Metal powder grinding temperature field online monitoring method and system

Abstract

The application relates to the technical field of temperature monitoring, and discloses a metal powder grinding temperature field online monitoring method and system, the method comprising the following steps: acquiring dust concentration and temperature data, obtaining temperature field data through convolution and correlation analysis; extracting local fluctuation values, screening temperature abnormal regions, and generating a temperature abnormality map. Through time series analysis, a temperature diffusion sequence is obtained, similarity analysis is carried out in combination with historical dust distribution, a temperature deviation value is calculated, and gradient descent optimization is adopted, so that an environment monitoring model is optimized. The optimized model is used to simulate the interaction between dust and temperature, the coordinates of the abnormal region are located, and the preliminary dust distribution map and the temperature abnormality map are updated, so that a closed-loop monitoring result is formed. The application can realize accurate online monitoring of the metal powder grinding temperature field under the condition of dynamic high-concentration dust interference.

Description

Technical Field

[0001] This invention relates to the field of temperature monitoring technology, and in particular to a method and system for online monitoring of the temperature field in metal powder grinding. Background Technology

[0002] Currently, in the metal powder grinding process, real-time monitoring of the temperature field inside the grinding chamber directly affects the finished product quality and the safe operation of the equipment. Because the grinding process involves intense mechanical friction and high concentrations of dust, accurately sensing the evolution trend of the actual temperature field under dust interference is crucial for preventing dust explosions and equipment damage.

[0003] In existing technologies, temperature monitoring in metal powder grinding often employs single-point temperature probes based on fixed thresholds or simple sensor networks, relying primarily on static analysis of absolute surface temperature values ​​to trigger alarms. When sudden conditions occur inside the grinding equipment, such as localized material accumulation or dry friction of the grinding media, the distribution of high-concentration metal dust inside undergoes drastic dynamic changes, significantly interfering with heat radiation and convection. Traditional static monitoring strategies, lacking in-depth analysis of the multimodal data coupling and interaction between dust and temperature, cannot promptly penetrate dust obstructions to capture early localized heat accumulation, leading to frequent missed alarms or severely delayed alarms.

[0004] In summary, existing technologies struggle to achieve accurate online monitoring of the temperature field during metal powder grinding under dynamic, high-concentration dust interference conditions, resulting in low monitoring accuracy. Summary of the Invention

[0005] This invention provides a method and system for online monitoring of the temperature field of metal powder grinding, so as to achieve accurate online monitoring of the temperature field of metal powder grinding under dynamic high-concentration dust interference conditions.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a method for online monitoring of the temperature field during metal powder grinding, comprising: Obtain dust concentration data and temperature data, and perform convolution processing on the dust concentration data to obtain a preliminary dust distribution map; Based on the temperature data, a temperature interference correlation analysis was performed on the preliminary dust distribution map to obtain temperature field data; Local fluctuation values ​​are extracted from the temperature field data, temperature anomaly areas with local fluctuation values ​​greater than a preset fluctuation threshold are filtered out, and the temperature anomaly areas are summarized to obtain a temperature anomaly map. A time-series analysis is performed on the temperature anomaly map to obtain a temperature diffusion sequence, historical dust distribution is obtained, a similarity analysis is performed on the preliminary dust distribution map and the historical dust distribution to obtain a similarity sequence, and a deviation calculation is performed on the temperature diffusion sequence and the similarity sequence to obtain a temperature deviation value. The temperature deviation value is optimized by gradient descent, and the pre-established environmental monitoring model is also optimized to obtain an optimized environmental model. The interaction between dust and temperature is simulated using the optimized environmental model to locate abnormal areas during the interaction process and obtain the coordinates of the abnormal areas. Based on the coordinates of the abnormal area, the preliminary dust distribution map and the temperature anomaly map are updated to obtain closed-loop monitoring results.

[0007] In one optional implementation, the step of acquiring dust concentration data and temperature data, and performing convolution processing on the dust concentration data to obtain a preliminary dust distribution map includes: Acquire dust concentration and temperature data collected by the sensor array; The dust concentration data is processed into a grid to obtain a two-dimensional distribution matrix; The particle distribution features and density gradient features of the two-dimensional distribution matrix are extracted through multi-layer convolution operations. The particle distribution characteristics and the density gradient characteristics are fused to obtain a preliminary dust distribution map.

[0008] In one optional implementation, the step of performing temperature interference correlation analysis on the preliminary dust distribution map based on the temperature data to obtain temperature field data includes: The preliminary dust distribution map is divided into multiple cells, and the dust concentration difference between adjacent cells is calculated. The cells where the dust concentration difference exceeds a preset gradient threshold are marked as boundary points, and the boundary points are summarized to obtain the uneven diffusion region. The dust distribution within the uneven diffusion region is clustered and divided using the K-means clustering algorithm to obtain dust accumulation areas; Spatial correlation analysis was performed on the dust accumulation area and the temperature data to obtain temperature field data.

[0009] In one optional implementation, the step of extracting local fluctuation values ​​from the temperature field data, filtering temperature anomaly areas where the local fluctuation values ​​are greater than a preset fluctuation threshold, and summarizing the temperature anomaly areas to obtain a temperature anomaly map includes: The temperature field data is divided into spatial grids, and the temperature change rate of each grid cell is calculated. The standard deviation of the temperature change rate is calculated to obtain the local fluctuation value. If the local fluctuation value is greater than the preset fluctuation threshold, it is marked as a potential anomaly. The spatial coordinates of the potential anomaly point are spatially superimposed with the spatial coordinates of the dust accumulation area to calculate the spatial proximity distance. If the spatial proximity distance is less than the preset proximity radius, the potential anomaly point is determined to be a temperature anomaly area. The local fluctuation values ​​of the temperature anomaly area and the dust concentration data are subjected to heat map visualization processing to generate a temperature anomaly map.

[0010] In one optional implementation, the step of performing time-series analysis on the temperature anomaly map to obtain a temperature diffusion sequence, acquiring historical dust distribution, performing similarity analysis on the preliminary dust distribution map and the historical dust distribution to obtain a similarity sequence, and calculating the deviation between the temperature diffusion sequence and the similarity sequence to obtain a temperature deviation value includes: Extract the temperature change trajectory of each grid cell in the temperature anomaly map over a continuous time period, and perform Markov weighted network propagation on the temperature change trajectory to obtain the temperature diffusion sequence. Historical dust distribution data is retrieved, and the historical dust distribution data is dynamically time-normalized and compared with the preliminary dust distribution map to obtain a similarity sequence. The energy decay value of the temperature diffusion sequence is calculated to obtain the temperature energy decay characteristics, and the change rate of the similarity sequence is calculated to obtain the dust distribution change characteristics. The temperature energy decay characteristic and the dust distribution change characteristic are fused by multiple linear weighted fusion to obtain the temperature deviation value.

[0011] In one optional implementation, the step of optimizing the temperature deviation value through gradient descent and optimizing the pre-established environmental monitoring model to obtain an optimized environmental model includes: The temperature deviation value is iteratively optimized using the gradient descent algorithm to obtain the optimized temperature deviation value. Based on the analysis of historical fault data obtained in advance, the relationship between equipment operating temperature and fault probability is analyzed to obtain the critical value of the equipment's potential hazard temperature. The optimized environmental monitoring model is optimized by backpropagation based on the optimized temperature deviation value and the critical temperature value of the potential hazard, thus obtaining the optimized environmental model.

[0012] In one optional implementation, the step of simulating the interaction between dust and temperature using the optimized environmental model, locating abnormal regions during the interaction process, and obtaining the coordinates of the abnormal regions includes: The interaction between the dust concentration data and the temperature data is analyzed using the optimized environmental model to obtain the probability of the interaction. Areas where the probability of interaction impact is greater than a preset impact threshold are identified as high-risk areas; Extract the center coordinates of the high-risk area to obtain the coordinates of the abnormal area.

[0013] In one optional implementation, updating the preliminary dust distribution map and the temperature anomaly map based on the coordinates of the abnormal area to obtain closed-loop monitoring results includes: Spatial smoothing is performed on the dust concentration data in the abnormal area to obtain an updated dust distribution map. The temperature data of the anomalous region is processed using kernel density estimation to obtain an updated temperature anomaly map; The updated dust distribution map and the updated temperature anomaly map are subjected to time-series analysis with a preset historical database to obtain the diffusion trend of the anomaly area; Based on the spread trend of the abnormal area, a preset dynamic adjustment mechanism is triggered to obtain closed-loop monitoring results.

[0014] Secondly, the present invention provides an online monitoring system for the temperature field of metal powder grinding, comprising: The data acquisition module is used to acquire dust concentration data and temperature data, and to perform convolution processing on the dust concentration data to obtain a preliminary dust distribution map; The temperature analysis module is used to perform temperature interference correlation analysis on the preliminary dust distribution map based on the temperature data to obtain temperature field data. Anomaly identification module is used to extract local fluctuation values ​​from the temperature field data, filter temperature anomaly areas where the local fluctuation values ​​are greater than a preset fluctuation threshold, summarize the temperature anomaly areas, and obtain a temperature anomaly map. The deviation calculation module is used to perform time-series analysis on the temperature anomaly map to obtain a temperature diffusion sequence, acquire historical dust distribution, perform similarity analysis on the preliminary dust distribution map and the historical dust distribution to obtain a similarity sequence, and perform deviation calculation on the temperature diffusion sequence and the similarity sequence to obtain a temperature deviation value. The model optimization module is used to optimize the temperature deviation value through gradient descent and optimize the pre-established environmental monitoring model to obtain an optimized environmental model. The coordinate output module is used to simulate the interaction between dust and temperature through the optimized environment model, locate abnormal areas in the interaction process, and obtain the coordinates of the abnormal areas. The closed-loop monitoring module is used to update the preliminary dust distribution map and the temperature anomaly map based on the coordinates of the abnormal area to obtain the closed-loop monitoring results.

[0015] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention obtains dust concentration data and temperature data, performs convolution processing on the dust concentration data to obtain a preliminary dust distribution map, and performs temperature interference correlation analysis based on the temperature data to obtain temperature field data. This series of operations effectively quantifies the degree of physical interference of high-concentration metal dust on heat radiation and convection, constructs a real thermodynamic expression that integrates the spatial characteristics of dust, thereby solving the problem that simple temperature measurement methods in the prior art cannot adapt to high-concentration dynamic dust environments, and improving the realism of temperature perception under complex working conditions.

[0016] (2) After obtaining the temperature field data, this invention extracts local fluctuation values ​​to screen temperature anomaly areas and generates a temperature anomaly map. At the same time, it performs time-series analysis on the anomaly map to obtain the temperature diffusion sequence, and calculates the temperature deviation value by combining it with the similarity sequence of historical dust distribution. Through this mechanism of introducing spatiotemporal dynamic evolution and historical benchmark comparison, it can accurately isolate reasonable heat fluctuations in the normal grinding process, identify local thermal anomalies that are truly caused by equipment failure in real time, and greatly improve the accuracy and sensitivity of temperature monitoring in responding to sudden failures.

[0017] (3) This invention optimizes the temperature deviation value through the gradient descent algorithm, fine-tunes the pre-established environmental monitoring model accordingly, and uses the optimized model to simulate the interaction between dust and temperature to locate the coordinates of abnormal areas. This process makes full use of the advantages of deep learning models in handling multivariate nonlinear coupling relationships, and realizes accurate spatial positioning of potential heat sources under complex turbulence and heavy dust cover, significantly improving the efficiency and spatial positioning accuracy of potential hazards in grinding equipment.

[0018] (4) Based on the coordinates of the abnormal area, this invention updates the preliminary dust distribution map and temperature anomaly map in real time, thereby obtaining closed-loop monitoring results. This closed-loop monitoring mechanism not only realizes full-process traceability from multimodal data acquisition and anomaly detection to precise positioning, but also continuously corrects the underlying monitoring view based on the positioning results, forming a dynamic and adaptive online monitoring capability. This effectively solves the problem of lack of linkage feedback in single alarms in traditional methods, and improves the inherent safety and self-healing capability of the metal powder grinding production process. Attached Figure Description

[0019] Figure 1 This is a schematic flowchart of an embodiment of an online monitoring method for the grinding temperature field of metal powder provided by the present invention; Figure 2 This is a schematic flowchart of an embodiment of the present invention for obtaining temperature deviation values; Figure 3This is a schematic diagram of an embodiment of an online monitoring system for the temperature field of metal powder grinding provided by the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Reference Figure 1 The first embodiment of the present invention provides a method for online monitoring of the temperature field during metal powder grinding, comprising the following steps: Step S11: Obtain dust concentration data and temperature data, and perform convolution processing on the dust concentration data to obtain a preliminary dust distribution map; Step S12: Perform temperature interference correlation analysis on the preliminary dust distribution map based on the temperature data to obtain temperature field data; Step S13: Extract local fluctuation values ​​from the temperature field data, filter out temperature anomaly areas where the local fluctuation values ​​are greater than a preset fluctuation threshold, summarize the temperature anomaly areas, and obtain a temperature anomaly map. Step S14: Perform time-series analysis on the temperature anomaly map to obtain a temperature diffusion sequence, acquire historical dust distribution, perform similarity analysis on the preliminary dust distribution map and the historical dust distribution to obtain a similarity sequence, and calculate the deviation between the temperature diffusion sequence and the similarity sequence to obtain a temperature deviation value. Step S15: Optimize the temperature deviation value through gradient descent and optimize the pre-established environmental monitoring model to obtain an optimized environmental model; Step S16: Simulate the interaction between dust and temperature using the optimized environmental model, locate abnormal areas in the interaction process, and obtain the coordinates of the abnormal areas; Step S17: Update the preliminary dust distribution map and the temperature anomaly map according to the coordinates of the abnormal area to obtain the closed-loop monitoring results.

[0022] In step S11, dust concentration data and temperature data are acquired, and the dust concentration data is convolved to obtain a preliminary dust distribution map. The specific implementation method is as follows: First, dust concentration and temperature data are acquired from a sensor array. Specifically, a high-precision sensor array is pre-deployed within the working chamber and around the exhaust vent of the metal powder grinding equipment (such as a ball mill or sand mill). This array consists of a cross arrangement of laser scattering dust sensors and infrared thermal imaging sensors. The laser scattering dust sensors can capture the scattered light intensity of suspended metal dust in the space in real time. Based on a pre-calibrated mapping relationship between scattered light intensity and dust mass concentration, the scattered light intensity is converted into a dust mass concentration value. The infrared thermal imaging sensors capture the infrared radiation energy at various physical points within the grinding area. Based on a pre-calibrated mapping relationship between infrared radiation energy and temperature, the infrared radiation energy is converted into a corresponding absolute temperature value. It should be noted that, in a sealed experimental chamber, standard dust of known mass concentration is introduced, and the scattered light intensity signal output by the sensors is simultaneously collected. By setting no fewer than five concentration gradient points (covering the operating range of the equipment), a least squares method is used to perform polynomial regression fitting to establish a fitting curve of scattered light intensity-mass concentration, thus obtaining the mapping relationship between scattered light intensity and dust mass concentration. A high-precision blackbody radiation source is used, with the blackbody set as multiple known temperature points (covering the sensor range). Under a stable radiation environment, the grayscale value or radiation charge number output by the detector is collected. Based on Planck's radiation law and Stefan-Boltzmann's law, a mapping relationship between radiation energy and absolute temperature is established through nonlinear fitting of polynomial functions. Driven by a unified hardware clock source, these multimodal sensors synchronously collect the real-time status of the workspace at an extremely high sampling frequency. The acquired one-dimensional and two-dimensional discrete data are transmitted in real time to the host computer system via industrial Ethernet, forming the original dust concentration data and temperature data stream.

[0023] In one specific embodiment, four sets of detection rings are arranged axially within the grinding chamber. Each set of detection rings uniformly houses four laser scattering dust sensors. One infrared thermal imaging sensor is located at both the exhaust inlet and outlet. The axial spacing between adjacent detection rings is 8 cm to 12 cm, and the circumferential angle between adjacent dust sensors within the same detection ring is 90°. The sensor probes are isolated from the grinding chamber through a wear-resistant quartz window. A compressed air purging port is located outside the window, with a purging pressure of 0.05 MPa to 0.1 MPa to reduce the impact of metal dust adhesion on the measurement. For example, the scattered light intensity I is fitted with a quadratic polynomial, with coefficients such as a0 = 0.02, a1 = 0.86, and a2 = 0.015; the infrared radiation response R is fitted with a quadratic polynomial, with coefficients such as b0 = 18.6, b1 = 0.74, and b2 = 0.006. The above coefficients were recalibrated using standard dust and blackbody radiation sources after equipment installation and then fixed into the host computer parameter table.

[0024] Subsequently, the dust concentration data is gridded to obtain a two-dimensional distribution matrix. Specifically, to transform the discretely distributed sensor node data in space into a continuous spatial representation that can be processed by a deep learning model, a virtual two-dimensional grid coordinate system that is mapped to the actual physical space inside the grinder at a fixed ratio is first constructed. The physical installation coordinates of each laser scattering dust sensor are extracted, and the real-time dust concentration values ​​collected by each sensor are mapped to the corresponding nodes of the virtual grid. For the blank areas of the grid where no sensors are installed, an inverse distance weighted interpolation algorithm is used to fill the values. Specifically, the Euclidean distance between the blank grid point and the surrounding known sensor nodes is used as the basis for weight calculation, and the distance attenuation parameter is set to 2, i.e., the weight function is constructed using an inverse square attenuation method. This interpolation method calculates the positional distance from the blank grid point to the surrounding known sensor nodes, assigning higher weights to nodes that are closer and lower weights to nodes that are farther away, and then performing a weighted summation calculation on the concentration values ​​of the known nodes to estimate the dust concentration at the blank grid point. After this spatial interpolation smoothing process, the originally sparse discrete sensor data is transformed into a tightly packed two-dimensional distribution matrix with a fixed number of rows and columns. Each element of this matrix represents the dust concentration in a small area of ​​the corresponding physical space.

[0025] Subsequently, the particle distribution features and density gradient features of the two-dimensional distribution matrix are extracted through multi-layer convolution operations. It should be noted that a pre-trained convolutional neural network model is used for feature extraction. This convolutional neural network model employs a three-layer cascaded two-dimensional convolutional structure. The obtained two-dimensional distribution matrix is ​​used as the input to the first convolutional layer, and multiple 3x3 convolutional kernels are used to perform sliding multiplication and addition operations on the matrix. The first convolutional layer is mainly used to capture the most basic concentration variation edges between adjacent grids, outputting a basic feature map. The first convolutional layer is set to have 32 3x3 convolutional kernels, corresponding to 32 output channels. Subsequently, the basic feature map enters the second convolutional layer, using 64 5x5 convolutional kernels to extract the aggregation morphology of dust in local areas with a larger receptive field, forming particle distribution features. Next, the particle distribution features are input into the third convolutional layer, which is configured with 64 orientation-sensitive convolutional kernels. A directional Sobel operator is used as the basic weight structure for the kernels, specifically designed to calculate the rate of change of the first derivative of dust concentration in the horizontal and vertical directions. This deeply reflects the dramatic spatial variation trend of dust concentration from high to low or from low to high, extracting density gradient features. A Rectified Linear Unit (ReLU) function is applied after each convolutional operation to introduce nonlinear expressive power.

[0026] The pre-training process of the convolutional neural network model is supervised learning. The training data comes from no fewer than 500 sets of metal dust diffusion samples synchronously collected in the laboratory using a high-speed camera and a high-density sensor array. The labeling criteria for dust aggregation morphology regions are grid sets where the dust concentration value is higher than 1.5 times the global mean in consecutive time frames and forms spatially connected regions. The labeling criteria for concentration change boundaries are the set of locations where the concentration gradient between adjacent grids exceeds a preset gradient threshold (mean plus 3 times the standard deviation). The labeling process combines the dust optical density distribution in the high-speed images with the sensor's measured concentration data, and is completed by cross-validation by at least two domain experts after spatial registration. Specifically, the training process uses historical two-dimensional dust distribution matrices as input samples and expert-labeled real dust aggregation morphology regions and concentration change boundaries as target labels. The input data undergoes min-max normalization to map the values ​​to the interval between zero and one. The mean squared error is used as the loss function, and the adaptive moment estimation (Adam) optimizer is used with an initial learning rate of 0.001. During training, a five-fold cross-validation method is used, and the weight coefficients of each convolutional kernel are continuously updated through the backpropagation algorithm. The maximum number of training rounds is set to 200. When the mean squared error loss value of the model on the validation set fluctuates within a range of less than 0.001 within 15 consecutive iterations and no longer reaches a new low, an early stopping mechanism is triggered. Finally, the network parameters with fixed weights are saved as the optimal parameters of the model used in this embodiment.

[0027] The sampling duration for each group of metal dust diffusion samples was 30 seconds, the sampling frequency was 10 Hz, and a uniform interpolation of 64×64 two-dimensional spatial grids was used. The training set, validation set, and test set were divided in a 7:2:1 ratio. The global mean refers to the arithmetic mean of the concentration values ​​of all grids within the current single normalized two-dimensional dust distribution matrix, and the historical mean across batches was not used as the labeling threshold. The specific structure of the convolutional neural network is as follows: the first layer has 32 convolutional kernels, a kernel size of 3×3, a stride of 1, and padding of 1, with an output dimension of 64×64×32; the second layer has 64 convolutional kernels, a kernel size of 5×5, a stride of 1, and padding of 2, with an output dimension of 64×64×64; the third layer uses 64 orientation-sensitive 3×3 convolutional kernels, a stride of 1, and padding of 1, with an output dimension of 64×64×64. No pooling layers are set between the layers to maintain a one-to-one correspondence between the grid coordinates and the physical space.

[0028] Finally, the particle distribution features and density gradient features are fused to obtain a preliminary dust distribution map. Specifically, the particle distribution feature matrix and density gradient feature matrix output from the above convolution operation are extracted. Since these two feature matrices originate from the same-dimensional convolution operation of the same original input, they have the same spatial resolution. The system uses channel concatenation technology to merge these two matrices in the depth dimension, forming a dual-channel comprehensive feature tensor. Subsequently, a dimensionality-reducing convolutional layer with a kernel size of 1x1 is used to linearly combine the channels of this dual-channel feature tensor. By assigning different adaptive learning weights to the particle distribution features and density gradient features, the dual channels are compressed back into a single-channel two-dimensional matrix. This operation preserves the spatial static distribution contour of the dust while highlighting the dynamic evolution trend of drastic concentration changes. The final output single-channel two-dimensional matrix is ​​the preliminary dust distribution map with high-order semantic information.

[0029] In step S12, temperature interference correlation analysis is performed on the preliminary dust distribution map based on the temperature data to obtain temperature field data. The specific implementation method is as follows: First, the preliminary dust distribution map is divided into multiple cells, and the dust concentration difference between adjacent cells is calculated. Specifically, based on the aerodynamic flow field distribution characteristics inside the grinding equipment, the entire preliminary dust distribution map is uniformly cut into several rectangular cells of equal area. For each central cell, its four directly adjacent cells in the four orthogonal directions (top, bottom, left, and right) are traversed, and their respective dust concentration values ​​are extracted. The dust concentration value of the central cell is subtracted from the concentration values ​​of these four adjacent cells, and the absolute values ​​are taken to obtain the dust concentration difference in the four directions. To comprehensively measure the degree of difference between the central cell and the surrounding environment, the arithmetic mean of these four absolute differences is calculated, and the average value is taken as the final comprehensive dust concentration difference of the central cell.

[0030] When performing threshold comparison, the dust concentration difference is further converted into a dust concentration difference per unit distance based on the center distance between adjacent cells, and the calculation formula is as follows: Where C represents the dust mass concentration of the cell, in mg / m³; subscript i represents the center cell, and subscript j represents the adjacent cell; l represents the physical distance between the center points of two cells, in cm. Therefore, the unit of G is mg / (m³·cm), which can be directly compared with the preset gradient threshold.

[0031] Subsequently, cells with dust concentration differences exceeding a preset gradient threshold are marked as boundary points, and these boundary points are aggregated to obtain the non-uniform diffusion region. Specifically, the calculated comprehensive dust concentration difference of each cell is compared one by one with the preset gradient threshold. If the dust concentration difference of a cell is greater than the preset gradient threshold, it indicates that the dust concentration at the cell's location has undergone an extremely steep abrupt change. The system then extracts the two-dimensional spatial coordinates of that cell and assigns it a Boolean-type boundary point label. All cells marked as boundary points are aggregated and connected in the spatial coordinate system. The polygonal interior region enclosed or delineated by these boundary points is the non-uniform diffusion region within the grinding chamber, where dust diffusion is extremely unstable and prone to turbulence.

[0032] It should be noted that, according to the calibration method of the present invention, in actual metal powder grinding equipment, when the grinding chamber plane is divided into centimeter-level cells, the preset gradient threshold obtained through statistical analysis is typically within an exemplary range of 0.5 mg / (m³·cm) to 5 mg / (m³·cm). This range coefficient is obtained by collecting all cell dust concentration difference samples generated under safe and stable operating conditions of the equipment over the past thirty days, constructing its probability density distribution histogram, and calibrating it based on the 95th quantile value of the sample concentration difference using quantile statistics. This range will vary depending on the equipment model and the grinding material.

[0033] The samples used for calibration are all dust concentration difference samples per unit distance, rather than simply the difference between two concentration values. When the grid size changes, the values ​​are first divided by the corresponding center distance before quantile statistics are performed to maintain consistent threshold dimensions.

[0034] Subsequently, the K-means clustering algorithm is used to cluster the dust distribution within the non-uniform diffusion region, resulting in dust aggregation areas. Specifically, within the defined non-uniform diffusion region, the dust concentration values ​​of all cells are extracted as the input sample set for cluster analysis. The number of clusters K in the K-means clustering algorithm is set to two, representing the two extreme states of high concentration aggregation and low concentration sparseness. The algorithm first randomly selects two concentration values ​​in the sample set as initial cluster centers. Next, the absolute distance between the dust concentration value of each cell in the region and these two initial cluster centers is calculated. Based on the principle of closest proximity, each cell is assigned to the corresponding cluster. After completing one full assignment, the arithmetic mean of the concentrations of all cells within these two clusters is calculated, and these two averages are updated as new cluster centers. The above assignment and center update process is iterated and repeated until the change in the new cluster centers compared to the previous iteration is extremely small, less than the set convergence accuracy, at which point the process stops. After clustering is completed, the final center values ​​of the two clusters are compared. All cells corresponding to the cluster with larger center values ​​are spatially contiguous, forming a dust accumulation area, while the cells corresponding to the cluster with smaller center values ​​form a sparse dust area.

[0035] Finally, spatial correlation analysis is performed on the dust accumulation area and the temperature data to obtain temperature field data. Specifically, the system extracts the original temperature data matrix obtained in step S11 and uses the spatial coordinate mask of the dust accumulation area to perform spatial projection clipping on the temperature data matrix, retaining only the set of temperature pixels that completely overlap with the spatial position of the dust accumulation area. During metal grinding, high dust accumulation is often accompanied by intense mechanical friction and kinetic energy conversion, thereby generating a large amount of heat. Therefore, the system extracts the dust concentration value and the corresponding local temperature value of each pixel in the area, and uses the Pearson correlation coefficient calculation formula to calculate the linear correlation between the concentration sequence and the temperature sequence. The calculated correlation coefficient, the absolute value of the local temperature at each coordinate point in the dust accumulation area, and the temperature gradient direction vector are structured and packaged to form a multi-dimensional data structure that integrates the physical barrier characteristics and thermodynamic radiation characteristics of dust, which is the final output temperature field data.

[0036] In step S13, local fluctuation values ​​are extracted from the temperature field data, temperature anomaly areas with local fluctuation values ​​greater than a preset fluctuation threshold are filtered out, and the temperature anomaly areas are summarized to obtain a temperature anomaly map. The specific implementation method is as follows: First, the temperature field data is spatially gridded, and the temperature change rate of each grid cell is calculated. Specifically, the generated temperature field data is finely divided into spatial grids according to a set physical resolution (e.g., one grid per square centimeter). To capture the dynamic evolution of temperature, the system introduces a five-second time sliding window. Furthermore, to adapt to the dynamic changes under different operating conditions, an adaptive window adjustment mechanism is introduced: when a sustained increase in the temperature change rate is detected, the sliding window length is shortened (e.g., adjusted to 2–3 seconds) to improve anomaly response sensitivity; when the system is in a stable operating state, the window length is appropriately extended (e.g., adjusted to 5–10 seconds) to reduce the impact of random noise on the calculation results, thereby achieving a balance between detection sensitivity and false alarm rate. At each moment, the latest temperature value at the end of the current time window of the grid cell is extracted, and the historical temperature value at the beginning of the window is subtracted to obtain the absolute value of the temperature difference within the time interval. Then, this absolute value of the temperature difference is divided by the length of the time sliding window (five seconds), and the quotient represents the instantaneous rate of temperature rise or fall of the grid cell during this period, which is defined as the temperature change rate of the grid cell.

[0037] Subsequently, the standard deviation of the temperature change rate is calculated to obtain the local fluctuation value. If the local fluctuation value is greater than a preset fluctuation threshold, it is marked as a potential anomaly. Specifically, the system selects a 3x3 local neighborhood matrix centered on each target grid cell. The temperature change rate values ​​of nine grid cells within this neighborhood matrix are extracted. First, the arithmetic mean of these nine values ​​is calculated to represent the overall temperature change trend of the local area. Next, the square of the difference between each of the nine values ​​and the average is calculated. The sum of these nine squared values ​​is divided by the number of grid cells, and finally, the square root of the quotient is taken. Through the statistical variance and standard deviation calculation process described above, the result obtained is the local fluctuation value of the target grid cell. The larger the local fluctuation value, the more uneven the rate of temperature rise or fall within the small area where the grid cell is located, indicating severe thermal oscillations. The system compares the calculated local fluctuation value with the preset fluctuation threshold. If the local fluctuation value is found to be greater than the preset fluctuation threshold, the spatial coordinates of the target grid cell are extracted and marked as a potential anomaly.

[0038] It should be noted that the preset fluctuation threshold is determined by combining grid search and cost-sensitive learning. Historical records of temperature fluctuations inside the grinding mill over the past 10,000 working hours are collected, and on-site process engineers precisely label the actual high-temperature fault times such as dry friction of the grinding media and localized material caking. Multiple candidate fluctuation thresholds ranging from 0.1 to 5.0 are traversed in the historical data pool. These fluctuation thresholds are dimensionless normalized parameters, representing the relative degree of temperature fluctuation deviating from the historical stable average. The system selects candidate thresholds sequentially in steps of 0.1. For each candidate threshold, the number of potential anomaly alarms triggered is counted and compared with manually labeled actual faults, calculating the false alarm rate and false alarm rate at that threshold. A comprehensive loss function is constructed, which is the false alarm rate multiplied by ten plus the false alarm rate multiplied by one. The candidate threshold that minimizes this comprehensive loss function value is selected as the final preset fluctuation threshold, ensuring that the monitoring system operates in the most economical and safest state.

[0039] Next, the spatial coordinates of the potential anomaly points are spatially superimposed with the spatial coordinates of the dust accumulation area to calculate the spatial proximity distance. If the spatial proximity distance is less than a preset proximity radius, the potential anomaly point is determined to be a temperature anomaly region. Specifically, the coordinates of all selected potential anomaly points are projected onto a two-dimensional plane coordinate system containing the outline of the dust accumulation area. For each potential anomaly point, the boundary point of the dust accumulation area closest to its physical location is found. Using the Euclidean distance calculation method, the coordinate differences between the potential anomaly point and the nearest boundary point on the horizontal and vertical axes are calculated respectively. The two differences are squared and added together, and the square root of the sum is taken. The result is the absolute spatial proximity distance from the potential anomaly point to the dust accumulation area. When the spatial proximity distance is less than the preset proximity radius, it is determined that the potential anomaly point and the dust accumulation area have a spatial coupling relationship, and the grid cell containing the potential anomaly point and its neighboring grid cells are marked as temperature anomaly regions.

[0040] It should be noted that the preset proximity radius is obtained through joint calibration using thermodynamic conduction formulas and the physical dimensions of the equipment. Specifically, the thermal conductivity of the metal material used in the inner wall of the grinder (typically ranging from 15 to 50 W / (m·K)) and the convective heat transfer coefficient of the internal mixed gas (typically ranging from 5 to 25 W / (m²·K)) are extracted. Assuming a standard dust-clogged heat source, the physical distance range required for its heat to conduct outwards and attenuate to a normal safe temperature under the current ambient temperature is calculated. Combining the actual physical gap width between the stator and rotor inside the grinder (e.g., between 3 mm and 5 mm), the smaller value between the heat attenuation calculation distance and the physical gap width is selected, plus an error compensation constant based on the measuring instrument (e.g., 0.5 mm).

[0041] Finally, the local fluctuation values ​​and dust concentration data of the temperature anomaly areas are visualized using a heatmap to generate a temperature anomaly map. Specifically, the center coordinates of all officially identified temperature anomaly areas, along with their corresponding local fluctuation values ​​and absolute dust concentration data, are extracted. After normalizing the local fluctuation values ​​and dust concentrations, they are multiplied and added together with weights of 0.7 and 0.3 to obtain a comprehensive heat value. Using a color mapping mechanism, this comprehensive heat value is mapped to a specific color in the RGB color space. The mapping rule is as follows: the safe area with the lowest heat value is rendered as dark blue, and as the heat value increases, the color gradually transitions smoothly from light blue, green, and yellow until the anomaly area with the highest heat value and extreme danger is rendered as a high-brightness dark red. A Gaussian blur kernel is used to smoothly diffuse these generated color nodes on a global two-dimensional canvas to eliminate harsh boundaries between color blocks, ultimately outputting a temperature anomaly map that intuitively displays the superimposed risks of temperature and dust inside the grinding chamber.

[0042] like Figure 2 As shown, in step S14, the time-series analysis of the temperature anomaly map is performed to obtain a temperature diffusion sequence, historical dust distribution is acquired, a similarity analysis is performed on the preliminary dust distribution map and the historical dust distribution to obtain a similarity sequence, and a deviation calculation is performed on the temperature diffusion sequence and the similarity sequence to obtain a temperature deviation value, including: Step S141: Extract the temperature change trajectory of each grid cell in the temperature anomaly map within a continuous time period, and perform Markov weighted network propagation on the temperature change trajectory to obtain the temperature diffusion sequence. Step S142: Retrieve historical dust distribution data and perform dynamic time-normalization comparison between the historical dust distribution data and the preliminary dust distribution map to obtain a similarity sequence. Step S143: Calculate the energy decay value of the temperature diffusion sequence to obtain the temperature energy decay characteristics; calculate the rate of change of the similarity sequence to obtain the dust distribution change characteristics. Step S144: The temperature energy decay characteristics and the dust distribution change characteristics are fused using a multivariate linear weighted fusion method to obtain the temperature deviation value.

[0043] First, the temperature change trajectory of each grid cell in the temperature anomaly map is extracted over a continuous time period. A Markov weighted network propagation is then performed on these temperature change trajectories to obtain a temperature diffusion sequence. Specifically, on a historical timeline of a preset length of thirty seconds, the continuously generated temperature anomaly maps are overlaid at a sampling frequency of ten frames per second to form a three-dimensional spatiotemporal data block. For each grid cell on the map, the comprehensive temperature heat value at its corresponding location within these thirty seconds is extracted and arranged in chronological order to form the temperature change trajectory specific to that grid. Then, the entire temperature anomaly map is considered as a complex graph network structure, with each grid cell as a network node. Information propagation edges exist between adjacent grid cells, with the temperature difference between nodes and the average dust concentration used as edge weights: the greater the temperature difference and the sparser the dust (facilitating heat convection), the higher the propagation weight.

[0044] Based on the state transition concept of Markov chains, the energy distribution simulation is performed on the hottest node state from the previous time step, according to the weight ratio of its associated edges, to its surrounding neighboring nodes. Specifically, the thermal state of each grid node at time t is denoted as... The process for defining the transfer probability from node i to neighboring node j is as follows: Extract the temperature difference between node i and node j, and the dust concentration at node j. Use the max-min normalization method to process the temperature difference and dust concentration, mapping their values ​​to the [0,1] interval to obtain the normalized temperature difference and normalized dust concentration. Calculate the product of the normalized temperature difference and the difference between one and the normalized dust concentration, using this as the initial transfer weight from node i to node j. Iterate through all neighboring nodes of node i, summing the initial transfer weights for all neighboring nodes to obtain a total value. Finally, divide the initial transfer weight for each neighboring node by this total value to obtain the final transfer probability. This step ensures that the sum of the transfer probabilities from node i to all neighboring nodes is 1.

[0045] At each moment, the state update formula is used. Perform iterative propagation, where This is the transition probability matrix. The propagation calculation terminates when the overall state change between two adjacent iterations is less than 0.001 or the maximum number of iterations (e.g., 50). Furthermore, in terms of boundary handling, adiabatic boundary conditions are applied to the equipment boundaries, meaning that boundary nodes do not transfer heat outwards but only participate in energy exchange with internal nodes. Through this dynamic propagation iteration of the weighted network, the displacement path of the centroid coordinates of the overall high-risk heat region and the expansion rate of the heat-affected area are recorded at different time sections. This time series array reflecting the temporal and spatial spread trend of heat is defined as the temperature diffusion sequence.

[0046] Subsequently, historical dust distribution data is retrieved and dynamically time-warped compared with the preliminary dust distribution map to obtain a similarity sequence. Specifically, standard historical dust distribution time-series data within a fault-free normal operation cycle with a deviation of ±10% from the current batch of grinding materials and equipment operating power is retrieved from the system's long-term stored database. At this point, the system faces two sequences of different lengths, which may have local time misalignments: one is the evolution sequence of the currently generated preliminary dust distribution map within a recent time window, and the other is the standard historical dust distribution sequence. To accurately measure the morphological differences between the two, the system employs the Dynamic Time Warping (DTW) algorithm. This algorithm does not require the two sequences to be absolutely equal in time length, but rather constructs a two-dimensional distance-cost matrix to calculate the spatial Euclidean distance between the morphology of each time point in the current sequence and the morphology of each time point in the historical sequence. Then, an optimal curved path from the upper left corner to the lower right corner is found in this matrix, minimizing the sum of all distance costs traversed along the path. During this process of finding the optimal matching path, the reciprocal of the local matching distance at each time step is recorded. Arrange these reciprocals in chronological order and normalize them to form a fluctuating sequence between zero and one. The closer the value in this sequence is to one, the more similar the current dust state is to the historical normal state. This sequence is defined as the similarity sequence.

[0047] Next, energy decay values ​​are calculated for the temperature diffusion sequence to obtain temperature energy decay characteristics, and change rate calculations are performed on the similarity sequence to obtain dust distribution change characteristics. Specifically, for the previously obtained temperature diffusion sequence, the area of ​​heat influence between adjacent time points in the sequence is extracted, and the relative area expansion rate is calculated according to the time interval. Simultaneously, the heat dissipation power Q is calculated based on the coolant mass flow rate, coolant specific heat capacity, and coolant inlet and outlet temperature difference. Then, the baseline heat dissipation power under safe and stable operating conditions is used as the benchmark. Normalization is performed, and the calibration coefficients are used. Converted to equivalent cooling shrinkage rate .in, The unit is 1 / s, and its value is obtained by calibrating the relationship between the natural fall rate of abnormal area and the heat dissipation power in the data of safe and stable operation. Therefore, the temperature energy decay characteristic is defined as... All three terms are in the dimension of 1 / s, allowing for numerical subtraction. For the similarity sequence, the first-order difference of each data point in the sequence relative to the previous data point is calculated, and temporal smoothing filtering is performed to obtain the dust distribution variation characteristics.

[0048] Finally, the temperature energy decay characteristic and the dust distribution change characteristic are fused using a multivariate linear weighted fusion to obtain the temperature deviation value. Specifically, considering the different physical characteristics of temperature thermodynamic hysteresis and dust motion instantaneity, a fixed weight determined based on the Analytic Hierarchy Process (AHP) is used to fuse the features of these two dimensions. In the AHP calculation during the system initialization phase, a judgment matrix is ​​constructed based on field experience, assigning a three-to-one importance ratio to the fault indication of "temperature energy accumulation" relative to "dust morphology change". The maximum eigenvalue and corresponding eigenvector of the judgment matrix are calculated, and a consistency check is performed. Specifically, the consistency ratio is calculated. When the consistency ratio is less than 0.1, the judgment matrix is ​​considered to have satisfactory consistency, and the obtained weight parameters are valid. If the consistency ratio is greater than or equal to 0.1, the judgment matrix needs to be adjusted until the consistency requirement is met, thereby ensuring the rationality and reliability of the weight allocation. The finally extracted eigenvectors are established as fixed weight parameters, for example, the weight of the temperature dimension is 0.75, and the weight of the dust dimension is 0.25. During real-time calculations, the current temperature energy decay characteristic value is multiplied by 0.75, and the current dust distribution change characteristic value is multiplied by 0.25. The products of these two values ​​are then arithmetically summed. The summation result is the temperature deviation value.

[0049] In step S15, optimizing the temperature deviation value through gradient descent and optimizing the pre-established environmental monitoring model to obtain an optimized environmental model includes: First, a gradient descent algorithm is used to smooth and optimize the temperature deviation value sequence within the sliding time window, resulting in optimized temperature deviation values. Specifically, the sequence of values ​​with a length of... The original temperature deviation value sequence is denoted as to The variable to be optimized is set as a smoothed deviation sequence within the same time window. to Instead of using a single, pre-computed scalar result as a model parameter, construct the objective function: The second term is the smoothing weight, which controls the importance of the "smoothing constraint". The larger the value, the stronger the second term's influence, and the more the system prioritizes smoothness at adjacent time points. It will be closer and the curve will be smoother, but it may also smooth out real sudden anomalies. The smaller the value, the weaker the influence of the second term, and thus the weaker the influence on the original data. It is more respectful and can preserve the original mutation information, but its noise resistance is weaker.

[0050] For example, =0.1 indicates mild smoothing; 1 indicates moderate smoothing; 5 or higher indicates strong smoothing. Generally, values ​​are between 0.1 and 1.0. The first term constrains the smoothing deviation sequence to not deviate from the original calculation result, and the second term constrains changes between adjacent time points to not jump excessively. A learning rate of 0.001 is used to iteratively update the partial derivatives of the objective function with respect to the smoothing deviation variable at each time point, with a maximum of 100 iterations; iteration stops when the absolute value of all partial derivatives is less than 0.0001. The output smoothing deviation value at the current time is the optimized temperature deviation value.

[0051] Subsequently, based on pre-acquired historical fault data, the relationship between equipment operating temperature and fault probability was analyzed to obtain the critical temperature value for potential hazards. Specifically, at least 50 downtime maintenance records, fire records, and corresponding microscopic sensor logs related to overheating of the grinding machine in the past were reviewed from the Equipment Asset Management System (EAM). The absolute temperature value and heating rate characteristics within 30 minutes prior to each fault were extracted. A survival analysis was performed on these historical destructive temperature data using a Weiper distribution statistical model. The shape and scale parameters of the Weiper distribution were calculated using the maximum likelihood estimation method, thereby establishing a continuous mathematical curve between equipment operating temperature and the probability of fatal equipment failure. On the fitted fault probability curve, the horizontal axis temperature value corresponding to a 90% cumulative failure probability risk was identified. This temperature scale, representing the point at which serious damage is highly probable once the equipment exceeds this limit, was strictly defined as the critical temperature value for potential hazards of the equipment.

[0052] Finally, the pre-established environmental monitoring model is optimized through backpropagation based on the optimized temperature deviation value and the critical temperature value of the potential hazard, resulting in an optimized environmental model. The pre-established environmental monitoring model is a deep time-series prediction model based on a Long Short-Term Memory (LSTM) network, designed to predict future temperature trends based on past environmental conditions. This model has been pre-trained offline using a massive amount of normal operating condition sequences. The input consists of a dust concentration matrix and a temperature field matrix for ten consecutive time steps, and the output is the expected temperature state for the next time step. The pre-training process of the model involves extracting multiple sets of dust concentration data sequences and temperature field data sequences from a historical operational database as training samples. Each training sample contains a dust concentration matrix and a corresponding temperature field matrix for multiple consecutive time steps, with the actual temperature field after the current time step serving as the supervision label. The dust concentration matrix sequence and temperature field matrix sequence are normalized and then input into a Long Short-Term Memory (LSTM) network model. The model includes an input layer, at least one LSTM hidden layer, and an output layer. The LSTM hidden layer has 64 neurons, and the time step length is set to 10 sampling periods. The mean square error between the predicted temperature field and the actual temperature field output by the model is used as the loss function, and an adaptive moment estimation optimization algorithm is employed to iteratively update the model parameters. During training, a maximum of 300 training epochs are set. Training stops when the mean square error of the validation set decreases by no more than 0.0001 within twenty consecutive preset epochs, and the corresponding model weight parameters are saved as the initial parameters of the online environmental monitoring model.

[0053] During the online operation phase, when the calculated optimized temperature deviation value continues to rise, and its corresponding absolute temperature approaches the critical value of the hazard temperature obtained in the previous step, the online fine-tuning mechanism of the model is triggered. The system normalizes both the optimized temperature deviation value and the critical value of the hazard temperature into dimensionless risk offsets, and then introduces the original mean square error loss function to form a composite loss function. .in, This represents the predicted temperature field tensor output by the model. This represents the measured temperature field tensor. Indicates the number of tensor elements. This represents the risk temperature calculated from the current optimized temperature deviation value. This indicates the critical temperature value indicating a potential hazard. The first term represents the penalty term weight, exemplified by a value of 0.1. The second term maintains the accuracy of the temperature field prediction, while the third term is a dimensionless regularization penalty term; the two can be weighted and added together. Subsequently, the backpropagation algorithm is used to calculate the gradient of the composite loss function with respect to each weight matrix of the LSTM model, and the Adam optimizer is used for fine-tuning. The first-order moment decay coefficient is set to 0.9, the second-order moment decay coefficient is set to 0.999, and the base learning rate is set to 0.001.

[0054] In step S16, the step of simulating the interaction between dust and temperature using the optimized environmental model, locating abnormal areas during the interaction process, and obtaining the coordinates of the abnormal areas includes: First, the interaction between the dust concentration data and the temperature data is analyzed using the optimized environment model to obtain the probability of this interaction. Specifically, the latest two-dimensional dust concentration distribution matrix and temperature field data matrix are simultaneously input into the newly fine-tuned optimized environment model (LSTM architecture). Within the model, the hidden states of the dust data and temperature data undergo deep feature space cross-fusion through a series of matrix multiplication operations and nonlinear activation function mappings in fully connected layers. A Sigmoid classification activation function layer is added at the very end of the network. This function nonlinearly maps and compresses the high-dimensional, complex interaction features into a continuous numerical range between zero and one. For each physical space cell in the grid, the model outputs a decimal between zero and one. This decimal represents the probability that, at that specific location, extremely high dust concentration hinders heat dissipation, leading to further temperature deterioration and loss of control—the probability of this interaction occurring—and is defined as the interaction probability of that cell.

[0055] Subsequently, regions with interaction probability greater than a preset impact threshold are identified as high-risk regions. Specifically, the system performs a global traversal scan of the interaction probability output for each cell in the 3D mapping mesh of the grinding chamber. Each interaction probability value is compared with a preset impact threshold pre-loaded in memory. For any mesh cell with a probability value greater than the preset impact threshold, its border is forcibly highlighted in red within the system's internal data structure, and all these consecutive, condition-compliant red cells are spatially merged. The resulting irregular topological polygonal region spanning multiple mesh cells is then officially upgraded by the system and labeled as a high-risk region.

[0056] It should be noted that the preset impact threshold is derived from the receiver operating characteristic curve (ROC curve) analysis of the validation set data. Specifically, during the offline training and validation phase of the model, no fewer than 1000 independent test samples containing known real-world hazardous interaction events were selected. The probability results predicted by the model for these test samples were traversed from 0.01 to 0.99 with a minimal step size as candidate thresholds. For each candidate threshold, the predicted true positive rate (i.e., the proportion of true hazards correctly identified) and false positive rate (i.e., the proportion of safe events mistakenly identified as hazardous events) were calculated. The ROC curve was plotted with the false positive rate on the horizontal axis and the true positive rate on the vertical axis. Subsequently, the point on the curve with the closest Euclidean distance to the upper left corner (0, 1) was found, which is the probability cutoff point where the difference between the true positive rate and the false positive rate is maximized. This optimal cutoff point value, which mathematically perfectly balances the interference of false alarms and the risk of missed alarms, was extracted and solidified as the preset impact threshold for online operation.

[0057] Finally, the center coordinates of the high-risk area are extracted to obtain the coordinates of the abnormal area. Specifically, for the irregular high-risk area formed by merging in the previous step, a weighted centroid calculation method based on probability intensity is adopted. The horizontal and vertical physical coordinates of each cell within the high-risk area, as well as the corresponding interaction probability value of each cell, are extracted. During calculation, the horizontal physical coordinate of each cell is multiplied by its interaction probability as a weight, and the product of all cells is summed and then divided by the sum of the interaction probabilities of all cells. The quotient is used as the final horizontal coordinate of the location. Similarly, the vertical physical coordinate of each cell is multiplied by its interaction probability, summed, and then divided by the sum of the probabilities. The quotient is used as the final vertical coordinate of the location. Through this mathematical transformation process similar to "centroid calculation" in physics, the located coordinates can be accurately biased towards the center point of the deepest danger and the highest probability. Outputting this pair of weighted coordinates yields the coordinates of the abnormal area.

[0058] In step S17, the preliminary dust distribution map and the temperature anomaly map are updated based on the coordinates of the abnormal area to obtain the closed-loop monitoring results. The specific implementation method is as follows: First, the dust concentration data in the abnormal region is spatially smoothed to obtain an updated dust distribution map. Specifically, when sensors detect severely abnormal regions, the irregular high-frequency scattering of dust often causes sharp noise in local concentration values. To restore the true physical concentration distribution, a two-dimensional Gaussian filter algorithm is used to spatially smooth the dust concentration matrix within the extracted abnormal region coordinates as the center point. The Gaussian filter constructs a weight matrix with the highest weight at the center and an exponential decay towards the surrounding areas. A sliding convolution operation is performed on the dust concentration data grid of the abnormal region, so that the concentration value of each grid point is replaced by the weighted average of its surrounding neighboring grid points. The kernel size and standard deviation parameters of the Gaussian smoothing filter are determined based on the area of ​​the abnormal region; specifically, the area A of the abnormal region is converted into an equivalent radius. Where A represents the area of ​​the abnormal region, in units of grid cell area; r represents the equivalent radius, in units of grid cells. The filtering parameters are determined according to the following rules: when the equivalent radius is less than 3 grid cells, a 3×3 convolution kernel is selected with a standard deviation σ of 1; when the equivalent radius is between 3 and 6 grid cells, a 5×5 convolution kernel is selected with σ ranging from 1.5 to 2; when the equivalent radius is greater than or equal to 6 grid cells, a 7×7 or larger convolution kernel is selected with σ ranging from 2 to 3. This yields an updated dust distribution map that removes high-frequency noise interference and more closely approximates the actual fluid dynamics gradient.

[0059] Subsequently, kernel density estimation is used to process the temperature data of the abnormal region to obtain an updated temperature anomaly map. Specifically, the system extracts the latest measured temperature data of the equipment and combines it with the denoised dust concentration data. Using the coordinates of the abnormal region obtained in step S16 as the radiation starting point, a heat map generation algorithm based on kernel density estimation is used. For dust concentration, the minimum concentration in the current safe operation sample of the equipment is first used. and maximum concentration After normalization, we get .in, , and The units are all mg / m³, after normalization It is a dimensionless value between 0 and 1. Then... Mapped to drag coefficient Therefore, the value of λ ranges from 1 to 1.5. After correcting the diffusion term in the standard Gaussian kernel function by λ, the high-dust region only shows a decrease in heat diffusion rate, without causing the attenuation factor to approach 0. The full-plane energy density distribution value after concentration damping correction is mapped onto the pseudo-color band, and the old image module on the original interface is refreshed to obtain an updated temperature anomaly map with higher accuracy and rendering boundaries that are more in line with physical reality.

[0060] Next, the updated dust distribution map and the updated temperature anomaly map are subjected to time-series analysis with a preset historical database to obtain the diffusion trend of the abnormal area. Specifically, the updated dust distribution map and the updated temperature anomaly map generated in the next second are aligned and compared frame by frame with the similar historical maps stored in the preset historical database within the last ten minutes. This preset historical database is a time-series buffer storage area that continuously scrolls and stores the dust distribution map and temperature anomaly map calculated within the last ten minutes. The area value of the bright red core region with extremely high temperature and high dust in this time-series roll is extracted. Using the time series as the independent variable (horizontal axis) and the extracted anomaly area value as the dependent variable (vertical axis), a univariate linear regression model is established using ordinary least squares (OLS). In terms of the formula, this means calculating the covariance of the independent variable and the dependent variable, dividing by the variance of the independent variable, and thus obtaining the slope of the fitted line. If the calculated slope is positive and continues to increase, it indicates that the abnormal region is irreversibly and rapidly deteriorating and expanding; if the slope is close to zero, it indicates that the risk is in a stalemate; if the slope is negative, it indicates that the risk is contracting and easing. The direction and absolute value of this slope constitute the trend of the abnormal region's spread in assessing the severity of the current situation.

[0061] Finally, a preset dynamic adjustment mechanism is triggered based on the abnormal area diffusion trend to obtain closed-loop monitoring results. Specifically, the system is equipped with a programmable logic controller (PLC) driven environmental control module, including an electromagnetic proportional valve for the cooling water jacket, a frequency converter for the exhaust fan, and a stepper motor for the grinding feed screw. The system analyzes the abnormal area diffusion trend obtained above: when the area expansion slope is found to be greater than zero and exceeds the safety buffer slope threshold, where the safety buffer slope threshold is a fixed parameter pre-calibrated based on the equipment's risk tolerance. The calibration process is as follows: extract the evolution curve of the abnormal area expansion slope from the historical fault database of each event that evolved from local overheating to equipment damage, take the minimum expansion slope value within the five minutes before all faults occur and multiply it by a safety factor of 0.8, and finally set the safety buffer slope threshold to 2cm² / s. This value indicates that when the abnormal area continues to expand at a rate exceeding 2cm² / s, the system determines that active physical intervention is necessary. The monitoring and control platform actively sends digital control commands to the underlying PLC. After receiving the instruction, the PLC, on the one hand, increases the frequency of the cooling water pump's inverter by 20% according to the proportional-integral-derivative (PID) control logic to increase the coolant circulation flow. The proportional controller is set to 2.5, the integral time constant to 30s, and the derivative time constant to 5s. Simultaneously, the opening of the exhaust and dust removal system's damper is increased. On the other hand, an intervention signal is sent to the material feeding system to forcibly reduce the raw material feeding speed, reducing frictional heat generation at the source. After executing the above electromechanical linkage physical intervention actions, the system continues to cycle back to step S11 for a new round of data acquisition and monitoring verification. When it is confirmed that the temperature and dust concentration curves have steadily fallen back to the normal range, the complete timeline log from "anomaly detection, analysis and location" to "physical intervention elimination" is packaged, encrypted, and stored, and a report is output, thus obtaining the final closed-loop monitoring result of intelligent self-healing.

[0062] In a series of sample verifications, dynamic high-concentration dust interference tests were conducted using aluminum alloy powder and stainless steel powder, respectively. The dust concentration ranged from 10 mg / m³ to 1000 mg / m³, and the grinding chamber temperature ranged from 25℃ to 180℃. A total of 30 sets of continuous operating data were collected. Using manually verified thermal anomaly locations as a reference, the average positioning error of the anomaly area center coordinates was 1.8 cm, the alarm response time was less than 3 seconds, the false alarm rate was 4.2%, and the missed alarm rate was 3.1%. Under the same operating conditions, compared with alarms based solely on single-point temperature thresholds, this embodiment can identify local heat accumulation earlier in areas with significant dust obscuration while maintaining a low false alarm rate. The above data are used to illustrate the feasibility and reliability of this embodiment under dynamic high-concentration dust interference conditions.

[0063] In summary, this invention discloses an online monitoring method for the temperature field of metal powder grinding, comprising: acquiring dust concentration data and temperature data; performing convolution processing on the dust concentration data to obtain a preliminary dust distribution map; performing temperature interference correlation analysis on the preliminary dust distribution map based on the temperature data to obtain temperature field data; extracting local fluctuation values ​​from the temperature field data, screening temperature anomaly regions where the local fluctuation values ​​are greater than a preset fluctuation threshold, and summarizing the temperature anomaly regions to obtain a temperature anomaly map; performing time-series analysis on the temperature anomaly map to obtain a temperature diffusion sequence, acquiring historical dust distribution, performing similarity analysis on the preliminary dust distribution map and the historical dust distribution to obtain a similarity sequence, calculating the deviation between the temperature diffusion sequence and the similarity sequence to obtain a temperature deviation value; optimizing the temperature deviation value through gradient descent and optimizing a pre-established environmental monitoring model to obtain an optimized environmental model; simulating the interaction between dust and temperature through the optimized environmental model, locating anomaly regions in the interaction process, and obtaining the coordinates of the anomaly regions; updating the preliminary dust distribution map and the temperature anomaly map based on the coordinates of the anomaly regions to obtain closed-loop monitoring results. This invention enables precise online monitoring of the temperature field during metal powder grinding under dynamic high-concentration dust interference conditions.

[0064] Reference Figure 3 The second embodiment of the present invention provides an online monitoring system for the temperature field of metal powder grinding, comprising: The data acquisition module is used to acquire dust concentration data and temperature data, and to perform convolution processing on the dust concentration data to obtain a preliminary dust distribution map; The temperature analysis module is used to perform temperature interference correlation analysis on the preliminary dust distribution map based on the temperature data to obtain temperature field data. Anomaly identification module is used to extract local fluctuation values ​​from the temperature field data, filter temperature anomaly areas where the local fluctuation values ​​are greater than a preset fluctuation threshold, summarize the temperature anomaly areas, and obtain a temperature anomaly map. The deviation calculation module is used to perform time-series analysis on the temperature anomaly map to obtain a temperature diffusion sequence, acquire historical dust distribution, perform similarity analysis on the preliminary dust distribution map and the historical dust distribution to obtain a similarity sequence, and perform deviation calculation on the temperature diffusion sequence and the similarity sequence to obtain a temperature deviation value. The model optimization module is used to optimize the temperature deviation value through gradient descent and optimize the pre-established environmental monitoring model to obtain an optimized environmental model. The coordinate output module is used to simulate the interaction between dust and temperature through the optimized environment model, locate abnormal areas in the interaction process, and obtain the coordinates of the abnormal areas. The closed-loop monitoring module is used to update the preliminary dust distribution map and the temperature anomaly map based on the coordinates of the abnormal area to obtain the closed-loop monitoring results.

[0065] It should be noted that the online monitoring system for the temperature field of metal powder grinding provided in this embodiment of the invention is used to execute all the process steps of the online monitoring method for the temperature field of metal powder grinding in the above embodiment. The working principle and beneficial effect of the two are one-to-one, so they will not be described again.

[0066] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0067] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for online monitoring of the temperature field during metal powder grinding, characterized in that, include: Obtain dust concentration data and temperature data, and perform convolution processing on the dust concentration data to obtain a preliminary dust distribution map; Based on the temperature data, a temperature interference correlation analysis is performed on the preliminary dust distribution map to obtain temperature field data. This includes: dividing the preliminary dust distribution map into multiple cells and calculating the dust concentration difference between adjacent cells; marking cells where the dust concentration difference exceeds a preset gradient threshold as boundary points, and summarizing these boundary points to obtain a non-uniform diffusion region; using a K-means clustering algorithm to cluster the dust distribution within the non-uniform diffusion region to obtain dust aggregation areas; and performing a spatial correlation analysis between the dust aggregation areas and the temperature data to obtain temperature field data. Local fluctuation values ​​are extracted from the temperature field data, temperature anomaly areas with local fluctuation values ​​greater than a preset fluctuation threshold are filtered out, and the temperature anomaly areas are summarized to obtain a temperature anomaly map. A time-series analysis is performed on the temperature anomaly map to obtain a temperature diffusion sequence, historical dust distribution is obtained, a similarity analysis is performed on the preliminary dust distribution map and the historical dust distribution to obtain a similarity sequence, and a deviation calculation is performed on the temperature diffusion sequence and the similarity sequence to obtain a temperature deviation value. The temperature deviation value is optimized by gradient descent, and the pre-established environmental monitoring model is also optimized to obtain an optimized environmental model. The interaction between dust and temperature is simulated using the optimized environmental model to locate abnormal areas during the interaction process and obtain the coordinates of the abnormal areas. Based on the coordinates of the abnormal area, the preliminary dust distribution map and the temperature anomaly map are updated to obtain closed-loop monitoring results.

2. The method for online monitoring of the temperature field during metal powder grinding according to claim 1, characterized in that, The process of acquiring dust concentration data and temperature data, and performing convolution processing on the dust concentration data to obtain a preliminary dust distribution map includes: Acquire dust concentration and temperature data collected by the sensor array; The dust concentration data is processed into a grid to obtain a two-dimensional distribution matrix; The particle distribution features and density gradient features of the two-dimensional distribution matrix are extracted through multi-layer convolution operations. The particle distribution characteristics and the density gradient characteristics are fused to obtain a preliminary dust distribution map.

3. The method for online monitoring of the temperature field during metal powder grinding according to claim 1, characterized in that, The process of extracting local fluctuation values ​​from the temperature field data, filtering out temperature anomaly areas where the local fluctuation values ​​are greater than a preset fluctuation threshold, and summarizing the temperature anomaly areas to obtain a temperature anomaly map includes: The temperature field data is divided into spatial grids, and the temperature change rate of each grid cell is calculated. The standard deviation of the temperature change rate is calculated to obtain the local fluctuation value. If the local fluctuation value is greater than the preset fluctuation threshold, it is marked as a potential anomaly. The spatial coordinates of the potential anomaly point are spatially superimposed with the spatial coordinates of the dust accumulation area to calculate the spatial proximity distance. If the spatial proximity distance is less than the preset proximity radius, the potential anomaly point is determined to be a temperature anomaly area. The local fluctuation values ​​of the temperature anomaly area and the dust concentration data are subjected to heat map visualization processing to generate a temperature anomaly map.

4. The method for online monitoring of the temperature field during metal powder grinding according to claim 1, characterized in that, The process involves performing time-series analysis on the temperature anomaly map to obtain a temperature diffusion sequence, acquiring historical dust distribution data, performing similarity analysis on the preliminary dust distribution map and the historical dust distribution to obtain a similarity sequence, and calculating the deviation between the temperature diffusion sequence and the similarity sequence to obtain a temperature deviation value, including: Extract the temperature change trajectory of each grid cell in the temperature anomaly map over a continuous time period, and perform Markov weighted network propagation on the temperature change trajectory to obtain the temperature diffusion sequence. Historical dust distribution data is retrieved, and the historical dust distribution data is dynamically time-normalized and compared with the preliminary dust distribution map to obtain a similarity sequence. The energy decay value of the temperature diffusion sequence is calculated to obtain the temperature energy decay characteristics, and the change rate of the similarity sequence is calculated to obtain the dust distribution change characteristics. The temperature energy decay characteristic and the dust distribution change characteristic are fused by multiple linear weighted fusion to obtain the temperature deviation value.

5. The method for online monitoring of the temperature field during metal powder grinding according to claim 1, characterized in that, The process of optimizing the temperature deviation value through gradient descent and optimizing the pre-established environmental monitoring model to obtain an optimized environmental model includes: The temperature deviation value is iteratively optimized using the gradient descent algorithm to obtain the optimized temperature deviation value. Based on the analysis of historical fault data obtained in advance, the relationship between equipment operating temperature and fault probability is analyzed to obtain the critical value of the equipment's potential hazard temperature. The optimized environmental monitoring model is optimized by backpropagation based on the optimized temperature deviation value and the critical temperature value of the hidden danger, thus obtaining the optimized environmental model.

6. The method for online monitoring of the temperature field during metal powder grinding according to claim 1, characterized in that, The step of simulating the interaction between dust and temperature using the optimized environmental model, locating abnormal areas during the interaction process, and obtaining the coordinates of the abnormal areas includes: The interaction between the dust concentration data and the temperature data is analyzed using the optimized environmental model to obtain the probability of the interaction. Areas where the probability of interaction impact is greater than a preset impact threshold are identified as high-risk areas; Extract the center coordinates of the high-risk area to obtain the coordinates of the abnormal area.

7. The method for online monitoring of the temperature field during metal powder grinding according to claim 1, characterized in that, The step of updating the preliminary dust distribution map and the temperature anomaly map based on the coordinates of the abnormal area to obtain closed-loop monitoring results includes: Spatial smoothing is performed on the dust concentration data in the abnormal area to obtain an updated dust distribution map. The temperature data of the anomalous region is processed using kernel density estimation to obtain an updated temperature anomaly map; The updated dust distribution map and the updated temperature anomaly map are subjected to time-series analysis with a preset historical database to obtain the diffusion trend of the anomaly area; Based on the spread trend of the abnormal area, a preset dynamic adjustment mechanism is triggered to obtain closed-loop monitoring results.

8. An online monitoring system for the temperature field of metal powder grinding, characterized in that, include: The data acquisition module is used to acquire dust concentration data and temperature data, and to perform convolution processing on the dust concentration data to obtain a preliminary dust distribution map; The temperature analysis module is used to perform temperature interference correlation analysis on the preliminary dust distribution map based on the temperature data to obtain temperature field data. This includes: dividing the preliminary dust distribution map into multiple cells and calculating the dust concentration difference between adjacent cells; marking cells where the dust concentration difference exceeds a preset gradient threshold as boundary points, and summarizing the boundary points to obtain a non-uniform diffusion region; using a K-means clustering algorithm to cluster the dust distribution within the non-uniform diffusion region to obtain dust aggregation areas; and performing spatial correlation analysis between the dust aggregation areas and the temperature data to obtain temperature field data. Anomaly identification module is used to extract local fluctuation values ​​from the temperature field data, filter temperature anomaly areas where the local fluctuation values ​​are greater than a preset fluctuation threshold, summarize the temperature anomaly areas, and obtain a temperature anomaly map. The deviation calculation module is used to perform time-series analysis on the temperature anomaly map to obtain a temperature diffusion sequence, acquire historical dust distribution, perform similarity analysis on the preliminary dust distribution map and the historical dust distribution to obtain a similarity sequence, and perform deviation calculation on the temperature diffusion sequence and the similarity sequence to obtain a temperature deviation value. The model optimization module is used to optimize the temperature deviation value through gradient descent and optimize the pre-established environmental monitoring model to obtain an optimized environmental model. The coordinate output module is used to simulate the interaction between dust and temperature through the optimized environment model, locate abnormal areas in the interaction process, and obtain the coordinates of the abnormal areas. The closed-loop monitoring module is used to update the preliminary dust distribution map and the temperature anomaly map based on the coordinates of the abnormal area to obtain the closed-loop monitoring results.

Images