fall into the feasible range. The parameter feasible domain table records the feasible range and adjusted values ​​of each parameter, including fields such as parameter name, minimum value, maximum value, initial value and adjusted value. This table clarifies the boundary conditions for parameter optimization to ensure that subsequent optimization results are practical and operational.

[0065] The data in the parameter feasible domain table is optimized and screened to generate a parameter optimization sequence. Optimization screening is the process of finding the optimal parameter combination in the parameter feasible domain according to a specific objective function. The optimization objectives usually include maximizing product quality, minimizing energy consumption, or maximizing production efficiency. The optimization method uses techniques such as gradient descent or genetic algorithm to search for the optimal solution in the parameter space. The control parameters of each node are optimized multiple times in its feasible domain until the parameter value that meets the optimization objective is found. The parameter optimization sequence records the parameter evolution path during the optimization process, including information such as the number of iterations, parameter values, and objective function values, and shows the convergence process of parameter optimization. Based on the parameter optimization sequence, the control parameters are integrated and adjusted to construct a maintenance control parameter table. Integrated adjustment is to systematically adjust and coordinate the optimized parameters to ensure the consistency and coherence between the parameters of each node. Adjustment methods include smoothing, trend unification, and conflict elimination. The results of the optimization sequence are integrated to ensure smooth parameter changes between adjacent nodes and a coherent and orderly overall maintenance process. The maintenance control parameter table is the control scheme, which contains precise control parameters of each node, such as temperature setting value, allowable fluctuation range, control priority and other information, which directly guides the temperature control and motion scheduling of the mold box.

[0066] In the application case of an aerated concrete production line, the residence time of 12 key nodes was processed. The original residence time ranged from 20 minutes to 90 minutes, and after normalization, a standardized time series between 0 and 1 was generated. For example, the node with an original residence time of 60 minutes has a standardized value of 0.57; the node with a standardized value of 90 minutes has a standardized value of 1.0. Based on the material process requirements, a 3×4 curing standard matrix was established, with rows representing the three stages of initial setting period, strength development period, and stable maturity period, and columns representing four parameters: temperature range, target temperature, minimum time, and maximum time. The standardized time series was mapped with the curing standard matrix to generate an initial control parameter set. For example, the initial control temperature of the 3rd node (initial setting period, standardized time 0.57) was mapped to 55°C, with a range of 53-57°C; the initial control temperature of the 8th node (strength development period, standardized time 0.94) was mapped to 50°C, with a range of 48-52°C. The boundary constraint check found that the temperature setting of the 5th node exceeded the upper limit of the equipment capability, so it was adjusted from 58°C to 56°C; the temperature changes between the 10th and 11th nodes were too large, so the transition parameters were adjusted to reduce the temperature jump.

[0067] During the optimization and screening stage, the parameters were optimized for multiple rounds of iterations with the goal of maximizing product quality. For example, the temperature of the 3rd node was optimized from the initial value of 55°C to 54.5°C, and the 8th node was adjusted from 50°C to 49.5°C to better match the development characteristics of the material. After integration and adjustment, a complete curing control parameter table was constructed to clarify the precise control parameters of each node. For example, the control parameters of the 3rd node are: temperature 54.5±1.5°C, residence time 65 minutes, humidity 65%; the 8th node is: temperature 49.5±1.0°C, residence time 85 minutes, humidity 60%. These precise control parameters ensure that aerated concrete can obtain the best curing conditions at each development stage, significantly improving the consistency and stability of product quality.

Claims

1. A method for intelligent monitoring of mold box materials based on machine vision, characterized in that: The machine vision-based intelligent monitoring method for mold box materials includes: The image of the material in the mold box and the surrounding environment is collected through a multi-dimensional visual acquisition network to generate a material image dataset and an environmental obstacle image dataset; Based on the material image data set and combined with the temperature sensor data, the improved Gaussian mixture background modeling algorithm is used to extract features to obtain the material consistency parameter matrix, hardness parameter matrix and development state parameter matrix; The material consistency parameter matrix, hardness parameter matrix and development state parameter matrix are input into the improved LSTM-Attention network for multimodal data fusion. At the same time, the environmental obstacle image dataset is combined to generate the material development feature vector and the mold box motion feature vector. According to the material development feature vector, an intelligent heating control strategy is constructed to adaptively adjust the insulation system and auxiliary heating system of the mold box to form a temperature control instruction set; An intelligent scheduling model is established based on the mold box motion feature vector and temperature control instruction set, and the mold box motion trajectory planning scheme and material maintenance parameters are output to realize unmanned intelligent scheduling of the mold box.

2. The method for intelligently monitoring mold box materials based on machine vision according to claim 1, characterized in that: The method of collecting images of the material in the mold box and the surrounding environment through a multi-dimensional visual collection network to generate a material image dataset and an environmental obstacle image dataset includes: Use a depth camera to scan the internal space of the mold box, organize the acquired point cloud data according to the spatial coordinate sequence, and obtain three-dimensional point cloud sequence data; Perform spatial registration and calibration on the 3D point cloud sequence data, separate the material area from the non-material area, and generate a material spatial distribution map; Through the material spatial distribution map, the material surface contour point set is extracted to establish the material surface height field data; The material surface height field data is fused and mapped with the time series image to generate a material dynamic change sequence; Combined with the temperature field data collected by the infrared sensor, the thermal distribution of the material dynamic change sequence is mapped to obtain the material temperature distribution characteristics; The material temperature distribution characteristics and the environmental scanning data are divided into regions to generate a material image data set and an environmental obstacle image data set respectively.

3. The method for intelligently monitoring mold box materials based on machine vision according to claim 1, characterized in that: The method uses an improved Gaussian mixture background modeling algorithm to extract features based on the material image data set and in combination with the temperature sensor data to obtain a material consistency parameter matrix, a hardness parameter matrix and a development state parameter matrix, including: Perform multi-scale image segmentation on the material image dataset and extract the material surface texture feature point set; Combined with the temperature sensing data, the material surface texture feature point set is thermally mapped to obtain the material temperature distribution matrix; The background layer of the material temperature distribution matrix is ​​separated by improving the Gaussian mixture background modeling algorithm to generate a material surface state feature map; Based on the material surface state characteristic diagram, the material consistency evaluation function is constructed and the material consistency parameter matrix is ​​calculated; Perform gradient calculation on the material surface state characteristic diagram, establish the material surface hardness evaluation criteria, and obtain the hardness parameter matrix; The material temperature distribution matrix and the material surface state characteristic map are combined, and data mapping is performed according to the material development standard to obtain the material consistency parameter matrix, hardness parameter matrix and development state parameter matrix.

4. The method for intelligently monitoring mold box materials based on machine vision according to claim 1, characterized in that: The material consistency parameter matrix, hardness parameter matrix and development state parameter matrix are input into the improved LSTM-Attention network for multimodal data fusion, and the material development feature vector and the mold box motion feature vector are generated by combining the environmental obstacle image data set, including: Perform time series alignment and calibration on the material consistency parameter matrix, hardness parameter matrix and development state parameter matrix to obtain material characteristic time series data; The material feature time series data is input into the time series encoding layer of the improved LSTM-Attention network for time step decomposition to generate a material state evolution sequence; The feature attention layer of the improved LSTM-Attention network is used to calculate the importance weight of the material state evolution sequence and extract the key time series features of the material; The environmental obstacle image dataset is input into the spatial encoding layer of the improved LSTM-Attention network to generate an environmental constraint feature map; The multimodal fusion layer of the improved LSTM-Attention network performs feature mapping on the key time series features of the materials and the environmental constraint feature graph to obtain a fusion feature matrix; The fused feature matrix is ​​input into the feature decoding layer of the improved LSTM-Attention network for feature decoupling to generate the material development feature vector and the mold box movement feature vector.

5. The method for intelligently monitoring mold box materials based on machine vision according to claim 1, characterized in that: The intelligent heating control strategy is constructed according to the material development feature vector, and the insulation system and auxiliary heating system of the mold box are adaptively adjusted to form a temperature control instruction set, including: Perform heating demand analysis based on material development feature vectors and generate a heating area distribution map; Perform temperature field mapping on the heating area distribution map to obtain the target temperature control point sequence; Allocate power to the insulation units of the insulation system according to the target temperature control point sequence and construct an insulation power curve; Calculate the temperature gradient based on the insulation power curve and establish an auxiliary heating power compensation table; Combine the insulation power curve with the auxiliary heating power compensation table to generate a heating control matrix; The control parameters of the insulation system and the auxiliary heating system are dynamically adjusted according to the heating control matrix to form a temperature control instruction set.

6. The method for intelligently monitoring mold box materials based on machine vision according to claim 1, characterized in that: The intelligent scheduling model is established based on the mold box motion feature vector and the temperature control instruction set, and the mold box motion trajectory planning scheme and material maintenance parameters are output to realize unmanned intelligent scheduling of the mold box, including: Decompose the mold box motion feature vector according to the time sequence to generate the mold box position change sequence, speed change sequence and acceleration change sequence; A three-dimensional thermal distribution map is established according to the temperature control instruction set, and the thermal distribution map is sliced ​​in time and space to obtain the temperature field constraint conditions at each time point; The position change sequence, velocity change sequence and acceleration change sequence are combined to construct a dynamic motion state space, and the temperature field constraint conditions are mapped into the motion state space; Perform collision detection on the motion state space, generate a safe obstacle avoidance area, and form a comprehensive constraint space in combination with the temperature field constraint conditions; Construct a trajectory cost function based on the comprehensive constraint space, optimize the path sampling of the cost function, and generate a candidate trajectory set that meets the temperature requirements and obstacle avoidance requirements; The key path nodes are extracted from the candidate trajectory set, the maintenance time window sequence is established, the node information is matched with the maintenance requirements, and the mold box motion trajectory planning scheme and material maintenance parameters are output.

7. The method for intelligently monitoring mold box materials based on machine vision according to claim 6, characterized in that: The key path nodes are extracted from the candidate trajectory set, a maintenance time window sequence is established, and the node information is matched with the maintenance requirements, and the mold box motion trajectory planning scheme and material maintenance parameters are output, including: Perform path fitting analysis on the candidate trajectory set and extract the key path node sequence through the curvature extreme points; Perform time segmentation processing according to the key path node sequence and construct the initial maintenance time window sequence; Perform temperature field matching on the initial curing time window sequence to obtain the curing temperature range for each time window; Calculate the dwell time of the key path node sequence based on the maintenance temperature range and generate a node dwell schedule; Match the node stay schedule with the material maintenance requirements in terms of temperature and time, and build a maintenance control parameter table; The trajectory of the key path node sequence is optimized according to the maintenance control parameter table, and the mold box motion trajectory planning scheme and material maintenance parameters are output.

8. The method for intelligently monitoring mold box materials based on machine vision according to claim 7, characterized in that: The time segmentation process is performed according to the critical path node sequence to construct the initial maintenance time window sequence, including: Calculate the distance of the key path node sequence and generate a distance distribution table between nodes; Divide the time window based on the distance distribution table between nodes to form an initial time segment sequence; Evaluate the material status for each segment in the initial time segment sequence to obtain a material maintenance requirement table; Combined with the material maintenance demand table, the window boundaries of the initial time segment sequence are adjusted to construct a time window boundary set; Spatially map the time window boundary set with the key path node sequence to generate a node time attribute matrix; The time windows are aligned according to the node time attribute matrix to construct the initial maintenance time window sequence.

9. The method for intelligently monitoring mold box materials based on machine vision according to claim 7, characterized in that: The temperature field matching is performed on the initial curing time window sequence to obtain the curing temperature range of each time window, including: Collect temperature data for each time window in the initial maintenance time window sequence to generate a window temperature sampling point set; Calculate the temperature gradient distribution based on the window temperature sampling point set to form a temperature change trend graph; The temperature change trend graph is segmented according to the time window boundary to obtain a segmented temperature curve set; Conduct temperature fluctuation analysis on the segmented temperature curve set and construct a temperature fluctuation range table; The segmented temperature curve set is divided into intervals in combination with the temperature fluctuation range table to generate a temperature interval mapping table; The temperature of each time window is matched through the temperature range mapping table to obtain the maintenance temperature range of each time window.

10. The method for intelligently monitoring mold box materials based on machine vision according to claim 7, characterized in that: The node stay schedule is matched with the material maintenance requirements in terms of temperature and time, and a maintenance control parameter table is constructed, including: Normalize the time data in the node stay schedule to generate a standardized time series; Establish a temperature-time correspondence table based on material maintenance requirements to form a maintenance standard matrix; Map the standardized time series to the maintenance standard matrix to obtain the initial control parameter set; Perform boundary constraint checks on the initial control parameter set and construct a parameter feasible domain table; Optimize and screen the data in the parameter feasible domain table to generate a parameter optimization sequence; The control parameters are integrated and adjusted based on the parameter optimization sequence to construct a maintenance control parameter table.

Images