Calcium carbide slag drying temperature control method and system based on multi-agent cooperation

Abstract

The application discloses a calcium carbide slag drying temperature control method and system based on multi-agent cooperation, relates to the technical field of industrial heat treatment, and comprises the following steps: S1, collecting real-time pressure data and temperature data through a sensor network arranged in a calcium carbide slag accumulation area to obtain a current bearing distribution diagram and a temperature distribution diagram of the calcium carbide slag accumulation; S2, processing the diagram data by adopting a convolutional neural network model according to the bearing distribution diagram and the temperature distribution diagram to determine a structural instability risk level of the calcium carbide slag accumulation; and S3, if the structural instability risk level is higher than a preset threshold, reducing the temperature input of the corresponding area by adjusting the power parameter of the heating mode to obtain an optimized heating distribution scheme. The calcium carbide slag drying temperature control method and system based on multi-agent cooperation significantly improve the safety and drying efficiency of the calcium carbide slag accumulation and reduce the instability and overheating risks through multi-model cooperation and dynamic feedback.

Description

Technical Field

[0001] This invention relates to the field of industrial heat treatment technology, specifically to a method and system for drying and controlling carbide slag based on multi-agent collaboration. Background Technology

[0002] In industrial production and resource processing, the drying of carbide slag is a crucial technology. Slag drying not only affects the efficiency of subsequent resource utilization but also directly relates to production safety and equipment lifespan. An efficient drying system needs to meet temperature requirements while ensuring the stability of the carbide slag stack structure, preventing collapse or equipment damage due to improper operation. Technological advancements in this field are of great significance for improving production efficiency and reducing safety risks. However, current drying technologies still face many challenges in balancing temperature control and structural stability, urgently requiring innovative breakthroughs.

[0003] Existing drying systems often employ a single heating method or static structural monitoring, making it difficult to adapt to the complex characteristics of carbide slag accumulation. The morphology and load-bearing distribution of carbide slag accumulation dynamically change due to material type, particle size, and stacking method. Traditional methods often overlook this dynamism, resulting in an inability to accurately respond to real-time changes in the stacked structure during heating. Furthermore, existing systems typically lack comprehensive consideration of structural stability in temperature control, and there is a lack of effective coordination between adjusting the heating method and monitoring the structural state. This makes the drying process prone to localized overheating or structural instability, affecting production safety and efficiency. The core technical challenge in carbide slag drying lies in achieving a dynamic balance between temperature control and structural stability. The primary technical factor is the highly heterogeneous load-bearing distribution of carbide slag accumulation. Differences in particle size, humidity, or packing density in different areas of carbide slag lead to uneven load-bearing capacity, and localized areas may experience sliding or collapse risks due to concentrated stress. This heterogeneity in load-bearing distribution further introduces a second technical factor: the direct impact of temperature changes on structural stability. Rapid heating or uneven heating can cause the internal moisture of carbide slag to evaporate rapidly, leading to volume expansion or contraction, which alters the load-bearing distribution and exacerbates the risk of structural instability. These two factors are intertwined, necessitating that the drying system dynamically monitor the load-bearing distribution while precisely controlling the heating method to avoid structural damage caused by temperature fluctuations.

[0004] Specifically, in actual production, the drying process of carbide slag piles often faces the following problem: when the heating system increases the temperature to accelerate the evaporation of moisture, cracks may be generated inside the carbide slag pile due to local overheating, resulting in a decrease in load-bearing capacity and even causing the pile to slip. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for temperature control of carbide slag drying based on multi-agent collaboration, which can accurately control the heating method under dynamically changing load distribution to avoid damage to the carbide slag accumulation structure caused by temperature changes.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a multi-agent collaborative method for drying and controlling the temperature of carbide slag, the method comprising: S1, collecting real-time pressure and temperature data through a sensor network deployed in the carbide slag accumulation area to obtain the current load-bearing distribution map and temperature distribution map of the carbide slag accumulation; S2, processing the data using a convolutional neural network model based on the load-bearing distribution map and temperature distribution map to determine the structural instability risk level in the carbide slag accumulation; S3, if the structural instability risk level is higher than a preset threshold, reducing the temperature input of the corresponding area by adjusting the power parameters of the heating method to obtain an optimized heating distribution scheme; S4, obtaining the optimized... After determining the heating distribution scheme, a support vector machine model is used to analyze the impact of the scheme on the overall drying efficiency to determine whether it meets the efficiency requirements under the production safety standard; S5. If the efficiency requirements under the production safety standard are met, the scheme is executed by adjusting the heating method in real time to obtain the updated temperature field and load-bearing distribution field of the carbide slag accumulation; S6. Based on the updated temperature field and load-bearing distribution field of the carbide slag accumulation, a recurrent neural network model is used to predict the structural stability change trend in the future time and determine the potential local thermal evolution path; S7. Through the potential local thermal evolution path, a targeted heating control command sequence is generated to obtain a dynamic monitoring cycle mechanism to maintain structural stability.

[0007] Preferably, step S1 includes: deploying a sensor network in a grid pattern over the carbide slag accumulation area to acquire pressure and temperature data, generating an initial dataset; using the initial dataset, removing noise data using a data cleaning method to obtain cleaned pressure and temperature data; using the cleaned pressure data, calculating the load-bearing distribution using an interpolation algorithm to generate a load-bearing distribution map of the carbide slag accumulation; using the cleaned temperature data, calculating the temperature distribution using an interpolation algorithm to generate a temperature distribution map of the carbide slag accumulation; if there are areas in the load-bearing distribution map where the pressure anomaly exceeds a preset threshold, then re-collecting regional data through the sensor network to obtain updated pressure data; if there are areas in the temperature distribution map where the temperature anomaly exceeds a preset threshold, then re-collecting regional data through the sensor network to obtain updated temperature data; and using the updated pressure and temperature data, fusing multiple sets of data using a weighted average algorithm to generate optimized load-bearing distribution and temperature distribution maps.

[0008] Preferably, step S2 includes acquiring load-bearing and temperature distribution maps of the carbide slag accumulation area through a sensor network to generate initial image data; standardizing the initial image data using a data preprocessing method to obtain standardized image data; extracting features from the standardized image data using a convolutional neural network to obtain feature vectors; calculating the probability of local overheating areas using a convolutional neural network based on the feature vectors to obtain a probability distribution; if there are areas in the probability distribution that are higher than a preset threshold, re-collecting the load-bearing and temperature data of that area through the sensor network to obtain updated image data; analyzing the updated image data using a convolutional neural network to calculate the structural instability risk level to obtain a risk assessment result; and classifying and labeling the carbide slag accumulation area using a region division method based on the risk assessment result to obtain a graded distribution map.

[0009] Preferably, step S3 includes acquiring temperature and load-bearing data of the carbide slag accumulation area through a sensor network to generate an initial distribution dataset; processing the initial distribution dataset using a data standardization method to obtain a standardized dataset; extracting features from the standardized dataset using a convolutional neural network to obtain a feature vector; if there are risk indicators in the feature vector that are higher than a preset threshold, then re-collecting temperature and load-bearing data of the corresponding area through the sensor network to generate an updated distribution dataset; analyzing the updated distribution dataset using a logistic regression model to determine the structural instability risk level; adjusting the heating power parameters according to the structural instability risk level to generate an optimized heating distribution scheme; and verifying the optimized heating distribution scheme through a sensor network to obtain a verification dataset and determine the stability of the temperature and load-bearing data.

[0010] Preferably, step S4 includes: acquiring temperature control parameters and load distribution data corresponding to the optimized heating distribution scheme through a sensor network to obtain an initial efficiency dataset; processing the initial efficiency dataset using a data standardization method to obtain a standardized efficiency dataset; performing classification analysis on the standardized efficiency dataset using a support vector machine model to obtain drying efficiency data; if the drying efficiency data is lower than the efficiency requirements preset by the production safety standard, then re-collecting the temperature control parameters and load distribution data through the sensor network to obtain an updated efficiency dataset; analyzing the updated efficiency dataset using a logistic regression model to determine risk assessment indicators; adjusting the temperature control parameters according to the risk assessment indicators to obtain an optimized adjustment strategy; verifying the optimized adjustment strategy through the sensor network, obtaining a verification dataset, and determining whether the drying efficiency data meets the production safety standard.

[0011] Preferably, step S5 includes acquiring updated temperature field data and load-bearing distribution field data of carbide slag accumulation through a sensor network to obtain an initial field distribution dataset; processing the initial field distribution dataset using a data standardization method to obtain a standardized field distribution dataset; performing classification analysis on the standardized field distribution dataset using a random forest model to obtain feature classification results for the temperature field and load-bearing distribution; if the feature classification results show that the temperature field or load-bearing distribution deviates from a preset stability threshold, then re-acquiring temperature field data and load-bearing distribution field data through the sensor network to obtain an updated field distribution dataset; performing dimensionality reduction processing on the updated field distribution dataset using principal component analysis to obtain a dimensionality-reduced feature dataset; analyzing the stability of the temperature field and load-bearing distribution using a logistic regression model based on the dimensionality-reduced feature dataset to obtain a stability assessment result; if the stability assessment result meets production safety standards, then adjusting the heating distribution parameters in real time through the sensor network to obtain optimized temperature field data and load-bearing distribution field data.

[0012] Preferably, step S6 includes acquiring temperature field data and load-bearing distribution field data of the carbide slag accumulation through a sensor network to obtain a real-time distribution dataset; processing the real-time distribution dataset using a data standardization method to obtain a standardized distribution dataset; analyzing the standardized distribution dataset using a recurrent neural network model to predict the structural stability change trend and obtain a stability trend prediction result; if the stability trend prediction result shows that the local thermal risk is higher than a preset threshold, then re-collecting temperature field data through the sensor network to obtain an updated temperature dataset; grouping the updated temperature dataset using a clustering analysis method to determine the distribution of areas where the local overheating risk is higher than the preset threshold and obtain a thermal region dataset; judging the local overheating path based on the thermal region dataset using a preset threshold to obtain the thermal evolution path distribution; and optimizing the thermal evolution path distribution by adjusting the heating distribution parameters to obtain optimized temperature field data and load-bearing distribution field data.

[0013] Preferably, step S7 includes acquiring temperature distribution data of the local overheated area through a sensor network to obtain a real-time temperature dataset; grouping the temperature distribution using K-means clustering analysis based on the real-time temperature dataset to determine the distribution of areas where the local temperature is higher than a preset threshold, thus obtaining an overheated area dataset; identifying temperature anomalies in the overheated area dataset using a preset threshold to generate an initial heating control command sequence, thus obtaining a control command dataset; and analyzing the control command dataset using a long short-term memory network model to predict the stable change trend of the structure, thus obtaining a stable trend prediction result.

[0014] Preferably, step S7 further includes: if the stable trend prediction result shows that the local thermal risk is higher than a preset threshold, then re-collecting temperature distribution data and load distribution data through a sensor network to obtain an updated distribution dataset; based on the updated distribution dataset, using a support vector machine model to classify the evolution path of the overheated region to obtain an optimized overheating evolution path distribution; and generating a dynamically adjusted heating control command sequence through the optimized overheating evolution path distribution to determine a real-time monitoring cycle mechanism.

[0015] A multi-agent collaborative temperature control system for drying carbide slag is disclosed, used to implement the steps of the aforementioned multi-agent collaborative temperature control method for drying carbide slag. The system includes a sensor network module that collects real-time pressure and temperature data through a sensor network deployed in the carbide slag accumulation area to obtain the current load-bearing distribution map and temperature distribution map of the carbide slag accumulation; an image processing module that processes the image data using a convolutional neural network model based on the load-bearing distribution map and temperature distribution map to determine the structural instability risk level in the carbide slag accumulation; a heating control module that, if the structural instability risk level is higher than a preset threshold, reduces the temperature input of the corresponding area by adjusting the power parameters of the heating method to obtain an optimized heating distribution scheme; and an efficiency evaluation module. The system employs several modules: a first module, the second module, and a third module. The first module, after obtaining an optimized heating distribution scheme, uses a support vector machine model to analyze the scheme's impact on overall drying efficiency and determine whether it meets the efficiency requirements under production safety standards. The third module, the fourth module, executes the scheme by adjusting the heating method in real time if the efficiency requirements are met, obtaining updated temperature and load-bearing distribution fields for the carbide slag accumulation. The fifth module, the sixth module, uses a recurrent neural network model to predict future structural stability trends based on the updated temperature and load-bearing distribution fields, identifying potential local thermal evolution paths. Finally, the seventh module, the eighth module, generates targeted heating control command sequences based on these potential local thermal evolution paths, thus establishing a dynamic monitoring and cyclic mechanism to maintain structural stability.

[0016] As can be seen from the above technical solution, the present invention has the following beneficial effects:

[0017] This multi-agent collaborative temperature control method and system for drying carbide slag addresses the challenges of localized overheating and structural instability in carbide slag accumulation. It utilizes a sensor network to collect pressure and temperature data in real time, generating load-bearing and temperature distribution maps. A convolutional neural network is then employed to analyze the probability of localized overheating and the risk of structural instability. If the risk exceeds a threshold, the invention optimizes the temperature input by adjusting the heating power, generating a heating distribution scheme. A support vector machine is used to assess its impact on drying efficiency, ensuring compliance with production safety standards. Subsequently, the invention updates the temperature and load-bearing distribution fields by executing the optimized scheme in real time. A recurrent neural network is then used to predict structural stability trends and overheating evolution paths, generating a dynamic control command sequence for continuous monitoring and regulation. Through multi-model collaboration and dynamic feedback, this invention significantly improves the safety and drying efficiency of carbide slag accumulation, reducing the risks of instability and overheating. Attached Figure Description

[0018] Figure 1 This is a flowchart of the temperature control method for drying carbide slag based on multi-agent collaboration according to the present invention;

[0019] Figure 2 This is a module connection diagram of the carbide slag drying temperature control system based on multi-agent collaboration of 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] like Figure 1As shown, this invention provides a technical solution: a multi-agent collaborative temperature control method for drying carbide slag, the method comprising: S1, collecting real-time pressure and temperature data through a sensor network deployed in the carbide slag accumulation area to obtain the current load-bearing distribution map and temperature distribution map of the carbide slag accumulation; S2, processing the data using a convolutional neural network model based on the load-bearing distribution map and temperature distribution map to determine the structural instability risk level in the carbide slag accumulation; S3, if the structural instability risk level is higher than a preset threshold, reducing the temperature input of the corresponding area by adjusting the power parameters of the heating method to obtain an optimized heating distribution scheme; S4, obtaining the optimized heating... After the distribution scheme is determined, a support vector machine model is used to analyze the impact of the scheme on the overall drying efficiency to determine whether it meets the efficiency requirements under the production safety standard. S5. If the efficiency requirements under the production safety standard are met, the scheme is implemented by adjusting the heating method in real time to obtain the updated temperature field and load-bearing distribution field of the carbide slag accumulation. S6. Based on the updated temperature field and load-bearing distribution field of the carbide slag accumulation, a recurrent neural network model is used to predict the structural stability change trend in the future time and determine the potential local thermal evolution path. S7. Through the potential local thermal evolution path, a targeted heating control command sequence is generated to obtain a dynamic monitoring cycle mechanism to maintain structural stability.

[0022] This implementation aims to improve the structural stability of carbide slag during the drying process. A sensor network is used to deploy pressure and temperature sensors in the carbide slag accumulation area to achieve real-time monitoring of the accumulation status. Two-dimensional or three-dimensional distribution maps are constructed from the collected temperature and load-bearing information and fed into a trained convolutional neural network model to identify high-risk areas and the overall structural stability level. When the system determines that there is a potential instability risk exceeding a safety threshold, a local temperature control optimization program is initiated to dynamically adjust the power output of the heating device, reducing the heat input to the corresponding high-risk area. Subsequently, the system uses a support vector machine model to evaluate the impact of this heating adjustment scheme on the overall drying efficiency, ensuring a balance between drying rate and safety. If the scheme meets safety standards, the adjusted temperature control strategy is applied in real time, updating the thermodynamic state of the accumulation area. Based on this, a recurrent neural network predicts future structural change trends based on the new temperature and load-bearing data, identifying potential local thermal evolution paths and providing decision-making basis for subsequent heating regulation, forming a complete closed-loop control chain. Finally, the system can generate a command sequence based on the prediction results to implement fine-grained control of the heating unit, ensuring the stability of the structure and dynamic optimization of the drying effect during the drying process.

[0023] Compared to traditional single-point heating or fixed-strategy heating systems, this implementation significantly improves structural safety and drying efficiency during the carbide slag drying process. By introducing multi-agent model collaborative control and machine learning algorithms, adaptive temperature control and risk perception are achieved during the drying process, effectively avoiding structural collapse or uneven heat distribution leading to decreased drying efficiency. Furthermore, the establishment of a dynamic monitoring and cyclic mechanism allows the system to adjust control strategies in real time to cope with changes in the field, exhibiting stronger robustness and environmental adaptability.

[0024] S1 includes: deploying a sensor network in a gridded manner in the carbide slag accumulation area to acquire pressure and temperature data, generating an initial dataset; using the initial dataset, removing noise data using a data cleaning method to obtain cleaned pressure and temperature data; using the cleaned pressure data, calculating the load-bearing distribution using an interpolation algorithm to generate a load-bearing distribution map of the carbide slag accumulation; using the cleaned temperature data, calculating the temperature distribution using an interpolation algorithm to generate a temperature distribution map of the carbide slag accumulation; if there are areas in the load-bearing distribution map where the pressure anomaly exceeds a preset threshold, then re-collecting regional data through the sensor network to obtain updated pressure data; if there are areas in the temperature distribution map where the temperature anomaly exceeds a preset threshold, then re-collecting regional data through the sensor network to obtain updated temperature data; and using the updated pressure and temperature data, fusing multiple sets of data using a weighted average algorithm to generate optimized load-bearing and temperature distribution maps.

[0025] In the implementation process, the carbide slag accumulation area is first deployed in a grid pattern using a sensor network. Specifically, the accumulation area is divided into several regular cell regions, and a set of pressure and temperature sensors is placed at the center of each cell. The grid density is set based on the overall area and structural complexity of the accumulation area. For example, if the accumulation area is 1,000 square meters, it is typically divided into 100 cells, each with an area of ​​10 square meters, ensuring comprehensive data coverage and sufficiently high spatial resolution. Each sensor collects data every five minutes, forming an initial dataset of pressure and temperature values. This dataset includes the current pressure and temperature values ​​at each grid location.

[0026] The collected initial dataset then proceeds to the data cleaning step. The specific method for data cleaning is as follows: First, a normal range of variation is defined for each type of sensor data. For example, the normal range for pressure values ​​is 0 to 500 kPa, and the normal range for temperature values ​​is 0 to 300 degrees Celsius. If a data point exceeds this range, it is marked as outlier and directly discarded. Then, median smoothing is performed on the remaining data. This involves sorting each data point with the values ​​of its eight surrounding grid points and taking the median value as the cleaned value for that point, thereby removing occasional spikes and interference to obtain the cleaned pressure and temperature data.

[0027] After obtaining the pressure data after cleaning, an interpolation algorithm is used to calculate the load-bearing distribution map. Specifically, for any point without sensors, the four nearest sensor data points are selected around that point. The pressure value of each point is calculated by weighting it according to its distance from the target point. The weighting principle is that the closer the distance, the higher the weight; the farther the distance, the lower the weight. The specific weights are determined according to an inverse relationship, i.e., proportional to the reciprocal of the distance. The value of each calculated point is the estimated pressure value at that location. The calculated values ​​of all points constitute the load-bearing distribution map of the carbide slag accumulation.

[0028] The generation process of temperature distribution map and load distribution Figure 1 Similarly, a four-point weighted interpolation calculation is used, with the data source being the cleaned temperature data. For the temperature value at any non-sensor location, a weight is calculated based on its distance from the four surrounding sensors, and then a weighted average is taken to calculate the temperature value at the target location, thus forming a complete temperature distribution map.

[0029] Subsequently, anomaly detection was performed on the load-bearing distribution map and temperature distribution map. The criteria for judging the degree of anomaly were as follows: in the load-bearing distribution map, if the pressure value of a certain area was more than 50% higher than the surrounding average, the area was considered to have a pressure anomaly; in the temperature distribution map, if the temperature of a certain area was more than 30 degrees Celsius higher than the surrounding average temperature, it was considered a temperature anomaly area. The above percentages and temperature differences are preset anomaly judgment thresholds, which are statistical averages derived from the analysis of historical accumulation and collapse cases, ensuring that potential anomalies can be detected while avoiding over-response.

[0030] Once an abnormal area is identified, the sensor network is immediately instructed to resample that area at a higher frequency. Specifically, this resampling involves deploying a higher density of sensors within the marked abnormal area or adjusting the sampling frequency of existing sensors to once per minute, with a sampling duration of ten minutes, to collect new pressure and temperature data.

[0031] After obtaining the updated regional data, a weighted average algorithm is used to merge the original cleaned data and the updated data. Specifically, different weights are assigned to the original and updated values ​​at the same location: the updated data has a weight of 70%, and the original value has a weight of 30%. The final optimized value is then calculated according to these weights. This method is used to process the data for all updated regions to obtain the final optimized load-bearing distribution map and temperature distribution map.

[0032] S2 includes acquiring load-bearing and temperature distribution maps of the carbide slag accumulation area through a sensor network to generate initial image data; standardizing the initial image data using a data preprocessing method to obtain standardized image data; extracting features from the standardized image data using a convolutional neural network to obtain feature vectors; calculating the probability of local overheating areas using a convolutional neural network based on the feature vectors to obtain a probability distribution; if there are areas in the probability distribution that are higher than a preset threshold, re-acquiring load-bearing and temperature data for that area through the sensor network to obtain updated image data; analyzing the updated image data using a convolutional neural network to calculate the structural instability risk level to obtain a risk assessment result; and classifying and labeling the carbide slag accumulation area using a region division method based on the risk assessment result to obtain a graded distribution map.

[0033] In this implementation, pressure and temperature data are first collected using a sensor network deployed in the carbide slag accumulation area. A two-dimensional grid structure is then constructed based on the spatial locations of the sensors. Each grid cell represents a sensor measurement point at a spatial location. Using the previously obtained load-bearing and temperature distribution maps, these are encoded into a two-dimensional matrix in image form. The load-bearing and temperature values ​​serve as two channels of information for the image, forming the initial image data. The resolution of this image data is consistent with the grid division, typically set to one pixel per meter. The specific image size depends on the actual area of ​​the accumulation area. For example, when the accumulation area is a rectangle 20 meters long and 10 meters wide, the image data size is 20 x 10 pixels.

[0034] Before being input into the neural network, the initial image data undergoes standardization. The standardization method is linear normalization, which involves linearly compressing the load-bearing and temperature values ​​of each pixel according to preset maximum and minimum values, mapping all pixel values ​​to floating-point numbers between 0 and 1. Specifically, the maximum load-bearing value is set to 500 kPa, and the minimum to 0 kPa; the maximum temperature value is set to 300 degrees Celsius, and the minimum to 0 degrees Celsius. Subtracting the minimum value from each pixel value and dividing by the difference between the maximum and minimum values ​​yields the standardized image data. This standardization step eliminates differences in data magnitude between different regions, ensuring good numerical stability of the neural network during processing.

[0035] Standardized image data is input into a convolutional neural network for feature extraction. The network structure consists of three convolutional layers and two pooling layers. The convolutional kernel size is fixed at 3x3 with a stride of 1. 2x2 max pooling is used, and the activation function is a uniform modified linear unit function. Through this network structure, local features are extracted from the input image, outputting a feature vector of length 256, representing the overall feature information of the temperature distribution and load-bearing structure in the image. This feature vector is input to the subsequent classification structure, a computational module consisting of fully connected layers and a softmax layer, to calculate the probability that each grid point is a local overheating region. The network's final output is a two-dimensional probability map with the same size as the input image, representing the probability value of each location being a local overheating region.

[0036] A threshold of 0.8 is set for determining local overheating. This means that if the overheating probability value of a grid point exceeds 0.8, that point is considered to have a serious overheating risk. This threshold is determined based on statistical data from historical accumulation accidents and field experience, exhibiting high sensitivity and a low false positive rate. Once an area exceeding the threshold is identified, the sensors within the corresponding grid area are immediately instructed to enter high-frequency sampling mode, sampling once per minute to continuously collect the latest pressure and temperature data for the next 5 minutes.

[0037] The newly collected data is re-standardized in the same way and converted into updated image data, which is then input back into the original convolutional neural network for risk analysis. After performing the aforementioned feature extraction and risk assessment processes on the updated image data, the network outputs the structural instability risk level. The risk level is defined as a continuous value between 0 and 1, and is divided into 5 risk level intervals based on this value: 0 to 0.2 is low risk, 0.2 to 0.4 is low-to-medium risk, 0.4 to 0.6 is medium risk, 0.6 to 0.8 is medium-to-high risk, and 0.8 to 1 is high risk. The risk level classification standard is based on the stability assessment model of carbide slag accumulation, and is comprehensively evaluated based on multiple characteristics such as regional thermal intensity gradient and local structural stress concentration to ensure the accuracy and guiding value of the level results.

[0038] Based on the aforementioned risk level results, a region division method is implemented, whereby the corresponding locations in the original image are marked with different colors according to the risk level of each grid point. The colors are set as follows: blue represents low risk, green represents low-to-medium risk, yellow represents medium risk, orange represents medium-to-high risk, and red represents high risk, forming the final graded distribution map. This image serves as the basis for subsequent temperature control adjustments, implementing corresponding heating power limitation strategies based on the risk level of different risk zones to ensure the overall safety and stability of the stacked structure.

[0039] S3 includes: acquiring temperature and load-bearing data of the carbide slag accumulation area through a sensor network to generate an initial distribution dataset; processing the initial distribution dataset using a data standardization method to obtain a standardized dataset; extracting features from the standardized dataset using a convolutional neural network to obtain feature vectors; if there are risk indicators in the feature vectors that are higher than a preset threshold, re-collecting temperature and load-bearing data of the corresponding area through the sensor network to generate an updated distribution dataset; analyzing the updated distribution dataset using a logistic regression model to determine the structural instability risk level; adjusting the heating power parameters according to the structural instability risk level to generate an optimized heating distribution scheme; and verifying the optimized heating distribution scheme through a sensor network to obtain a verification dataset and determine the stability of the temperature and load-bearing data.

[0040] In this implementation, temperature and load-bearing data are first collected from each grid cell using a sensor network deployed within the carbide slag accumulation area. Temperature sensors are installed at a density of one per square meter, and pressure sensors at a density of one per two square meters, ensuring comprehensive and dense data coverage. All sensors sample every 5 minutes. After a complete sampling cycle, the data collected by all sensors is organized according to grid location to construct an initial distribution dataset. Each data point corresponds to a grid location and includes two attributes: temperature and load-bearing value. This initial distribution dataset is then standardized using linear normalization. Specifically, the temperature range is set to 0°C to 300°C. Each temperature value is subtracted by 0°C and then divided by 300, resulting in a floating-point number between 0 and 1 as the standardized result. The load-bearing value range is set to 0 kPa to 500 kPa. This is done by subtracting 0 from each load-bearing value and then dividing by 500, also resulting in a standardized result between 0 and 1. This process ensures that all data maintains a uniform numerical scale, eliminating interference from different physical quantity units.

[0041] Next, the standardized dataset is input into the trained convolutional neural network model. This model consists of three convolutional layers, each using a 3x3 convolutional kernel for sliding computation to extract spatial features. Each kernel performs a weighted summation operation on the input data within its effective range with a fixed stride. The calculated local features are then processed by an activation function and output to the next layer. Following the convolutional layers are two max-pooling layers, each using a 2x2 sliding window to retain the maximum value within the region, thus reducing data dimensionality while preserving key features. The final output of the neural network is a feature vector of length 256, representing the coupling characteristics of the structural stress and thermal distribution in the entire stacked region at the current moment.

[0042] The feature vector is analyzed item by item, containing several risk indicators representing the risk of structural anomalies. If any indicator value exceeds a set risk threshold of 0.7, a risk area re-collection mechanism is triggered. This threshold of 0.7 is derived from statistical analysis of the neural network outputs of hundreds of cases of stacking anomalies and represents the critical probability value for structural instability. At this point, the sensor network is instructed to re-sample the grid area corresponding to the risk indicator at a high frequency, adjusting the sampling frequency to once every minute and the sampling duration to 5 minutes to ensure the acquisition of the latest and densest data updates. After collection, an updated distribution dataset is formed, whose structure is consistent with the initial distribution dataset, but with a higher sampling frequency.

[0043] The updated distribution dataset, after undergoing the same standardization process, is input into the established logistic regression model. The model uses the updated standardized temperature and load-bearing values ​​as independent variables to calculate the probability of structural instability. Internally, the model performs an inner product operation between the input vector and model parameters based on the training weights, and then uses the Sigmoid function to map the output to a continuous value between 0 and 1. This value represents the current risk level of structural instability. First, upon identifying a risk indicator in the feature vector exceeding a preset threshold, the sensor network is immediately instructed to re-collect temperature and load-bearing data for the corresponding area and its adjacent grids at a frequency of once per minute for 5 minutes, forming an updated distribution dataset. Each data point contains two fields: the temperature value (in degrees Celsius) and the load-bearing value (in kilopascals) at that grid location. After collection, the dataset is standardized by dividing the temperature value by the maximum set value of 300 degrees Celsius and the load-bearing value by the maximum set value of 500 kilopascals, yielding standardized temperature and load-bearing values ​​between 0 and 1, forming the standardized updated distribution dataset. Subsequently, this standardized dataset is input into the logistic regression model line by line. The logistic regression model is a binomial probabilistic model, consisting of two parts: a linear combiner and a probability mapping function. For each data point, the model first multiplies the standardized temperature and load-bearing values ​​by their corresponding weight parameters in the logistic regression model, sums the two products, and adds a constant bias term to form a linear combination value. This value is input as an independent variable to the Sigmoid function, which converts it into a continuous value between 0 and 1, representing the structural instability risk level corresponding to that data point. The Sigmoid function is a probability mapping function commonly used in logistic regression models; its function is to compress any real-valued input into a numerical range between 0 and 1, thereby converting the model's output into a probabilistic result. In practical applications, the standardized temperature and load-bearing values ​​are first weighted and summed with the weight coefficients obtained from model training, and then a bias term is added to form a linear combination value. The linear combination value serves as the input to the Sigmoid function. After nonlinear mapping, it outputs a floating-point number between 0 and 1. This value represents the probability assessment of whether the current input data falls under the risk category of "structural instability" or "insufficient drying efficiency." The Sigmoid function exhibits significant monotonicity and an S-shaped curve characteristic. When the input value is large, its output is close to 1, indicating that the risk is considered highly likely; when the input value is small, the output is close to 0, indicating that the risk is considered extremely small; when the input value is close to 0, the output value is 0.5, indicating no significant bias towards whether the risk will occur. This risk level represents the probability of structural instability occurring in this grid region under the current temperature and load-bearing conditions.The risk level of each point is divided into risk intervals: a risk value less than or equal to 0.2 is considered low risk; greater than 0.2 and less than or equal to 0.4 is considered low-to-medium risk; greater than 0.4 and less than or equal to 0.6 is considered medium risk; greater than 0.6 and less than or equal to 0.8 is considered medium-to-high risk; and greater than 0.8 is considered high risk. These grading standards are established based on retrospective data of historical structural instability events in industrial sites, combined with laboratory simulations of stacked structure evolution curves, ensuring that each risk level corresponds to a clear structural state boundary. The training samples for the logistic regression model come from pre-collapse data collected from past actual stacking operations. Fixed weights and bias parameters are obtained through supervised learning. The model's stability and discrimination accuracy are confirmed through cross-validation and field testing. In actual operation, a complete logistic regression prediction process is performed on each sample in the updated distribution dataset to obtain the structural instability risk level of all points in the entire grid map. Finally, a risk level result map with spatial distribution significance is generated to guide the design of subsequent heating power parameter adjustment schemes.

[0044] After determining the structural instability risk level, the heating power parameters are automatically adjusted. The initial heating power is 100% of full power. The heating power is reduced progressively according to the risk level, decreasing to 80% in high-risk areas, 90% in medium-to-high-risk areas, and maintaining the original power in medium-risk and lower-risk areas. The adjusted power parameters are then reorganized according to the grid positions to form an optimized heating distribution scheme. This scheme defines the power output percentage for each grid unit, which is then used by the heating device.

[0045] After implementing the optimized heating scheme, the sensor network is activated again to collect temperature and load-bearing values ​​under the post-heating state. The sampling frequency is restored to once every 5 minutes, and the sampling duration is 10 minutes, forming a verification dataset. The temperature fluctuation at each point in the verification dataset must not exceed ±5 degrees Celsius of the pre-optimization temperature value, and the load-bearing value fluctuation must not exceed ±20 kPa of the pre-optimization value. If the above stability criteria are met, the optimized heating scheme is considered successful, and the process proceeds to the next cycle of control. If not, the failed area is recorded, and the process returns to the risk level judgment step. The heating power in that area is readjusted, and iterative optimization continues until the data fluctuation is within a stable threshold, ensuring that both structural safety and thermal efficiency meet the standards.

[0046] A stability assessment mechanism based on sensor data was designed, and objective quantitative analysis of temperature and load-bearing data was performed through multi-dimensional data processing and evaluation methods to ensure the feasibility and objectivity of the assessment results. First, after the system executes the optimized heating scheme, a sensor network deployed within the carbide slag accumulation area continuously collects real-time temperature and load-bearing data from multiple key areas. The acquisition frequency and period are set by the control platform, for example, once per second for ten consecutive minutes, generating the dataset required for stability verification. After data acquisition, the system analyzes the temperature and load-bearing changes at each monitoring node in chronological order. For temperature data, the system determines whether the rate of change is stable throughout the observation period, i.e., whether the temperature fluctuates drastically within a unit of time. Similarly, for load-bearing data, the system analyzes the magnitude of change within a unit of time to determine if there are drastic increases or decreases. If temperature or load fluctuates rapidly within a short period, it indicates thermal or mechanical instability in the area. Next, the system performs statistical analysis on the degree of fluctuation in temperature and load-bearing data. Specifically, it calculates the temperature and pressure change ranges for each monitoring point throughout the observation period to determine whether they are within a preset stable range. If the data from a monitoring point shows minimal fluctuation over time, it indicates that the point is relatively stable; otherwise, it is considered to pose a potential risk. Simultaneously, to avoid local anomalies affecting the overall judgment, the system also performs horizontal comparisons of data differences between adjacent monitoring points. If the temperature or load-bearing value of a monitoring point differs significantly from multiple surrounding monitoring points, it indicates a potential concentration of local anomalies and will be marked as an unstable area. After completing the above analysis, the system generates a stability assessment result for each monitoring point from multiple dimensions. These dimensions include the rate of temperature change, the rate of load-bearing change, data volatility, and spatial consistency, each with its own set of judgment criteria and stability thresholds. The system scores the results of each dimension and comprehensively scores the stability of the entire carbide slag accumulation area. If the overall score reaches the preset safety standard, it indicates that the current temperature and load-bearing status are stable, and the optimized heating scheme is acceptable; otherwise, the system will automatically adjust the heating power parameters and repeat the data acquisition and stability verification process until the stability assessment meets the standard. Through the above methods, the system achieves continuous monitoring, stability judgment and quantitative analysis of the thermal state of the carbide slag accumulation structure, ensuring that the heating process meets the drying efficiency while ensuring structural safety.

[0047] S4 includes: acquiring temperature control parameters and load distribution data corresponding to the optimized heating distribution scheme through a sensor network to obtain an initial efficiency dataset; processing the initial efficiency dataset using a data standardization method to obtain a standardized efficiency dataset; performing classification analysis on the standardized efficiency dataset using a support vector machine model to obtain drying efficiency data; if the drying efficiency data is lower than the efficiency requirements preset by the production safety standard, re-collecting temperature control parameters and load distribution data through the sensor network to obtain an updated efficiency dataset; analyzing the updated efficiency dataset using a logistic regression model to determine risk assessment indicators; adjusting the temperature control parameters according to the risk assessment indicators to obtain an optimized adjustment strategy; and verifying the optimized adjustment strategy through a sensor network to obtain a verification dataset to determine whether the drying efficiency data meets the production safety standard.

[0048] In this implementation, firstly, based on the optimized heating distribution scheme implemented in the previous stage, the latest temperature control parameters and load-bearing distribution data are collected from temperature and pressure sensors deployed in the carbide slag accumulation area. The sampling frequency is set to once every 5 minutes, covering all heating control units and corresponding structural units throughout the entire accumulation area. The temperature control parameters are the real-time output power values ​​of each heating unit, expressed as a percentage, ranging from 0 to 100; the load-bearing distribution data are the actual load-bearing values ​​corresponding to each grid unit, expressed as kPa, ranging from 0 to 500. After the above two types of data are collected, an initial efficiency dataset is constructed according to spatial correspondence. Each sample contains two fields: the percentage of temperature control power and the load-bearing value at the corresponding point.

[0049] The initial efficiency dataset then undergoes a standardization process. Standardization employs linear normalization, dividing the temperature-power percentage by 100 to obtain standardized temperature values ​​between 0 and 1; and dividing the load-bearing value by 500 to obtain standardized load-bearing values ​​between 0 and 1, thus forming the standardized efficiency dataset. This process ensures that data of different scales can be simultaneously input into the support vector machine model, avoiding learning bias caused by differences in magnitude.

[0050] Standardized datasets are input into a pre-trained support vector machine (SVM) model for classification analysis. This SVM model constructs a nonlinear interface in a high-dimensional space using a Gaussian radial basis function to determine whether each set of input data meets the drying efficiency requirements. The model's training objective is to determine whether the current temperature control parameters and load conditions are sufficient to reduce the moisture content of the carbide slag from the initial 30% to below 10% within 48 hours. This determination is based on drying curves measured under actual laboratory and industrial production conditions. The model output label indicates whether the efficiency requirements are met or not.

[0051] A standardized efficiency dataset was prepared as input. This dataset consists of multiple samples, each containing two numerical fields: standardized temperature control parameter values ​​and standardized load distribution values. The temperature control parameter values ​​range from 0 to 1, corresponding to the original temperature power percentage divided by 100; the load distribution values ​​range from 0 to 1, corresponding to the original pressure values ​​divided by 500. Standardization ensures that all input data have consistent dimensions and values ​​within a uniform scale, which is beneficial to the stability and accuracy of model calculations. Subsequently, this standardized dataset was input into a pre-trained support vector machine model. This model uses a Gaussian radial basis function as the kernel function, the core purpose of which is to construct a hyperplane in high-dimensional space to classify data samples into two categories: one category representing "meets drying efficiency requirements" and the other representing "does not meet drying efficiency requirements." During the model training phase, a large amount of labeled historical data was used for learning, including labels for drying results under different combinations of temperature power and structural load. The labels were defined as follows: after the carbide slag drying process lasted for 48 hours, if the moisture content of the target area dropped below 10%, it was marked as "meets requirements"; if it was above 10%, it was marked as "does not meet requirements." During the model inference phase, each standardized input sample is first multiplied by the model parameter vector, and then the model bias is added. The result is fed into a kernel function mapping to construct a high-dimensional spatial distance metric. The model determines the category of a sample based on its distance from the interface. The final output is a binary classification result: if a sample falls into the "satisfied" category, the output label is 1, indicating that the parameter combination has the ability to meet the drying efficiency requirements under the current conditions; if it falls into the "unsatisfied" category, the output label is 0, indicating insufficient efficiency. This process performs classification analysis on all samples in the entire dataset one by one, and the sum of all output labels forms the drying efficiency data. These results are statistically analyzed, for example, by calculating the percentage of samples with an output of 1, and determining whether it is greater than 80% as the overall judgment criterion. If the percentage standard is met, it indicates that the overall heating strategy meets the drying efficiency target under the production safety standard; if it is not met, the efficiency is considered too low, and subsequent optimization steps are required.

[0052] If the model output does not meet the efficiency requirements, a data update process is initiated. Temperature control parameters and load distribution data for the same area are collected again via the sensor network at a higher frequency. The sampling frequency is increased to once every 2 minutes, and the sampling duration is 10 minutes to ensure the data contains sufficient information on temporal variations. The new data is used to construct an updated efficiency dataset, which, after being processed using the same linear normalization method, is input into the logistic regression model for risk assessment.

[0053] The logistic regression model calculates a weighted sum of standardized temperature and load values ​​with model parameters, then maps the result to a risk probability value between 0 and 1 using a sigmoid function. This value serves as the risk assessment index. This index reflects the potential impact of adjustments to the current heating scheme on the stability of the stacked structure. Based on empirical parameter settings, when the index is less than 0.3, it indicates good structural stability, and the current heating power can be increased by 10%; when the index is between 0.3 and 0.6, it indicates a certain degree of structural sensitivity, and the heating power can be increased by 5%; when the index is higher than 0.6, the structure enters a high-risk state, and the heating power remains unchanged to avoid further exacerbating structural load concentration. First, based on the latest temperature control parameters and load distribution data obtained from the aforementioned supplementary sampling steps, an updated efficiency dataset is constructed. Each data record includes two fields: one is the temperature control parameter, ranging from 0 to 100, corresponding to the current output power percentage of each heating unit; the other is the load value, in kPa, ranging from 0 to 500, corresponding to the current structural bearing pressure of each grid region. The updated efficiency dataset was then standardized using linear normalization. Temperature control parameters were divided by 100, and load-bearing values ​​by 500, converting them into standardized values ​​between 0 and 1, generating the standardized updated efficiency dataset. The standardized data were then input into an established logistic regression model for analysis. The logistic regression model is a binary classification probability prediction model with a weighted linear combiner structure. The calculation process is as follows: for each input sample, the standardized temperature and load-bearing values ​​are multiplied by two weight parameters obtained from model training. The results are summed and then added to a bias parameter to obtain an intermediate value. This intermediate value is used as an independent variable input to the Sigmoid function, which maps it to a floating-point number between 0 and 1. The output is the risk assessment index. This risk assessment index represents the safe margin for adjusting heating power while maintaining structural stability under the current combination of input parameters. A lower assessment value indicates higher structural stability and a larger allowable power adjustment space; a higher assessment value indicates that the structure is on the edge of high risk, and further increases in temperature and power must be strictly limited. Quantitative zoning is performed based on the risk assessment index values: if the risk assessment value is less than 0.3, it indicates low structural risk, and the current temperature and power can be increased by 10%; if the assessment value is between 0.3 and 0.6, it indicates moderate structural sensitivity, and the temperature and power can only be increased by 5%; if the assessment value is higher than 0.6, it is determined to be high structural risk, and power cannot be increased, and the current heating level should be maintained.

[0054] Based on the risk assessment indicators output by the logistic regression model, the temperature control parameters at each grid location are adjusted to generate a new temperature control configuration table. The adjustment results are then reorganized into an optimization strategy. This strategy applies control commands to each heating unit to achieve a new power output configuration. A sensor network is simultaneously activated to continuously collect temperature control parameters and load distribution data over 24 hours at a frequency of once every 5 minutes, constructing a validation dataset.

[0055] The validation dataset is then input back into the original support vector machine model, and the same efficiency judgment process as described above is executed. If the model determines that the efficiency requirements are met, it means that the current optimization and adjustment strategy has met the drying performance target under the premise of structural safety; if it still does not meet the requirements, the current risk parameters are retained, and the logistic regression evaluation, power adjustment, and validation process are executed again until the efficiency judgment meets the production safety standards, ensuring that the final implementation plan has safety, feasibility, and stability in actual industrial operation.

[0056] S5 includes: acquiring updated temperature field and load-bearing distribution field data of carbide slag accumulation through a sensor network to obtain an initial field distribution dataset; processing the initial field distribution dataset using a data standardization method to obtain a standardized field distribution dataset; performing classification analysis on the standardized field distribution dataset using a random forest model to obtain feature classification results for the temperature field and load-bearing distribution; if the feature classification results show that the temperature field or load-bearing distribution deviates from a preset stability threshold, then re-acquiring temperature field and load-bearing distribution field data through the sensor network to obtain an updated field distribution dataset; performing dimensionality reduction processing on the updated field distribution dataset using principal component analysis to obtain a dimensionality-reduced feature dataset; analyzing the stability of the temperature field and load-bearing distribution using a logistic regression model based on the dimensionality-reduced feature dataset to obtain a stability assessment result; if the stability assessment result meets production safety standards, then adjusting the heating distribution parameters in real time through the sensor network to obtain optimized temperature field and load-bearing distribution field data.

[0057] In this implementation, a network of temperature and pressure sensors is first deployed within the carbide slag accumulation area. At a density of at least one temperature sensor per square meter and one pressure sensor per two square meters, sampling data is collected at a 5-minute interval for each grid location within the accumulation area. Each grid cell contains two values: a temperature value (in degrees Celsius, ranging from 0 to 300) and a load-bearing value (in kilopascals, ranging from 0 to 500). All data are arranged spatially to form an initial field distribution dataset. This dataset is then standardized using linear normalization, dividing each temperature value by 300 and each load-bearing value by 500 to convert them into standardized data between 0 and 1, generating a standardized field distribution dataset. This standardization process aims to unify the numerical range and prevent differences in the numerical values ​​of different physical quantities from affecting the model's judgment results during subsequent modeling and analysis.

[0058] Subsequently, the standardized field distribution dataset was input into a random forest classification model. This random forest model consists of 100 decision trees. During the training phase, each decision tree randomly extracts feature combinations from the samples, forming multiple feature discrimination paths. During analysis, each tree independently makes path judgments on the standardized temperature and load-bearing values ​​of a specific grid point, outputting a "stable" or "unstable" result. The model ultimately uses majority voting to determine the judgment result for each grid point. The preset stability thresholds used by the model are: temperature variation range should not exceed 5 degrees Celsius, meaning the temperature fluctuation at a single point within the sampling period should not exceed ±5 degrees Celsius of the previous sampling value; and load-bearing fluctuation range should not exceed 20 kPa, meaning the difference between the current load-bearing value and the historical average load-bearing value should not exceed ±20 kPa. These two stability thresholds were derived through statistical analysis of the fluctuation ranges of 200 stable stacking operation records over the past 3 years, ensuring the reliability and applicability of the judgment results.

[0059] If the random forest model determines that the temperature field or load-bearing distribution deviates from the aforementioned stability threshold, a high-frequency supplementary sampling mechanism is executed. The sampling frequency is increased from once every 5 minutes to once every 2 minutes, the sampling duration is 10 minutes, and the sampling area covers all grid cells determined to be unstable and their eight adjacent grid cells. The sampled data again constitutes the updated field distribution dataset and is converted to a standardized format using the same standardization method.

[0060] The feature classification results include the following categories: First, in terms of temperature field, the feature classification results include the following: (1) Local high temperature area identification: The system detects that the temperature value of some areas is significantly higher than that of the surrounding areas or the historical average temperature, and is initially classified as local overheating risk; (2) Temperature gradient abnormal area identification: There is an abnormal temperature difference between adjacent areas, such as a sudden rise or fall in temperature, which may lead to uneven thermal expansion, structural deformation and other hidden dangers; (3) Temperature fluctuation area identification: The temperature of a certain area fluctuates frequently in a short period of time, which may mean that the heating control is unstable or the heat transfer is abnormal; (4) Temperature uniformity level assessment: Whether the overall temperature field is uniformly distributed. If there is a large-scale thermal unevenness, it may affect the drying efficiency or the safety of the local structure.

[0061] Secondly, in terms of load distribution, the feature classification results mainly include: (1) Local high pressure area identification: The system identifies that the load in some areas is significantly higher than that in other areas, indicating that there are concentrated stress points, and it is necessary to be alert to the risk of local settlement or structural damage; (2) Uneven load area identification: The load difference within the area is large, and there is a serious problem of uneven stress distribution, which can easily lead to local slippage or compaction abnormalities; (3) Stress change area detection: The load change is drastic between adjacent areas, indicating that there may be sudden settlement, suspension, fracture or structural misalignment; (4) Structural stress symmetry analysis: The system performs geometric symmetry assessment on the stress of the overall stacked area. If obvious eccentricity or unilateral concentrated stress is found, it indicates that the overall stability of the structure may decrease.

[0062] In addition, the classification results also include cross-dimensional comprehensive identification, such as: (1) thermal coupling anomaly zone identification: areas with both high temperature and high pressure characteristics are marked as high coupling risk areas and need to be monitored; (2) structural edge risk zone determination: the temperature and load-bearing characteristics of the edge position are classified separately to assess whether there are problems with poor edge heat dissipation or weak support; (3) interference anomaly identification: if extreme values ​​or data continuity anomalies occur, the classification model will mark them as possible data acquisition failures or external disturbances, prompting that resampling is required.

[0063] The updated standardized dataset is then input into the principal component analysis (PCA) module to perform dimensionality reduction. The specific steps are as follows: First, covariance is calculated for all sample matrices in the updated dataset to generate a covariance matrix; then, eigenvalue decomposition is performed on the covariance matrix to extract principal component vectors; based on the principle that the cumulative contribution rate is greater than 95%, the first few principal components are selected, typically the first 3 to 5 principal components, as feature data, forming a dimensionality-reduced feature dataset. This dimensionality reduction step aims to reduce model complexity and redundant information while retaining the main feature information that effectively expresses the trends of temperature and load-bearing changes.

[0064] The dimensionality-reduced feature dataset is then input into a logistic regression model for stability probability prediction. The principal component feature vector of each sample is multiplied term-by-term with the weight coefficients in the logistic regression model, summed, and then a bias term is added to form a linear combination value. This value is then input into the Sigmoid function to convert it into a probability value between 0 and 1, called the stability assessment index. This value represents the probability level that the point is in a structurally stable state under the current temperature and load-bearing conditions. The stability threshold is set to 0.7. If the stability index of a grid point is higher than or equal to 0.7, the point is considered to be in a safe and stable state; otherwise, it is considered to have insufficient stability. This threshold is set based on the physical model of the stacked structure and industrial safety boundary experience to ensure that the classification results neither miss high-risk points nor misjudge normal states.

[0065] After stability assessment of all grid points, if the overall stability meets production safety standards (i.e., all point indices are not lower than 0.7), the control execution process is immediately initiated. Based on the stability indices of each region, a differentiated heating parameter adjustment strategy is implemented: if the stability indices of all grids in a region are higher than 0.85, the heating power can be increased by 10%; if the stability indices are between 0.7 and 0.85, the original power remains unchanged; if individual grids are close to the threshold (i.e., between 0.7 and 0.75), the heating power is reduced by 10%. A new heating parameter configuration file is generated based on the above strategy, and adjustment commands are sent to each heating unit via the network to achieve real-time dynamic adjustment of the heating power. Subsequently, the sensor network continues to monitor the temperature field and load distribution under the new heating state every 5 minutes to construct the initial field distribution dataset for the next round, achieving closed-loop control and real-time response.

[0066] The stability assessment result refers to a comprehensive judgment made by the system regarding the current thermal stability state of the structure after multi-dimensional analysis of temperature field distribution data and load-bearing distribution data of the carbide slag accumulation area. The goal of this judgment is to identify whether the current heating process will cause structural instability, local instability, or other safety risks, and whether further adjustments to the heating strategy are needed to ensure continuous and stable drying.

[0067] After implementing the optimized heating scheme, the system acquires the latest temperature field distribution data and load-bearing field distribution data of the area through a sensor network deployed in the carbide slag accumulation area. These data reflect the thermal and stress states at different spatial locations, forming a complete, multi-dimensional "field" information map.

[0068] First, a random forest model is used to perform preliminary classification analysis on the above data. This model extracts features from the temperature and load-bearing values ​​at each current location point and determines whether there are risk areas exceeding the "normal stable range." Its output will clearly indicate whether there are local areas with significantly higher or lower temperatures or pressures than other areas, potentially leading to localized thermal expansion, structural displacement, or stress concentration.

[0069] If abnormally high temperatures or uneven load distribution are detected in certain areas during this stage, it indicates a potential unstable trend. To further confirm whether this abnormal state persists, the system will conduct a second round of data collection on these abnormal areas, acquiring more concentrated or higher-frequency temperature and load data to construct an "updated field distribution dataset." Because the original data has a very high dimensionality (i.e., each area may contain multiple time points, multiple data types, and multiple physical indicators), the system uses principal component analysis to reduce the dimensionality of the updated data. The purpose of this step is to extract the core factors or features that best represent the overall trend, making subsequent judgments more accurate and efficient.

[0070] After dimensionality reduction, the system will use a logistic regression model to comprehensively analyze these core features and determine whether the current temperature and load distribution show a "stable trend" or an "unstable trend." The judgment is based on the following aspects: temporal continuity: whether the temperature and load data have changed significantly in a short period of time; spatial consistency: whether there are drastic differences between adjacent points in the same area; trend direction: whether the temperature or load has been rising or falling continuously; structural symmetry: whether the force on the entire platform is balanced; historical comparability: whether the current data shows a deteriorating trend compared to the previous operating state.

[0071] The system will ultimately output a clear stability assessment result. This result can be a qualitative judgment (such as "stable" or "unstable"), a multi-level classification (such as "safe", "metastable", "critical" or "instable"), or a score or rating value.

[0072] S6 includes acquiring temperature field data and load-bearing distribution field data of carbide slag accumulation through a sensor network to obtain a real-time distribution dataset; processing the real-time distribution dataset using a data standardization method to obtain a standardized distribution dataset; analyzing the standardized distribution dataset using a recurrent neural network model to predict the structural stability change trend and obtain a stability trend prediction result; if the stability trend prediction result shows that the local thermal risk is higher than a preset threshold, then re-acquiring temperature field data through the sensor network to obtain an updated temperature dataset; grouping the updated temperature dataset using a clustering analysis method to determine the distribution of areas with local overheating risk higher than the preset threshold, obtaining a hot region dataset; judging the local overheating path based on the hot region dataset using a preset threshold to obtain the thermal evolution path distribution; and optimizing the thermal evolution path distribution by adjusting the heating distribution parameters to obtain optimized temperature field data and load-bearing distribution field data.

[0073] In this implementation, a network of temperature and pressure sensors deployed in the carbide slag accumulation area is first used to collect temperature and load-bearing data for each grid point within the entire accumulation area every 5 minutes, constructing a real-time distributed dataset. Temperature data for each grid point is in degrees Celsius, with a maximum value set to 300 degrees Celsius; load-bearing data is in kilopascals, with a maximum value set to 500 kilopascals. After data collection, each data point in the dataset is standardized by dividing the temperature value by 300 and the load-bearing value by 500, converting all data into floating-point values ​​between 0 and 1, thus forming a standardized distributed dataset. This standardization process ensures data dimensionality uniformity, which is beneficial to the stability and accuracy of subsequent model analysis.

[0074] Next, the standardized distribution dataset is fed into a pre-trained recurrent neural network model. This model employs a two-layer long short-term memory network structure, with each layer containing 128 neurons, used to analyze time-series data from the past 48 hours and predict structural evolution trends over the next 12 hours. The input data consists of temperature and load-bearing data sequences collected every 5 minutes over the previous 48 hours, and the output is the hourly rate of temperature change and load-bearing growth rate. The set temperature change risk threshold is a temperature rise exceeding 5 degrees Celsius per hour, and the load-bearing growth risk threshold is a pressure increase exceeding 30 kPa per hour. If the temperature or load-bearing growth rate at any grid point in the prediction results exceeds these thresholds, a high-risk local thermal evolution trend is considered to exist, and a high-frequency data re-collection process is initiated.

[0075] After the supplementary data collection process is initiated, temperature data is collected from the aforementioned high-risk area and its eight adjacent grid points at a frequency of once per minute for 10 minutes. After the supplementary collection is completed, the data undergoes the same standardization process to form an updated temperature dataset. This dataset is then input into the clustering analysis module for group identification. The clustering analysis employs a density-based spatial clustering algorithm, specifically finding adjacent points within a 2-meter radius for each data point and dividing them into independent cluster regions based on the density of the points. The defined threshold for local overheating is: if at least 80% of the points in a cluster region have a temperature value higher than 250 degrees Celsius, and the temperature change trend is positively increasing, then the region is marked as a "locally overheated area".

[0076] After identifying high-temperature regions, a thermal evolution path distribution is further constructed. This process is achieved using the adjacent grid tracking method. Starting with the grid with the highest temperature within the thermal region, each point is checked to see if its eight adjacent grids exhibit a temperature increase with a time-series temperature rise rate exceeding 3 degrees Celsius per hour. If these conditions are met, the point is included in the thermal evolution path. This method is continuously expanded outwards until the path expansion conditions are no longer met, ultimately forming one or more continuous thermal evolution path distribution maps. This map illustrates the diffusion trend and direction of local heat within the accumulation region, serving as a crucial basis for formulating control strategies.

[0077] Based on the thermal evolution path distribution map, heating parameters are optimized and adjusted for heating units within the path. The adjustment strategy is as follows: if a grid is at the starting point of the thermal evolution path and the current heating power exceeds 80%, the power at that point is reduced to 60%; if it is in the middle of the thermal evolution path and the temperature is still rising, the power is reduced to 50%; if it is at the end of the path, the power remains unchanged; no adjustments are made to other areas not covered by the path. All adjustment results are organized into an optimized heating parameter configuration table according to grid location, and the control module sends it to each heating execution unit in real time for implementation.

[0078] After adjustments are made, the optimized temperature field and load-bearing distribution field are continuously monitored at a sampling frequency of once every 5 minutes, and the complete dataset is updated. The new data is compared with the evolution path distribution to determine whether the temperature control intervention effect has been achieved. If the thermal evolution path terminates, it indicates that the heating optimization strategy is effective; otherwise, the next round of adjustments is initiated.

[0079] S7 includes: acquiring temperature distribution data of local overheated areas through a sensor network to obtain a real-time temperature dataset; grouping the temperature distribution using K-means clustering analysis based on the real-time temperature dataset to determine the distribution of areas where the local temperature exceeds a preset threshold, thus obtaining an overheated area dataset; identifying temperature anomalies in the overheated area dataset using the preset threshold to generate an initial heating control command sequence, thus obtaining a control command dataset; analyzing the control command dataset using a long short-term memory network model to predict the structural stability change trend, thus obtaining a stability trend prediction result; if the stability trend prediction result shows that the local thermal risk is higher than the preset threshold, re-collecting temperature distribution data and load-bearing distribution data through the sensor network to obtain an updated distribution dataset; classifying the evolution path of the overheated area using a support vector machine model based on the updated distribution dataset, thus obtaining an optimized overheating evolution path distribution; and generating a dynamically adjusted heating control command sequence based on the optimized overheating evolution path distribution to determine the real-time monitoring cycle mechanism.

[0080] In this implementation, a temperature sensor network deployed in the carbide slag accumulation area is first used to collect temperature data from all grids within the local overheating risk area at a frequency of once per minute, constructing a real-time temperature dataset. Each record in this dataset contains a grid location number and its corresponding temperature value, in degrees Celsius, ranging from 0 to 300. After data collection, the K-means clustering analysis method is applied to the dataset to group the temperature distribution. During clustering analysis, the number of clusters K is initially set to 5, a value determined based on the common temperature gradient stratification characteristics and thermal diffusion range divisions observed in actual working conditions. In the initial stage of the K-means algorithm, five temperature values ​​are randomly selected from the dataset as cluster centers, and then each data point is classified according to its Euclidean distance from the cluster center. After all points are classified, the center temperature of each category is recalculated, and the classification operation is repeated until the classification of all points no longer changes, ultimately forming five temperature distribution regions. Groups with cluster centers above 250 degrees Celsius are marked as "overheated regions," and the relevant grids are extracted to form an overheated region dataset. The 250-degree Celsius threshold is set based on the safe upper limit of temperature measured during previous drying processes of carbide slag. This value is the highest temperature at which long-term heating will not cause damage to the internal structure under the premise of structural stability.

[0081] Subsequently, temperature anomalies within the overheated region are identified based on two criteria. The first criterion is whether the current grid temperature exceeds 270 degrees Celsius; this threshold is a safe upper limit determined statistically from boundary data of thermal instability experiments on structural materials. The second criterion is whether the temperature at this point has risen by more than 10 degrees Celsius in the last 5 minutes. When both conditions are met, the grid point is marked as a temperature anomaly, and an initial heating control command is generated for it. The command includes: the grid number, the current heating power value (in percentage form), the suggested control action (e.g., reducing power by 20%), and the command timestamp. Control commands for all anomalies are compiled into a control command dataset.

[0082] The dataset of regulatory commands is input into a pre-trained Long Short-Term Memory (LSTM) neural network model for structural stability trend prediction. The model consists of a two-layer LSTM network and a fully connected output layer. The inputs are historical temperature sequences, heating power adjustment records, and load-bearing changes for each grid cell. The model predicts the probability of structural stability for each grid cell within the next 6 hours, with the output being a floating-point number between 0 and 1. A stability risk threshold of 0.3 is set; this value is optimized based on historical failure prediction accuracy. It indicates that if the predicted stability probability for a grid cell is below 0.3, there is a high risk of structural thermal instability in that area. If any grid cell in the model output has a stability probability below this threshold, a supplementary data collection process is immediately initiated, re-collecting temperature and load-bearing values ​​for the area over the past 10 minutes at a frequency of once per minute, generating an updated distribution dataset.

[0083] The updated distribution dataset is input into a Support Vector Machine (SVM) model for evolutionary path classification analysis. The SVM model uses a radial basis function kernel for high-dimensional mapping and has been trained in a previous supervised learning process using real overheating path data as samples. The input for each sample is a temperature value, a load-bearing value, and its rate of change; the output is a classification label indicating whether the point belongs to a thermal evolution path. After processing each data point, the model outputs the classification result, constructs path segments based on spatial adjacency for all grid points classified as "belonging to a thermal evolution path," and merges them to form an optimized overheating evolution path distribution map.

[0084] Finally, a dynamically adjusted heating control command sequence is regenerated based on the optimized path results. This sequence is determined as follows: if a grid is the starting point of the path, its original heating power is reduced by 30%; if it is located in the path extension zone, the power is reduced by 20%; if it is at the end of the path or in an area not included in the path, the current power remains unchanged. All commands are combined to form a new control command set, which is distributed to each heating unit in real time by the central control unit. The control is set to a 10-minute cycle. After each round of control execution, new temperature and load data are collected again, and the judgment and prediction process is restarted, forming an adaptive closed-loop control mechanism based on the optimized path. This ensures that local thermal anomalies are identified and intervened in a timely manner, thereby maintaining the thermal stability and safe operation of the entire stacked structure.

[0085] like Figure 2 As shown, a multi-agent collaborative temperature control system for drying carbide slag is also provided to implement the steps of the multi-agent collaborative temperature control method for drying carbide slag. The system includes a sensor network module that collects real-time pressure and temperature data through a sensor network deployed in the carbide slag accumulation area to obtain the current load-bearing distribution map and temperature distribution map of the carbide slag accumulation; an image processing module that processes the image data using a convolutional neural network model based on the load-bearing distribution map and temperature distribution map to determine the structural instability risk level in the carbide slag accumulation; and a heating control module that, if the structural instability risk level is higher than a preset threshold, reduces the temperature input of the corresponding area by adjusting the power parameters of the heating method to obtain an optimized heating distribution scheme; efficiency. The evaluation module, after obtaining the optimized heating distribution scheme, uses a support vector machine model to analyze the impact of the scheme on the overall drying efficiency and determine whether it meets the efficiency requirements under production safety standards. The execution control module, if the efficiency requirements under production safety standards are met, executes the scheme by adjusting the heating method in real time to obtain the updated temperature field and load-bearing distribution field of the carbide slag accumulation. The stability prediction module, based on the updated temperature field and load-bearing distribution field of the carbide slag accumulation, uses a recurrent neural network model to predict the structural stability change trend in the future and determine potential local thermal evolution paths. The instruction generation module, through the potential local thermal evolution paths, generates a targeted heating control instruction sequence to obtain a dynamic monitoring cycle mechanism to maintain structural stability.

[0086] In this system implementation, the sensor network module first performs real-time data acquisition. All temperature and pressure sensors are connected to the central processing unit via wireless or wired communication, and synchronously acquire data every 5 minutes. Each sensor corresponds to a physical grid cell in the stacked area. The system automatically maps the acquired data into a two-dimensional temperature distribution map and a load distribution map. The image processing module receives the distribution map data and uses a convolutional neural network model to extract image features. Through multi-layer convolution, pooling, and fully connected structures, it analyzes the areas of local abnormal heating or structural concentration in the image and outputs a risk level, a floating-point number between 0 and 1, representing the probability of current structural instability. When the risk level exceeds the system-set threshold of 0.7, the heating control module is activated. Based on the high-risk areas marked in the image, it identifies the corresponding heating units and reduces their output power by 20% to 30% by controlling the heating power parameters, generating an optimized heating distribution scheme.

[0087] The efficiency evaluation module receives the heating distribution scheme and, combined with the drying progress data within the current temperature control period and sampling cycle, inputs it into a pre-trained support vector machine model. The model outputs a judgment of "meets" or "does not meet the efficiency requirements under the production safety standard." The system's efficiency standard is to reduce the moisture content to no more than 10% within 48 hours. When the output is "meets," the execution control module distributes the optimized heating scheme to each heating unit. The controller adjusts the temperature output in real time according to the new configuration, while the sensor network continuously feeds back the latest temperature field and load-bearing field change data, forming a new closed loop.

[0088] The updated field distribution is input into the stability prediction module. This module uses a recurrent neural network model to perform time-series analysis on data from multiple consecutive periods. The model structure includes long short-term memory units, and the output is the probability value of structural stability for each region over the next 6 hours. If the future temperature rise rate in a local area within the predicted path exceeds 5 degrees Celsius per hour or the load fluctuation exceeds 30 kPa, the system identifies it as a potential thermal evolution path. The instruction generation module then generates a heating control instruction sequence based on this path structure, combined with the system's heating unit location and adjustment capabilities. Each instruction includes the unit location, current status, suggested control amplitude, and execution period. All instructions are sent to the execution module through the control channel to complete dynamic control. The control system updates the execution strategy every 10 minutes, achieving a closed-loop cycle of "sensing—identification—prediction—adjustment," ensuring dynamic coordinated operation of drying efficiency and stacked structure safety.

[0089] In this embodiment, a "locally overheated area" refers to a specific region in the temperature field data collected by a sensor network, identified by cluster analysis methods (such as K-means), where the temperature is higher than the overall temperature distribution, representing an abnormal temperature distribution phenomenon under the current state. "Local thermal risk exceeding a preset threshold," on the other hand, is a risk assessment result obtained by analyzing the trend of the stability of the carbide slag accumulation structure over a future time period using a predictive model (such as a recurrent neural network or support vector machine) based on the aforementioned temperature and load-bearing data. It characterizes the unsafe possibility of structural instability or thermal evolution in this region. Therefore, the two are not technically equivalent. "Locally overheated area" focuses on the anomaly identification of the current spatial state, while "local thermal risk exceeding a preset threshold" is a dynamic prediction of the future evolution trend of the abnormal region. A causal relationship can exist between the two, but they cannot replace each other. In implementation, the system typically first identifies locally overheated areas, then assesses their thermal risk level, and thus decides whether to optimize heating control commands and dynamically monitor and adjust them.

[0090] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for temperature control in drying carbide slag based on multi-agent collaboration, characterized in that, The method includes: S1. Real-time pressure and temperature data are collected by a sensor network deployed in the carbide slag accumulation area to obtain the current load-bearing distribution map and temperature distribution map of the carbide slag accumulation. S2. Based on the load distribution map and temperature distribution map, a convolutional neural network model is used to process the map data to determine the structural instability risk level in the carbide slag accumulation. S3. If the structural instability risk level is higher than the preset threshold, the temperature input of the corresponding area is reduced by adjusting the power parameters of the heating method to obtain an optimized heating distribution scheme. S4. After obtaining the optimized heating distribution scheme, use the support vector machine model to analyze the impact of the scheme on the overall drying efficiency and determine whether it meets the efficiency requirements under the production safety standard. S5. If the efficiency requirements under the production safety standards are met, the updated temperature field and load-bearing distribution field of the carbide slag accumulation can be obtained by adjusting the heating method in real time. S6. Based on the updated temperature field and load-bearing distribution field of carbide slag accumulation, a recurrent neural network model is used to predict the structural stability change trend in the future and determine potential local thermal evolution paths. S7. By generating a targeted heating control command sequence through potential local thermal evolution paths, a dynamic monitoring cycle mechanism to maintain structural stability is obtained.

2. The method for temperature control in drying carbide slag based on multi-agent collaboration according to claim 1, characterized in that: S1 includes: A sensor network is deployed in a grid-like manner in the area where carbide slag accumulates to acquire pressure and temperature data and generate an initial dataset. Based on the initial dataset, data cleaning methods are used to remove noisy data, resulting in cleaned pressure and temperature data. By using the pressure data after cleaning, an interpolation algorithm is used to calculate the load-bearing distribution and generate a load-bearing distribution map of the carbide slag accumulation. Using the temperature data after cleaning, an interpolation algorithm is used to calculate the temperature distribution and generate a temperature distribution map of the carbide slag accumulation. If there are areas in the load distribution map where the pressure anomaly is higher than the preset threshold, the area data will be re-collected through the sensor network to obtain updated pressure data. If there are areas in the temperature distribution map where the temperature anomaly is higher than the preset threshold, the area data will be re-collected through the sensor network to obtain updated temperature data. Based on the updated pressure and temperature data, a weighted average algorithm is used to fuse multiple sets of data to generate optimized load distribution maps and temperature distribution maps.

3. The method for temperature control in drying carbide slag based on multi-agent collaboration according to claim 1, characterized in that: S2 includes: Initial image data is generated by acquiring load distribution and temperature distribution maps of the carbide slag accumulation area through a sensor network. The initial image data was standardized using data preprocessing methods to obtain standardized image data. Feature vectors are obtained by extracting features from standardized image data using a convolutional neural network. Based on the feature vectors, a convolutional neural network is used to calculate the probability of the occurrence of local overheating regions, and the probability distribution is obtained. If there is a region in the probability distribution that is higher than the preset threshold, the load-bearing and temperature data of that region will be re-acquired through the sensor network to obtain updated image data. By analyzing the updated image data using a convolutional neural network, the risk level of structural instability is calculated, and the risk assessment result is obtained. Based on the risk assessment results, the carbide slag accumulation areas were classified and labeled using a regional division method, resulting in a graded distribution map.

4. The method for temperature control of carbide slag drying based on multi-agent collaboration according to claim 1, characterized in that: S3 includes: Temperature and load-bearing data of the carbide slag accumulation area are acquired through a sensor network to generate an initial distribution dataset. The initial distributed dataset is processed using data normalization methods to obtain a standardized dataset. Feature vectors are obtained by extracting features from a standardized dataset using a convolutional neural network. If there are risk indicators in the feature vector that are higher than the preset threshold, then the temperature data and load data of the corresponding area are re-collected through the sensor network to generate an updated distribution dataset. Logistic regression model is used to analyze the updated distribution dataset to determine the level of structural instability risk; Based on the structural instability risk level, the heating power parameters are adjusted to generate an optimized heating distribution scheme. The optimized heating distribution scheme is validated using a sensor network, and a validation dataset is obtained to determine the stability of temperature and load-bearing data.

5. The method for temperature control of carbide slag drying based on multi-agent collaboration according to claim 1, characterized in that: S4 includes: The initial efficiency dataset is obtained by acquiring the temperature control parameters and load distribution data corresponding to the optimized heating distribution scheme through a sensor network. The initial efficiency dataset is processed using a data standardization method to obtain a standardized efficiency dataset. Drying efficiency data is obtained by classifying and analyzing the standardized efficiency dataset using a support vector machine model. If the drying efficiency data is lower than the efficiency requirement preset by the production safety standard, the temperature control parameters and load distribution data will be re-collected through the sensor network to obtain an updated efficiency dataset. Logistic regression model was used to analyze the update efficiency dataset and determine risk assessment indicators; The temperature control parameters are adjusted based on risk assessment indicators to obtain an optimized adjustment strategy. By verifying and optimizing the adjustment strategy through sensor networks, a verification dataset is obtained to determine whether the drying efficiency data meets production safety standards.

6. The method for temperature control of carbide slag drying based on multi-agent collaboration according to claim 1, characterized in that: S5 includes: The initial field distribution dataset is obtained by acquiring updated temperature field data and load-bearing distribution field data of carbide slag accumulation through a sensor network. The initial field distribution dataset is processed using a data standardization method to obtain a standardized field distribution dataset; The standardized field distribution dataset was classified and analyzed using a random forest model to obtain the characteristic classification results of temperature field and load-bearing distribution. If the feature classification results show that the temperature field or load distribution deviates from the preset stable threshold, the temperature field data and load distribution field data are re-collected through the sensor network to obtain an updated field distribution dataset. Principal component analysis was used to reduce the dimensionality of the updated field distribution dataset to obtain a dimensionality-reduced feature dataset. Based on the dimensionality-reduced feature dataset, the stability of the temperature field and load distribution is analyzed using a logistic regression model to obtain stability assessment results; If the stability assessment results meet the production safety standards, the heating distribution parameters are adjusted in real time through the sensor network to obtain optimized temperature field data and load distribution field data.

7. The method for temperature control of carbide slag drying based on multi-agent collaboration according to claim 1, characterized in that: S6 includes: The temperature field data and load-bearing distribution field data of the carbide slag accumulation are acquired by a sensor network to obtain a real-time distribution dataset. A data standardization method is used to process the real-time distributed dataset to obtain a standardized distributed dataset; By analyzing a standardized distribution dataset using a recurrent neural network model, the stable trend of structural changes is predicted, and the stable trend prediction results are obtained. If the stable trend prediction results show that the local thermal risk is higher than the preset threshold, the temperature field data will be re-acquired through the sensor network to obtain an updated temperature dataset. Cluster analysis was used to group the updated temperature dataset to determine the distribution of areas with local overheating risk higher than a preset threshold, thus obtaining a thermal region dataset. Based on the thermal region dataset, local overheating paths are determined by a preset threshold to obtain the thermal evolution path distribution; By adjusting the heating distribution parameters to optimize the thermal evolution path distribution, optimized temperature field data and load-bearing distribution field data are obtained.

8. The method for temperature control of carbide slag drying based on multi-agent collaboration according to claim 1, characterized in that: S7 includes: By acquiring temperature distribution data of local overheated areas through a sensor network, a real-time temperature dataset is obtained. Based on the real-time temperature dataset, the temperature distribution is grouped using the K-means clustering method to determine the distribution of areas where the local temperature is higher than the preset threshold, thus obtaining the overheated area dataset. By identifying temperature anomalies in the overheated area dataset using a preset threshold, an initial heating control command sequence is generated, resulting in a control command dataset. A long short-term memory network model is used to analyze the control command dataset, predict the stable change trend of the structure, and obtain the stable trend prediction results.

9. The method for temperature control of carbide slag drying based on multi-agent collaboration according to claim 8, characterized in that: The S7 also includes: If the stable trend prediction results show that the local thermal risk is higher than the preset threshold, then the temperature distribution data and load distribution data are re-collected through the sensor network to obtain an updated distribution dataset; Based on the updated distribution dataset, a support vector machine model is used to classify the evolution paths of the overheated region, resulting in the optimized distribution of overheated evolution paths. By optimizing the overheating evolution path distribution, a dynamically adjusted heating control command sequence is generated, and a real-time monitoring cycle mechanism is determined.

10. A temperature control system for drying carbide slag based on multi-agent collaboration, used to implement the steps of the temperature control method for drying carbide slag based on multi-agent collaboration as described in any one of claims 1-9, characterized in that, The system includes: The sensor network module collects real-time pressure and temperature data through a sensor network deployed in the carbide slag accumulation area, and obtains the current load-bearing distribution map and temperature distribution map of the carbide slag accumulation. The image processing module uses a convolutional neural network model to process the image data based on the load distribution map and temperature distribution map to determine the structural instability risk level in the carbide slag accumulation. If the structural instability risk level is higher than a preset threshold, the heating control module reduces the temperature input of the corresponding area by adjusting the power parameters of the heating method, thereby obtaining an optimized heating distribution scheme. The efficiency assessment module, after obtaining the optimized heating distribution scheme, uses a support vector machine model to analyze the impact of the scheme on the overall drying efficiency and determine whether it meets the efficiency requirements under the production safety standards. If the efficiency requirements under the production safety standards are met, the execution control module will adjust the heating method in real time to obtain the updated temperature field and load-bearing distribution field of the carbide slag accumulation. The stability prediction module uses a recurrent neural network model to predict the structural stability change trend in the future time based on the updated temperature field and load distribution field of carbide slag accumulation, and determines the potential local thermal evolution path. The instruction generation module generates a targeted heating control instruction sequence through potential local thermal evolution paths, thereby obtaining a dynamic monitoring cycle mechanism to maintain structural stability.

Images