Low-delay control method and system for gangue selecting robot based on edge computing

Abstract

The application discloses a gangue selecting robot low-delay control method and system based on edge calculation, and relates to the technical field of edge calculation. The method comprises the following steps: acquiring belt operation parameters and coal-gangue multi-modal data; determining an optimal compression ratio based on the belt operation parameters, compressing the multi-modal data, and generating a compressed data packet; sending the compressed data packet to an edge node of the gangue selecting robot, dynamically allocating edge calculation power, and obtaining an allocation result containing data analysis calculation power and parameter optimization calculation power; analyzing the compressed data packet based on the data analysis calculation power, and acquiring a coal-gangue identification result; iteratively optimizing robot control parameters based on the coal-gangue identification result by using the parameter optimization calculation power, generating an optimized control scheme, and realizing low-delay accurate control of the gangue selecting robot. Through dynamic data compression, elastic calculation power allocation and parallel processing, the application reduces data processing and decision-making delay, and improves the real-time performance, accuracy and working condition adaptability of the separation system.

Description

Technical Field

[0001] This invention relates to the field of edge computing technology, and specifically to a low-latency control method and system for a coal sorting robot based on edge computing. Background Technology

[0002] Coal gangue sorting is a crucial step in ensuring coal quality. Machine vision-based gangue sorting robots are gradually replacing manual sorting. They analyze material images on the conveyor belt in real time and drive robotic arms to perform sorting. This process has extremely stringent requirements for the real-time performance of data processing and decision control. Even a slight delay in recognition or control can lead to missed gangue or misoperation, directly affecting sorting efficiency and coal quality.

[0003] Existing technologies typically employ centralized cloud computing or fixed-mode edge computing architectures to handle coal and gangue identification and control tasks. Centralized solutions upload the collected multimodal data entirely to cloud servers for analysis. While this enables complex calculations, significant network transmission latency makes it difficult to meet the real-time requirements of high-speed conveyor belt scenarios. Although some edge computing solutions offload computational tasks to field devices, they often use static resource allocation strategies and fixed data compression methods. They fail to dynamically adjust computational resources and processing flows based on conveyor belt operating status, data characteristics, and identification difficulty. This leads to uneven distribution of computing power, processing bottlenecks, or response delays at edge nodes when there are sudden changes in conveyor belt speed, complex coal and gangue textures, or a surge in data volume. Consequently, these issues affect identification accuracy and control timeliness, limiting the reliability and adaptability of coal and gangue sorting robots in real-world industrial environments. Summary of the Invention

[0004] This invention addresses the technical problems of insufficient real-time performance and low sorting efficiency of existing coal sorting robots caused by data processing delays, rigid edge computing power allocation, and lag in control parameter response. It provides a low-latency control method and system for coal sorting robots based on edge computing.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: In a first aspect, the present invention provides a low-latency control method for a coal sorting robot based on edge computing, comprising: Acquire conveyor belt operating parameters and coal and gangue multimodal data; Based on the belt conveyor operating parameters, the optimal compression ratio is obtained, and the coal gangue multimodal data is compressed to obtain a compressed data package; The compressed data packet is sent to the edge node of the coal sorting robot, and the computing power allocation ratio is obtained based on the size of the compressed data packet and the multimodal data characteristics of coal gangue. The computing power of the edge node is allocated, and the computing power allocation result is obtained. The computing power allocation result includes data analysis computing power and parameter optimization computing power. Using the aforementioned data analysis computing power, the compressed data packet is analyzed to obtain coal gangue identification results; The computing power is optimized using the parameters, and the control parameters of the gangue sorting robot are iteratively optimized based on the coal gangue identification results to obtain an optimized control scheme and perform low-latency control of the gangue sorting robot.

[0006] Secondly, the present invention provides a low-latency control system for a coal sorting robot based on edge computing, comprising: The data acquisition module is used to acquire conveyor belt operating parameters and multimodal data of coal and gangue. The data compression module is used to obtain the optimal compression ratio based on the belt operation parameters, compress the coal gangue multimodal data, and obtain a compressed data package; The computing power allocation module is used to send the compressed data packet to the edge node of the coal gangue sorting robot, and obtain the computing power allocation ratio based on the size of the compressed data packet and the multimodal data characteristics of coal gangue, allocate the computing power of the edge node, and obtain the computing power allocation result, wherein the computing power allocation result includes data analysis computing power and parameter optimization computing power; The data analysis module is used to perform data analysis on the compressed data packet using the data analysis computing power to obtain coal gangue identification results; The optimization and control module is used to optimize computing power using the parameters, iteratively optimize the control parameters of the coal gangue sorting robot based on the coal gangue identification results, obtain an optimized control scheme, and perform low-latency control of the coal gangue sorting robot.

[0007] The beneficial effects of this invention are: Compared to existing technologies, this invention first dynamically determines the data compression ratio by integrating the belt's operating status, effectively reducing the amount of transmitted data and lowering network transmission latency while ensuring recognition accuracy. Secondly, it flexibly allocates edge computing power based on real-time data characteristics and task requirements, avoiding waste or bottlenecks in computing resources and improving resource utilization efficiency and system responsiveness. Thirdly, it decouples and parallelizes the recognition task and control parameter optimization task at the computing power level, shortening the overall link time from perception to decision-making. Finally, it iteratively optimizes control parameters online based on the recognition results, enabling the robot to adapt to changes in working conditions and achieving stable and low-latency precise control, comprehensively improving the efficiency and reliability of coal sorting operations. Attached Figure Description

[0008] Figure 1 A flowchart illustrating the low-latency control method for a coal sorting robot based on edge computing provided by this invention; Figure 2 This is a schematic diagram of the low-latency control system for a coal sorting robot based on edge computing provided by the present invention.

[0009] In the attached diagram, the components represented by each number are as follows: Data acquisition module 11, data compression module 12, computing power allocation module 13, data analysis module 14, optimization and control module 15. Detailed Implementation

[0010] Example 1, as Figure 1 As shown, this embodiment of the invention provides a low-latency control method for a coal sorting robot based on edge computing, including: S10: Obtain belt operation parameters and coal gangue multimodal data; Specifically, acquiring conveyor belt operating parameters and multimodal data of coal and gangue includes: Obtain belt operating parameters, wherein the belt operating parameters include belt operating speed and belt operating fluctuation coefficient; Obtain multimodal data of coal gangue, wherein the multimodal data of coal gangue includes coal gangue texture information and coal gangue grayscale information.

[0011] First, acquire the belt operating parameters. Belt operating parameters are physical quantities characterizing the real-time working state of the conveyor belt, including belt speed and belt ripple coefficient. Belt speed refers to the linear displacement of the material conveyed by the belt per unit time, directly related to the frequency of image acquisition and data processing time. The belt ripple coefficient is a dimensionless index used to quantify the smoothness of belt operation, reflecting the degree to which the belt's speed or position deviates from its theoretical steady-state value due to factors such as mechanical vibration, load changes, or drive instability.

[0012] Secondly, multimodal data of coal and gangue is acquired. Multimodal data refers to multi-dimensional sensory information collected by one or more sensors deployed along the conveyor belt, characterizing the different physical properties of coal and gangue. This multimodal data includes coal and gangue texture information and coal and gangue grayscale information. Coal and gangue texture information originates from visual or tactile perception of the material's surface structure, describing the differences in spatial distribution characteristics of coal and gangue in terms of surface roughness, pattern, and cracks. Coal and gangue grayscale information reflects the intensity of reflection or absorption of light of a specific wavelength on the material's surface, usually presented as a sequence of brightness values ​​of image pixels, and is the basis for distinguishing the color and spectral characteristics of coal and gangue. Coal and gangue texture information and coal and gangue grayscale information together constitute the original data foundation for identifying and classifying the target material.

[0013] S20: Based on the belt operation parameters, obtain the optimal compression ratio, compress the coal gangue multimodal data, and obtain a compressed data packet; Specifically, based on the belt conveyor operating parameters, the optimal compression ratio is obtained, and the coal and gangue multimodal data is compressed to obtain a compressed data package, including: Construct a compression ratio optimization model; Input the belt running parameters into the compression ratio optimization model to obtain the speed compression ratio; Based on the coal and gangue multimodal data, the proportion of coal and gangue texture information is obtained, and based on the proportion of coal and gangue texture information, the mass compression ratio is obtained; The velocity compression ratio and the mass compression ratio are combined and calculated to compress the coal gangue multimodal data and obtain a compressed data package.

[0014] First, a compression ratio optimization model is constructed. This model is a mathematical model trained using historical operating data and performance targets. Its function is to establish a nonlinear mapping relationship between belt operating parameters and recommended data compression ratios. It can be used to dynamically adjust the data compression strategy based on the real-time belt operating status, minimizing the loss of effective features caused by data compression while meeting low-latency control requirements.

[0015] Specifically, a compression ratio optimization model is constructed, including: Obtain the sample belt operation parameter set, and obtain the data packet compression ratio of the control delay of the coal selection robot under the sample belt operation parameters that is less than or equal to the control delay threshold, as the sample compression ratio set; A compression ratio optimization model is constructed, using the sample belt running parameter set as input and the sample compression ratio set as supervision, and the compression ratio optimization model is trained until convergence.

[0016] First, a sample set of belt operation parameters was obtained. This set was compiled through systematic collection and organization of historical operation records. It contains multiple sets of belt operation parameters actually recorded under different working conditions. Each set of belt operation parameters consists of two specific dimensions: belt speed and belt fluctuation coefficient. The sample set of belt operation parameters aims to comprehensively cover all possible working states of the belt, providing a sufficient range of feature variations for model learning.

[0017] Secondly, a sample compression ratio set is obtained. Establishing this set requires associating it with each specific parameter group in the sample conveyor belt operation parameter set. Specifically, for each set of sample conveyor belt operation parameters, historical operation data is backtracked and filtered to identify the real-time data packet compression ratio values ​​that have been successfully adopted and verified as effective when the total delay time generated by the sorting robot from perception to execution of the complete control process under the corresponding actual working conditions is less than or equal to the control delay threshold. The control delay threshold is a pre-set performance boundary based on the real-time requirements of the sorting process, used to define the maximum acceptable control delay, for example, set to 300 milliseconds. Using this method, a feasible compression ratio that has been practically verified and meets the low-latency requirement is matched to each set of sample conveyor belt operation parameters, thus forming the sample compression ratio set.

[0018] The final set of sample belt operation parameters and the set of sample compression ratios together constitute a set of training data pairs with a clear correspondence. Each pair of data reflects the effective data compression ratio that can be used to achieve low-latency control under a specific belt operation state.

[0019] Further, a compression ratio optimization model is constructed and trained. Preferably, the compression ratio optimization model can be a machine learning model structure with nonlinear fitting capabilities, such as a neural network or support vector regression model. During training, the set of sample belt running parameters is used as the input features of the model, and the corresponding set of sample compression ratios is used as the target supervised values. The internal parameters of the model are continuously adjusted through an iterative optimization algorithm, so that the error between the model output value, i.e., the predicted compression ratio, and the supervised sample compression ratio gradually decreases. Training continues until convergence, obtaining the trained compression ratio optimization model. The convergence condition can be set according to the needs of model training, for example, it can be set as the mean squared error on the validation set decreasing by less than 0.001 within 10 consecutive training epochs, or the maximum number of training epochs can be set to 1000 epochs. When the model meets any convergence condition, the training process terminates, and the model obtained at this time is the trained compression ratio optimization model. This compression ratio optimization model has the ability to accurately and quickly infer the recommended compression ratio that meets the low-latency control requirements based on the real-time input belt running parameters.

[0020] For example, as a feasible implementation, the compression ratio optimization model can be constructed using a neural network structure with nonlinear mapping capabilities. Since the relationship between belt operating parameters and the optimal compression ratio typically exhibits complex nonlinear characteristics, neural networks can effectively learn and fit this complex relationship through their multi-layered nonlinear transformations.

[0021] Specifically, the constructed compression ratio optimization model can include an input layer, several hidden layers, and an output layer. The input layer receives a standardized vector of belt operation parameters, which contains at least two dimensions: belt speed and belt fluctuation coefficient. The number of hidden layers and the number of neurons in each layer can be configured according to the data scale and task complexity; for example, two hidden layers with 16 and 8 neurons per layer, respectively. The activation function in the hidden layers can be the ReLU function to introduce non-linear computational power and enhance the model's ability to express complex relationships. To improve the model's generalization ability and prevent overfitting on the training data, Dropout regularization layers can be introduced between or after the hidden layers, with the dropout rate set in the range of 0.1 to 0.3. The output layer uses a linear activation function, ultimately outputting a continuous numerical value, which is the optimal compression ratio predicted by the model.

[0022] Model training requires configuring appropriate hyperparameters. For example, the learning rate can be initialized to 0.005 to strike a balance between training stability and convergence speed; the total number of training epochs can be set to 200 epochs to ensure that the model has enough opportunities to learn patterns in the data; and the size of each training batch can be set to 32 to balance computational efficiency and gradient update stability.

[0023] The training process employs a supervised learning paradigm. In practice, the set of sample belt running parameters obtained in the preceding steps is used as the input feature set, and the corresponding set of sample compression ratios is used as the target value set. These are then divided into training subsets, validation subsets, and test subsets according to a preset ratio, such as 7:2:1.

[0024] During the training phase, the belt conveyor operating parameter vectors from the training subset are input into the model, and the model outputs the predicted compression ratio. The difference between the predicted compression ratio and the true target value in the sample compression ratio set is calculated using the mean squared error loss function. The Adam optimization algorithm is employed, and all weight parameters in the model are iteratively updated based on this loss value through backpropagation. The validation subset is used to monitor the training process and evaluate the model's performance on unseen data. When the mean squared error on the validation set decreases by less than 0.001 over 10 consecutive training epochs, the model is considered to have converged, and the training process is terminated. Finally, a fully trained compression ratio optimization model is obtained, which has the ability to efficiently and accurately infer the corresponding optimal compression ratio based on real-time belt conveyor operating parameters.

[0025] Furthermore, the real-time collected belt running parameters are input into the trained compression ratio optimization model to obtain the speed compression ratio. Specifically, the belt running speed directly affects the time window for material to pass through the identification area. The faster the speed, the shorter the time allowed for data processing and decision-making. Therefore, the model tends to output a higher compression ratio recommendation value, aiming to reduce the overall delay risk by reducing the total amount of data to be transmitted. At the same time, the belt running fluctuation coefficient reflects the stability of the conveying process. The larger the coefficient, the more significant the instantaneous speed deviation from the expectation or the mechanical vibration. Unstable operating conditions require subsequent control algorithms to have stronger robustness and adaptability. To provide a more sufficient data foundation for analysis and decision-making, the model may recommend a relatively low compression ratio in this case, so as to retain more original information reflecting the details of the operating conditions in the compressed data packet.

[0026] The final output compression ratio is a dimensionless value between 0 and 1, representing the maximum compression allowed from a real-time perspective while meeting a predetermined control delay threshold. This compression ratio primarily sets a theoretical upper limit for subsequent data compression operations from the perspective of ensuring the timeliness of control actions.

[0027] Furthermore, based on the coal and gangue multimodal data, the proportion of coal and gangue texture information is obtained, and the quality compression ratio is obtained based on this proportion. First, feature analysis is performed on the real-time acquired coal and gangue multimodal data to calculate the proportion of coal and gangue texture information in the total amount of information in the multimodal data, i.e., the proportion of coal and gangue texture information. This proportion is a value between 0 and 1, used to objectively measure the share of texture information in the overall dataset's information capacity.

[0028] Because coal and gangue texture information carries crucial surface morphological details that distinguish coal from gangue, although the extraction and recognition of such features involves high computational complexity, it plays a decisive role in improving the final classification accuracy. Therefore, texture information must be carefully protected during data compression to minimize its distortion.

[0029] Specifically, the quality compression ratio is calculated based on the proportion of coal gangue texture information. The principle is that a higher proportion of coal gangue texture information indicates richer high-value details crucial to the recognition task in the current dataset; to reduce the risk of information loss during compression, a more cautious compression strategy should be adopted, i.e., a smaller compression ratio should be assigned. Specifically, this can be quantified through a functional relationship, for example, defining the quality compression ratio as 1 - α × proportion of coal gangue texture information, where α is an adjustment coefficient determined experimentally, ranging from 0.1 to 0.5. A larger proportion of coal gangue texture information results in a smaller quality compression ratio, and a milder compression operation. This quality compression ratio primarily sets a lower limit for data compression to ensure the accuracy of subsequent coal gangue recognition algorithms, preventing the loss of key features due to over-compression.

[0030] Finally, the velocity compression ratio and mass compression ratio are fused together to compress the coal gangue multimodal data and obtain a compressed data package. Specifically, after obtaining the velocity compression ratio and mass compression ratio, they need to be fused to determine the optimal compression ratio for the final compression operation. The fusion strategy needs to comprehensively consider real-time control requirements and recognition accuracy requirements, such as performing a weighted average between the two or taking the median of their intersection range.

[0031] Preferably, the optimal compression ratio = β × speed compression ratio + (1-β) × mass compression ratio, where β is a weighting coefficient, and its value is set according to the priority between real-time control and recognition accuracy in actual applications. For example, in the case where sorting speed is emphasized, β can be set to 0.7; while in the scenario where coal and gangue are difficult to distinguish and recognition accuracy is extremely important, β can be set to 0.3.

[0032] After determining the optimal compression ratio, the original coal gangue multimodal data, including coal gangue texture information and coal gangue grayscale information, is processed using a specified lossy or lossless compression algorithm based on this ratio. For example, a JPEG compression algorithm based on discrete cosine transform, a compression method based on wavelet transform, or a compression coding technique specifically designed for image feature preservation can be used. The selection of the compression algorithm must balance compression efficiency and feature preservation capability. While achieving a significant reduction in data volume, priority should be given to ensuring the integrity of key classification features such as coal gangue texture information through algorithm parameter configuration or preprocessing methods to avoid introducing excessive distortion due to compression, which would affect the subsequent recognition accuracy.

[0033] After the compression process is complete, the generated data unit is the compressed data packet. This compressed data packet integrates the compressed multimodal information of coal and gangue, and its data volume is significantly reduced compared to the original data, thereby reducing network transmission load and latency. This compressed data packet will be sent to the edge computing node deployed near the gangue sorting robot for subsequent real-time analysis and decision control.

[0034] S30: Send the compressed data packet to the edge node of the coal gangue sorting robot, and obtain the computing power allocation ratio based on the size of the compressed data packet and the multimodal data characteristics of coal gangue, allocate the computing power of the edge node, and obtain the computing power allocation result, wherein the computing power allocation result includes data analysis computing power and parameter optimization computing power; Specifically, the compressed data packet is sent to the edge node of the coal sorting robot, and the computing power allocation ratio is obtained based on the size of the compressed data packet and the multimodal data characteristics of the coal gangue. The computing power of the edge node is then allocated, and the computing power allocation result is obtained. The computing power allocation result includes data analysis computing power and parameter optimization computing power, including: Based on the multimodal data characteristics of the coal gangue, the characteristic computing power allocation ratio is obtained; Based on the ratio of the compressed data packet size to the standard compressed data packet size, the data volume computing power allocation ratio is obtained; The computing power allocation ratio is obtained by weighting the feature computing power allocation ratio and the data volume computing power allocation ratio. The computing power is allocated to the edge nodes using the aforementioned computing power allocation ratio, and the computing power allocation results are obtained. The computing power allocation results include data analysis computing power and parameter optimization computing power.

[0035] Specifically, the aforementioned compressed data packet is sent to the edge node of the coal sorting robot. The edge node refers to an embedded or dedicated computing device deployed on or near the coal sorting robot, possessing certain computing, storage, and communication capabilities. Because the edge node has limited hardware computing resources and needs to simultaneously handle two computationally intensive tasks—real-time data analysis and control parameter optimization—its computing power must be dynamically and rationally allocated to balance recognition accuracy requirements with decision-making response speed, thereby ensuring the entire coal sorting control process operates stably and efficiently under strict time constraints.

[0036] First, based on the multimodal data characteristics of coal gangue, the characteristic computing power allocation ratio is obtained.

[0037] Specifically, based on the multimodal data characteristics of the coal gangue, the feature computing power allocation ratio is obtained, including: Based on the coal and gangue identification results of the previous moment, coal and gangue multimodal data features are obtained, wherein the coal and gangue multimodal data features include the proportion of historical coal identification. Based on the historical coal identification ratio and the coal gangue texture information ratio, the identification difficulty coefficient is obtained; Based on the recognition difficulty coefficient, the feature computing power allocation ratio is obtained.

[0038] First, based on the coal and gangue identification results from the previous time step, the multimodal data features of the current batch of materials are extracted and quantified. Since the materials on the conveyor belt typically exhibit continuity and batch similarity, the ratio of coal to gangue identified in the previous time step—that is, the historical coal identification ratio—can reflect, to some extent, the basic compositional characteristics of the material to be processed. This historical coal identification ratio serves as a priori feature for inferring the potential complexity of the current identification task.

[0039] Secondly, the identification difficulty coefficient is calculated by combining the historical coal identification ratio with the current coal and gangue texture information ratio obtained from real-time analysis. The identification difficulty coefficient is a comprehensive indicator used to quantify the complexity of the current coal and gangue sorting task. Specifically, if the historical coal identification ratio indicates a high degree of mixing of coal and gangue in the material, the initial classification will be more difficult. Simultaneously, the coal and gangue texture information ratio directly reflects the amount of detailed, computationally intensive texture features in the data; a higher ratio means more refined feature extraction and analysis are required. Therefore, the identification difficulty coefficient is positively correlated with both the historical coal identification ratio and the coal and gangue texture information ratio, and can be calculated through weighted summation or multiplication.

[0040] Preferably, the recognition difficulty coefficient = γ × historical coal identification ratio + δ × coal gangue texture information ratio, where γ and δ are the corresponding weighting coefficients, and their values ​​can be calibrated according to the actual influence of the two features on the recognition result in historical data, for example, they can be set to 0.6 and 0.4 respectively. The higher the recognition difficulty coefficient, the more computational resources are required to complete accurate recognition.

[0041] Finally, based on the calculated recognition difficulty coefficient, the feature computing power allocation ratio is dynamically determined. The feature computing power allocation ratio quantitatively defines the share of computing resources that should be allocated to real-time data analysis tasks within the total available computing resources of the current edge node. Specifically, the higher the recognition difficulty coefficient, the more complex the visual features representing the coal gangue material to be processed, and the higher the recognition uncertainty. Correspondingly, more computing resources need to be invested in the data analysis phase to ensure high-precision feature extraction and classification within a limited time. Therefore, a positively correlated functional relationship is established between the feature computing power allocation ratio and the recognition difficulty coefficient.

[0042] Specifically, quantification can be achieved through a preset mapping function, such as a linear function: Feature computing power allocation ratio = Proportion coefficient × Recognition difficulty coefficient + Basic computing power allocation constant. Here, the proportion coefficient adjusts the sensitivity of computing resources to changes in recognition difficulty, while the basic computing power allocation constant represents the minimum computing power share required to maintain the basic operation of the data analysis task, ensuring that even with an extremely low recognition difficulty coefficient, the data analysis process can still obtain the necessary basic computing resources to guarantee its continuous operation.

[0043] The specific values ​​of the proportionality coefficient and the basic computing power allocation constant can be calibrated through experimental simulation and optimization based on the actual computing power configuration of the edge nodes, the overall system performance requirements, and historical task load data. For example, the proportionality coefficient can be set to 0.6 and the basic computing power allocation constant to 0.1. This quantification mechanism can adaptively and accurately allocate the computing power quota for coal gangue identification tasks based on the inherent complexity of the analysis task directly reflected by the material characteristics.

[0044] Furthermore, the data volume computing power allocation ratio is obtained based on the ratio of the compressed data packet size to the standard compressed data packet size. This ratio reflects the computational resource requirements of the current data packet size. It is calculated by dividing the real-time compressed data packet size by a predefined standard compressed data packet size. The standard compressed data packet size is a baseline value set based on historical average data volume or system design capacity, for example, 50KB. A larger ratio indicates a higher data volume computing power allocation ratio, meaning the current compressed data packet size exceeds the normal level significantly, requiring more computational resources and time to process, and therefore a higher proportion of computing power should be allocated to it.

[0045] Furthermore, the aforementioned calculated feature computing power allocation ratio and data volume computing power allocation ratio are weighted to obtain the final computing power allocation ratio used to guide resource allocation. Specifically, the computing power allocation ratio = ω1 × feature computing power allocation ratio + ω2 × data volume computing power allocation ratio, where ω1 and ω2 are the weight coefficients assigned to the two allocation ratios, and satisfy ω1 + ω2 = 1. The values ​​of the weight coefficients ω1 and ω2 can be adjusted according to the preference for recognition accuracy and processing throughput in actual applications. For example, in scenarios emphasizing accuracy, ω1 can be set to 0.7 and ω2 to 0.3.

[0046] Finally, the calculated computing power allocation ratio is used to divide the total available computing resources of the edge nodes, obtaining specific computing power allocation results. Specifically, multiplying the calculated computing power allocation ratio by the real-time available total computing power of the edge nodes yields the computing power allocated to the data analysis task, i.e., the data analysis computing power. The remaining computing power of the edge nodes is automatically allocated to the parameter optimization task, i.e., the parameter optimization computing power. The resulting computing power allocation results clearly define the amount of computing resources used for coal gangue identification and analysis, and the amount of computing resources used for iterative optimization of control parameters within the current computing cycle, thereby achieving a dynamic, reasonable, and quantitative allocation of limited edge computing power between the two tasks.

[0047] S40: Using the data analysis computing power, perform data analysis on the compressed data packet to obtain coal gangue identification results; Specifically, the data analysis computing power is used to perform data analysis on the compressed data packet to obtain coal gangue identification results, including: Obtain a coal gangue identification plugin, wherein the coal gangue identification plugin includes K identification branches; Based on the ratio of the data analysis computing power to the preset data analysis computing power, the number of adaptive identification branches Q is obtained. Q identification branches are called in the coal gangue identification plugin to perform data analysis on the compressed data package and obtain the coal gangue identification result, wherein Q is less than or equal to K.

[0048] First, a pre-installed coal and gangue identification plugin needs to be acquired. This plugin is an integrated identification software module that encapsulates K structurally independent and fully functional identification branches. Each branch has the ability to extract features from the input data and classify coal and gangue. This plugin can flexibly schedule different numbers of identification branches to participate in parallel inference based on real-time allocated computing resources, thereby maximizing identification accuracy while meeting low latency constraints and achieving a dynamic optimal match between computing resource utilization and task execution efficiency.

[0049] Specifically, the coal gangue identification plugin includes: Based on a convolutional neural network, construct K recognition branches; Obtain a set of sample compressed data packets, wherein the sample compressed data packets include sample coal gangue texture information and sample coal gangue grayscale information; Based on the sample compressed data package set, a sample identification result set is obtained, wherein the sample identification result includes coal identification and gangue identification; The sample compressed data package set and the sample recognition result set are randomly sampled with replacement to obtain K training datasets of the same size. The K recognition branches are trained separately until convergence. The K identification branches are integrated using an average integration strategy to obtain the coal gangue identification plugin.

[0050] First, based on the convolutional neural network model architecture, K recognition branches with identical or similar structures but independent parameters are constructed in parallel. Each recognition branch is a complete convolutional neural network classification model, with its input layer adapted to the format of compressed data packets and its output layer corresponding to the labels of the two categories: coal and gangue. The structural parameters of the recognition branches, such as depth and kernel size, are uniformly set according to the requirements of the recognition task. For example, the network can be set to contain 5 convolutional layers, each using a 3×3 kernel and employing the ReLU activation function, and finally connected to a fully connected layer with 2 neurons as the output, thereby ensuring that each branch has the same basic feature learning capability.

[0051] Secondly, a set of compressed sample data packages is obtained. This set of compressed sample data packages originates from a collection of accurately labeled compressed data packages accumulated during historical production processes. Each compressed sample data package contains compressed texture information and grayscale information of the sample coal and gangue. Simultaneously, each compressed sample data package is paired with its actual material category identifier, forming a sample identification result set. Each sample identification result in this set includes a coal identifier and a gangue identifier, clearly indicating whether the material in the corresponding compressed sample data package is coal or gangue.

[0052] Furthermore, a random sampling method with replacement is employed to sample from the compressed sample data package set and the corresponding sample recognition result set. This process is repeated K times, each time independently sampling a new dataset of the same size as the original set but with potentially different compositions, ultimately generating K training datasets of consistent size but with subtle differences in their internal sample distributions. Specifically, this sampling method aims to provide slightly different data perspectives for each recognition branch, thereby helping to improve the model's generalization ability and robustness during subsequent ensemble analysis.

[0053] Then, the K generated training datasets are used to independently train the K recognition branches. Each recognition branch, on its dedicated dataset, optimizes its network parameters using backpropagation through a standard supervised learning process. During training, the model's performance on a reserved validation set is continuously monitored. When the classification accuracy on the validation set no longer improves and fluctuates within a range of less than 0.5% over several consecutive training epochs (e.g., 20 epochs), and the cross-entropy loss function value drops below 0.01 and remains stable, the model for that branch is considered converged, and its training process is terminated. After training, K recognition branches, each with independent classification capabilities, are obtained.

[0054] Finally, the trained K recognition branches are integrated using an average ensemble strategy. Specifically, the structure, parameters, and calling interfaces of the K branches are encapsulated within a unified software framework, forming a coal gangue recognition plugin. During plugin runtime, for a given input data, the internally integrated average ensemble strategy automatically calls a specified number of branches for inference and performs an arithmetic average of the classification probabilities or confidence scores output by all called branches, using the final integrated result as the plugin's output.

[0055] The resulting coal gangue identification plugin integrates the identification capabilities of multiple differentiated models, enabling improved stability and accuracy through dynamic invocation.

[0056] Furthermore, based on the currently available data analysis computing power, the number Q of recognition branches to be activated in this recognition task is dynamically determined. Specifically, the ratio of the current data analysis computing power to the preset data analysis computing power is calculated. The preset data analysis computing power represents the minimum computing resources required to fully invoke a recognition branch and complete a single inference within a specified time, and is set based on the floating-point operation complexity of a single forward inference of the selected convolutional neural network model and the clock speed of the single-core processor of the edge node. Based on the calculated ratio, the number of compatible recognition branches Q that can be invoked in parallel can be determined by rounding up, where Q does not exceed the total number of branches K within the plugin.

[0057] Finally, after determining the number Q of suitable identification branches, the coal gangue identification plugin calls Q identification branches to perform parallel analysis on the received compressed data packets. Each called branch independently performs feature decompression, feature extraction, and classification inference processes on the same compressed data packet, producing a preliminary identification result. Finally, by integrating the outputs of the Q branches, such as using majority voting or probability averaging strategies, a final, consensus-based coal gangue identification result is generated.

[0058] S50: Using the parameters to optimize computing power, based on the coal gangue identification results, iteratively optimize the control parameters of the gangue sorting robot to obtain an optimized control scheme and perform low-latency control of the gangue sorting robot.

[0059] Specifically, the computing power is optimized using the aforementioned parameters. Based on the coal and gangue identification results, the control parameters of the gangue sorting robot are iteratively optimized to obtain an optimized control scheme, and low-latency control of the gangue sorting robot is performed, including: The computing power is optimized based on the parameters, the optimization round limit is obtained, and the iterative optimization step size is obtained based on the coal identification ratio of the coal gangue identification results. The effectiveness of control parameters is calculated based on the control response time and selection efficiency of the coal sorting robot. Based on the optimization round limit and the iteration optimization step size, the control parameters are iteratively optimized to obtain the control parameters with the greatest effectiveness, which are then used as the optimized control scheme for low-latency control of the coal selection robot.

[0060] First, based on the currently allocated computing power for parameter optimization, the maximum number of iterations allowed in this optimization process is determined, i.e., the optimization iteration limit. Specifically, more computing power for parameter optimization means more abundant computing resources available for parameter search, and the number of allowed optimization iterations also increases accordingly, making it possible to conduct a more comprehensive and detailed exploration of the parameter space.

[0061] Optionally, the optimization round limit is equal to the parameter optimization computational power divided by the average computational cost per iteration, with the result rounded down. Here, parameter optimization computational power refers to the amount of computing resources currently allocated to the parameter optimization task, quantified in floating-point operations per second; the average computational cost per iteration refers to the average computing resources consumed in executing one complete control parameter iteration optimization operation, obtained through prior offline performance profiling and calibration of the selected optimization algorithm, also in floating-point operations.

[0062] For example, if the allocated computing power for parameter optimization is 5 billion floating-point operations per second, and the average computational cost per iteration is 200 million floating-point operations per second, then the number of optimization iterations is limited to 25. This quantification method ensures that the optimization process is strictly executed within the allocated computing power budget, avoiding overspending of computing resources, while ensuring that the parameter space is explored as fully as possible within limited resources, thereby achieving precise control and efficient utilization of computing resources.

[0063] Simultaneously, based on the real-time coal and gangue identification results, the percentage of coal-identified material in the total identified material is extracted, i.e., the coal identification percentage. The iterative optimization step size is then set based on this coal identification percentage. Specifically, a high coal identification percentage indicates that the current material is predominantly coal, the operating conditions are relatively stable, and a smaller optimization step size can be used for fine-tuning; conversely, a low percentage indicates a high proportion of gangue or mixed materials, the operating conditions are complex, and a larger optimization step size may be needed to quickly adjust the control strategy to adapt to changes.

[0064] Optional, optimize step size s=s max -(s max -s min ) × Coal labeling percentage. Among them, s max With s min These are the pre-defined maximum and minimum allowable optimization step sizes, respectively. For example, s can be set. max It is 0.5, s min The step size is set to 0.1. When the coal identification percentage is 1, the optimization step size is set to the minimum value of 0.1 for fine adjustment; when the coal identification percentage is 0, the optimization step size is set to the maximum value of 0.5 for rapid response. This quantification mechanism allows the optimization step size to be adaptively adjusted according to the real-time changes in material composition, thereby pursuing control accuracy under stable operating conditions and prioritizing adjustment speed under complex operating conditions.

[0065] Secondly, the effectiveness of the control parameters is calculated based on the control response time and sorting efficiency of the sorting robot. This control parameter effectiveness is a quantitative indicator that comprehensively evaluates control performance, and its calculation integrates two key performance parameters of the sorting robot: control response time and sorting efficiency. Control response time refers to the delay from issuing a control command to the execution mechanism completing the action; sorting efficiency combines the number of correctly sorted materials per unit time with the missorting rate. Preferably, the control parameter effectiveness = η × E - θ × T. Here, E represents the normalized sorting efficiency value, T represents the normalized control response time, and η and θ are the weighting coefficients for efficiency and response time, respectively, used to adjust their relative importance in the comprehensive evaluation. Their values ​​are set according to the priority requirements of efficiency and speed in actual production; for example, η can be set to 0.8 and θ to 0.2. A higher control parameter effectiveness means that the set of control parameters can achieve higher sorting efficiency in a shorter response time.

[0066] Finally, within the set optimization round limit, starting from the current control parameters, a guided search is performed in the multidimensional space of control parameters according to the determined iterative optimization step size. After each iteration to adjust the control parameters, the expected impact on control response time and coal selection efficiency is simulated or quickly evaluated based on the updated parameters, and the corresponding control parameter effectiveness is calculated. By comparing the effectiveness values ​​under different parameter combinations, historical optimal values ​​and their corresponding parameters are tracked and retained. This iterative process continues until the optimization round limit is reached. At this point, the control parameters corresponding to the maximum control parameter effectiveness obtained in the entire search process are determined as the final output of this optimization, i.e., the optimized control scheme.

[0067] Finally, the parameters in the optimized control scheme are deployed to the underlying controller of the sorting robot in real time, driving the robot to perform sorting operations based on the latest recognition results and the optimized action strategy.

[0068] In summary, this closed-loop optimization mechanism ensures that the control parameters can adapt to the real-time changes in material composition and operating conditions, achieving near-optimal low-latency control within limited computing resources and time.

[0069] In summary, the embodiments of this application have at least the following technical effects: Compared to existing technologies, this invention achieves adaptive and precise compression of multimodal coal and gangue data by constructing a dynamic optimization model between conveyor belt operation parameters and data compression ratio, effectively reducing network transmission load and latency. Simultaneously, it flexibly allocates edge computing power based on the characteristics and size of compressed data packets, ensuring efficient collaboration of limited computing resources between identification analysis and parameter optimization tasks. Furthermore, this invention maximizes coal and gangue identification accuracy under strict latency constraints through a scalable identification plug-in structure and a computing power-driven branch scheduling mechanism. Moreover, iterative optimization of control parameters based on real-time identification results enables the robot to quickly adapt to complex and changing material conditions, ultimately achieving an optimized balance between low latency, high accuracy, and high robustness in the gangue sorting process, improving the overall efficiency and stability of the sorting system.

[0070] Example 2, as Figure 2 As shown, based on the same inventive concept as the low-latency control method for a coal sorting robot based on edge computing provided in Embodiment 1, this embodiment of the invention also provides a low-latency control system for a coal sorting robot based on edge computing, including: Data acquisition module 11 is used to acquire belt conveyor operating parameters and coal gangue multimodal data; Data compression module 12 is used to obtain the optimal compression ratio based on the belt running parameters, compress the coal gangue multimodal data, and obtain a compressed data packet; The computing power allocation module 13 is used to send the compressed data packet to the edge node of the coal gangue sorting robot, and obtain the computing power allocation ratio based on the size of the compressed data packet and the multimodal data characteristics of coal gangue, allocate the computing power of the edge node, and obtain the computing power allocation result. The computing power allocation result includes data analysis computing power and parameter optimization computing power. Data analysis module 14 is used to perform data analysis on the compressed data packet using the data analysis computing power to obtain coal gangue identification results; The optimization and control module 15 is used to optimize computing power using the parameters, iteratively optimize the control parameters of the gangue sorting robot based on the coal gangue identification results, obtain an optimized control scheme, and perform low-latency control of the gangue sorting robot.

[0071] Specifically, the data acquisition module 11 is used for: Acquire conveyor belt operating parameters and multimodal data of coal and gangue, including: Obtain belt operating parameters, wherein the belt operating parameters include belt operating speed and belt operating fluctuation coefficient; Obtain multimodal data of coal gangue, wherein the multimodal data of coal gangue includes coal gangue texture information and coal gangue grayscale information.

[0072] Specifically, the data compression module 12 is used for: Based on the belt conveyor operating parameters, the optimal compression ratio is obtained, and the coal and gangue multimodal data is compressed to obtain a compressed data package, including: Construct a compression ratio optimization model; Input the belt running parameters into the compression ratio optimization model to obtain the speed compression ratio; Based on the coal and gangue multimodal data, the proportion of coal and gangue texture information is obtained, and based on the proportion of coal and gangue texture information, the mass compression ratio is obtained; The velocity compression ratio and the mass compression ratio are combined and calculated to compress the coal gangue multimodal data and obtain a compressed data package.

[0073] Specifically, a compression ratio optimization model is constructed, including: Obtain the sample belt operation parameter set, and obtain the data packet compression ratio of the control delay of the coal selection robot under the sample belt operation parameters that is less than or equal to the control delay threshold, as the sample compression ratio set; A compression ratio optimization model is constructed, using the sample belt running parameter set as input and the sample compression ratio set as supervision, and the compression ratio optimization model is trained until convergence.

[0074] Specifically, the computing power allocation module 13 is used for: The compressed data packet is sent to the edge node of the coal sorting robot, and the computing power allocation ratio is obtained based on the size of the compressed data packet and the multimodal data characteristics of the coal gangue. The computing power of the edge node is allocated, and the computing power allocation result is obtained. The computing power allocation result includes data analysis computing power and parameter optimization computing power, including: Based on the multimodal data characteristics of the coal gangue, the characteristic computing power allocation ratio is obtained; Based on the ratio of the compressed data packet size to the standard compressed data packet size, the data volume computing power allocation ratio is obtained; The computing power allocation ratio is obtained by weighting the feature computing power allocation ratio and the data volume computing power allocation ratio. The computing power is allocated to the edge nodes using the aforementioned computing power allocation ratio, and the computing power allocation results are obtained. The computing power allocation results include data analysis computing power and parameter optimization computing power.

[0075] Specifically, based on the multimodal data characteristics of the coal gangue, the feature computing power allocation ratio is obtained, including: Based on the coal and gangue identification results of the previous moment, coal and gangue multimodal data features are obtained, wherein the coal and gangue multimodal data features include the proportion of historical coal identification. Based on the historical coal identification ratio and the coal gangue texture information ratio, the identification difficulty coefficient is obtained; Based on the recognition difficulty coefficient, the feature computing power allocation ratio is obtained.

[0076] Specifically, the data analysis module 14 is used for: Using the aforementioned data analysis computing power, the compressed data packet is analyzed to obtain coal gangue identification results, including: Obtain a coal gangue identification plugin, wherein the coal gangue identification plugin includes K identification branches; Based on the ratio of the data analysis computing power to the preset data analysis computing power, the number of adaptive identification branches Q is obtained. Q identification branches are called in the coal gangue identification plugin to perform data analysis on the compressed data package and obtain the coal gangue identification result, wherein Q is less than or equal to K.

[0077] Specifically, the coal gangue identification plugin includes: Based on a convolutional neural network, construct K recognition branches; Obtain a set of sample compressed data packets, wherein the sample compressed data packets include sample coal gangue texture information and sample coal gangue grayscale information; Based on the sample compressed data package set, a sample identification result set is obtained, wherein the sample identification result includes coal identification and gangue identification; The sample compressed data package set and the sample recognition result set are randomly sampled with replacement to obtain K training datasets of the same size. The K recognition branches are trained separately until convergence. The K identification branches are integrated using an average integration strategy to obtain the coal gangue identification plugin.

[0078] The optimization and control module 15 is specifically used for: Using the aforementioned parameters to optimize computing power, and based on the coal and gangue identification results, the control parameters of the gangue sorting robot are iteratively optimized to obtain an optimized control scheme, enabling low-latency control of the gangue sorting robot, including: The computing power is optimized based on the parameters, the optimization round limit is obtained, and the iterative optimization step size is obtained based on the coal identification ratio of the coal gangue identification results. The effectiveness of control parameters is calculated based on the control response time and selection efficiency of the coal sorting robot. Based on the optimization round limit and the iteration optimization step size, the control parameters are iteratively optimized to obtain the control parameters with the greatest effectiveness, which are then used as the optimized control scheme for low-latency control of the coal selection robot.

Claims

1. A low-latency control method for a coal sorting robot based on edge computing, characterized in that, include: Acquire conveyor belt operating parameters and coal and gangue multimodal data; Based on the belt conveyor operating parameters, the optimal compression ratio is obtained, and the coal and gangue multimodal data is compressed to obtain a compressed data package, including: Construct a compression ratio optimization model; Input the belt running parameters into the compression ratio optimization model to obtain the speed compression ratio; Based on the coal and gangue multimodal data, the proportion of coal and gangue texture information is obtained, and based on the proportion of coal and gangue texture information, the mass compression ratio is obtained; The velocity compression ratio and the mass compression ratio are combined and calculated to compress the coal gangue multimodal data and obtain a compressed data package; The compressed data packet is sent to the edge node of the coal sorting robot, and the computing power allocation ratio is obtained based on the size of the compressed data packet and the multimodal data characteristics of coal gangue. The computing power of the edge node is allocated, and the computing power allocation result is obtained. The computing power allocation result includes data analysis computing power and parameter optimization computing power. Using the aforementioned data analysis computing power, the compressed data packet is analyzed to obtain coal gangue identification results; The computing power is optimized using the parameters, and the control parameters of the gangue sorting robot are iteratively optimized based on the coal gangue identification results to obtain an optimized control scheme and perform low-latency control of the gangue sorting robot.

2. The low-latency control method for a coal sorting robot based on edge computing according to claim 1, characterized in that, Acquire conveyor belt operating parameters and multimodal data of coal and gangue, including: Obtain belt operating parameters, wherein the belt operating parameters include belt operating speed and belt operating fluctuation coefficient; Obtain multimodal data of coal gangue, wherein the multimodal data of coal gangue includes coal gangue texture information and coal gangue grayscale information.

3. The low-latency control method for a coal sorting robot based on edge computing according to claim 1, characterized in that, Construct a compression ratio optimization model, including: Obtain the sample belt operation parameter set, and obtain the data packet compression ratio of the control delay of the coal selection robot under the sample belt operation parameters that is less than or equal to the control delay threshold, as the sample compression ratio set; A compression ratio optimization model is constructed, using the sample belt running parameter set as input and the sample compression ratio set as supervision, and the compression ratio optimization model is trained until convergence.

4. The low-latency control method for a coal sorting robot based on edge computing according to claim 1, characterized in that, The compressed data packet is sent to the edge node of the coal sorting robot, and the computing power allocation ratio is obtained based on the size of the compressed data packet and the multimodal data characteristics of the coal gangue. The computing power of the edge node is allocated, and the computing power allocation result is obtained. The computing power allocation result includes data analysis computing power and parameter optimization computing power, including: Based on the multimodal data characteristics of the coal gangue, the characteristic computing power allocation ratio is obtained; Based on the ratio of the compressed data packet size to the standard compressed data packet size, the data volume computing power allocation ratio is obtained; The computing power allocation ratio is obtained by weighting the feature computing power allocation ratio and the data volume computing power allocation ratio. The computing power is allocated to the edge nodes using the aforementioned computing power allocation ratio, and the computing power allocation results are obtained. The computing power allocation results include data analysis computing power and parameter optimization computing power.

5. The low-latency control method for a coal sorting robot based on edge computing according to claim 4, characterized in that, Based on the multimodal data characteristics of the coal gangue, the feature computing power allocation ratio is obtained, including: Based on the coal and gangue identification results of the previous moment, coal and gangue multimodal data features are obtained, wherein the coal and gangue multimodal data features include the proportion of historical coal identification. Based on the historical coal identification ratio and the coal gangue texture information ratio, the identification difficulty coefficient is obtained; Based on the recognition difficulty coefficient, the feature computing power allocation ratio is obtained.

6. The low-latency control method for a coal sorting robot based on edge computing according to claim 1, characterized in that, Using the aforementioned data analysis computing power, the compressed data packet is analyzed to obtain coal gangue identification results, including: Obtain a coal gangue identification plugin, wherein the coal gangue identification plugin includes K identification branches; Based on the ratio of the data analysis computing power to the preset data analysis computing power, the number of adaptive identification branches Q is obtained. Q identification branches are called in the coal gangue identification plugin to perform data analysis on the compressed data package and obtain the coal gangue identification result, wherein Q is less than or equal to K.

7. The low-latency control method for a coal sorting robot based on edge computing according to claim 1, characterized in that, Obtain the coal gangue identification plugin, including: Based on a convolutional neural network, construct K recognition branches; Obtain a set of sample compressed data packets, wherein the sample compressed data packets include sample coal gangue texture information and sample coal gangue grayscale information; Based on the sample compressed data package set, a sample identification result set is obtained, wherein the sample identification result includes coal identification and gangue identification; The sample compressed data package set and the sample recognition result set are randomly sampled with replacement to obtain K training datasets of the same size. The K recognition branches are trained separately until convergence. The K identification branches are integrated using an average integration strategy to obtain the coal gangue identification plugin.

8. The low-latency control method for a coal sorting robot based on edge computing according to claim 1, characterized in that, Using the aforementioned parameters to optimize computing power, and based on the coal and gangue identification results, the control parameters of the gangue sorting robot are iteratively optimized to obtain an optimized control scheme, enabling low-latency control of the gangue sorting robot, including: The computing power is optimized based on the parameters, the optimization round limit is obtained, and the iterative optimization step size is obtained based on the coal identification ratio of the coal gangue identification results. The effectiveness of control parameters is calculated based on the control response time and selection efficiency of the coal sorting robot. Based on the optimization round limit and the iteration optimization step size, the control parameters are iteratively optimized to obtain the control parameters with the greatest effectiveness, which are then used as the optimized control scheme for low-latency control of the coal selection robot.

9. A low-latency control system for a coal sorting robot based on edge computing, characterized in that, The method for executing the low-latency control of a coal sorting robot based on edge computing as described in any one of claims 1-8 includes: The data acquisition module is used to acquire conveyor belt operating parameters and multimodal data of coal and gangue. The data compression module is used to obtain the optimal compression ratio based on the belt conveyor operating parameters, compress the coal and gangue multimodal data, and obtain a compressed data package, including: Construct a compression ratio optimization model; Input the belt running parameters into the compression ratio optimization model to obtain the speed compression ratio; Based on the coal and gangue multimodal data, the proportion of coal and gangue texture information is obtained, and based on the proportion of coal and gangue texture information, the mass compression ratio is obtained; The velocity compression ratio and the mass compression ratio are combined and calculated to compress the coal gangue multimodal data and obtain a compressed data package; The computing power allocation module is used to send the compressed data packet to the edge node of the coal gangue sorting robot, and obtain the computing power allocation ratio based on the size of the compressed data packet and the multimodal data characteristics of coal gangue, allocate the computing power of the edge node, and obtain the computing power allocation result, wherein the computing power allocation result includes data analysis computing power and parameter optimization computing power; The data analysis module is used to perform data analysis on the compressed data packet using the data analysis computing power to obtain coal gangue identification results; The optimization and control module is used to optimize computing power using the parameters, iteratively optimize the control parameters of the coal gangue sorting robot based on the coal gangue identification results, obtain an optimized control scheme, and perform low-latency control of the coal gangue sorting robot.

Images