Self-optimized injection molding prediction and monitoring method

Abstract

The invention discloses a self-optimized injection molding prediction and monitoring method. The method comprises a data acquisition module, a data analysis module, a remote monitoring module, an appearance detection module, a fault processing module and a process adjustment module. And the data analysis module trains a process parameter prediction model and a defect prediction model by using historical production data to realize real-time prediction of key process parameters and early warning of defects. And the process adjustment module adaptively optimizes the process parameter setting of the injection molding machine according to the prediction result, so that the stability of the product quality is ensured, and human intervention is reduced. And the remote monitoring module is connected with a smart phone and a PC terminal of a user through the Internet, so that a technician can check the production state at any time and intervene in time when the system cannot be automatically adjusted, long-time shutdown of equipment and damage to the product quality are avoided, and the stability and efficiency of a production line are further improved.

Description

Technical Field

[0001] The present invention relates to the technical field of injection molding, and in particular to a self-optimizable injection molding prediction and monitoring method. Background Art

[0002] Amidst the evolving technology landscape for remote monitoring and data sharing of injection molding machines, research and application of remote control primarily focuses on enabling real-time transmission of equipment data and operational status query and feedback. However, current research and system development remain insufficient to fully explore and utilize the deeper information value inherent in this real-time data. During operation, injection molding machines continuously generate a vast amount of data closely related to product quality, equipment operating status, and process parameters. This data not only forms the core foundation for the effective implementation of remote monitoring systems but also contains crucial information on equipment operating trends, fault evolution, and potential for process adjustments. Therefore, further integrating data mining techniques with intelligent predictive models within existing remote control system architectures could significantly enhance the system's intelligence and autonomous decision-making capabilities. It could also provide more scientific support for equipment operation and maintenance management, the establishment of fault warning mechanisms, and the optimization of production decisions. This development is expected to enhance the overall reliability, management, and production efficiency of injection molding equipment, ultimately enabling the transformation and upgrading of production systems towards intelligent systems.

[0003] In traditional injection molding, equipment typically relies on preset, fixed process parameters to complete the injection molding process. However, in real-world production environments, due to factors such as fluctuating environmental conditions, inconsistent raw material properties, and dynamic changes in equipment operating conditions, these fixed parameter settings struggle to accurately adapt to complex and constantly changing operating conditions. This leads to fluctuations in product quality and makes it difficult to consistently ensure high-quality manufacturing. This fixed parameter control model places a high degree of reliance on operators, requiring them to frequently intervene and adjust parameters during the production process to address anomalies or changes. This not only increases labor input and reduces overall production efficiency, but also poses quality risks such as product defects or equipment failures caused by improper parameter settings due to human judgment errors. To address these issues, there is an urgent need to develop an injection molding prediction and monitoring method with self-optimizing adjustment capabilities. This method should be based on real-time data acquisition and analysis, integrating intelligent prediction models with dynamic optimization algorithms to achieve self-adjustment and closed-loop control of injection molding process parameters. This approach can not only effectively improve the consistency and stability of product quality, but also significantly enhance the automation and intelligence level of injection molding production, thereby achieving the goal of efficient, stable, and low-error intelligent manufacturing. Summary of the Invention

[0004] To solve the above problems, the present invention provides a self-optimizing injection molding prediction and monitoring method.

[0005] A self-optimizing injection molding prediction and monitoring method includes a data acquisition module, a data analysis module, a remote monitoring module, an appearance inspection module, a fault handling module, and a process adjustment module. The remote monitoring module connects to the user's smartphone and PC via the internet, enabling automated production and testing of injection molding processes. This avoids long equipment downtime and product quality impairments, thereby improving the stability and efficiency of the production line. The specific technical solution is as follows: Step S1: This step is performed before actual production. Based on historical production record data, the injection molding machine data corresponding to different injection molding products are extracted, including process setting parameters, parameters collected by various sensors, and defect information of molded products. These data are used to train the parameter prediction model and the defect prediction model in advance. Step S2: The data acquisition module periodically collects parameters of the injection molding machine, plastic particles, and the environment through various sensors, and transmits the data to the data analysis module on the PC through a wireless network; Step S3: The data analysis module preprocesses the input data, including removing noise and outliers in the data, in order to improve the quality and usability of the data. The processed data is sent to the remote monitoring module and the parameter prediction model respectively; Step S4: The remote monitoring module displays the received data in real time on the user side (such as a smartphone and a PC) in the form of a line chart; Step S5: The preprocessed data is first standardized and normalized, and then principal component analysis (PCA) is used for dimensionality reduction. The original data is converted into a set of independent principal components through linear transformation, and the principal components with a cumulative variance contribution rate exceeding 90% are selected based on the size of the eigenvalues, thereby reducing the data dimension, reducing redundant features, and improving the efficiency of model training. Step S6: Utilizing the parameter prediction model trained in Step S1, the data is subjected to parameter prediction analysis to obtain predicted data. Subsequently, the predicted data is re-entered into the trained defect prediction model for fault trend analysis to obtain fault trend data. The fault handling module classifies the status (normal, warning, or alert) based on the fault prediction data and sends the predicted data and fault level to the remote monitoring module, which then visualizes this data to the user end and the process adjustment module. Step S7: The appearance inspection module takes photos of the completed injection molded product, determines the quality of each part, and uploads images or reports any anomalies. The appearance inspection module detects defects such as no defects, color mixing, black spots, bubbles, shrinkage, cracks, burrs, poor demolding, weld lines, and short shots. Step S8: The process parameter adjustment module adjusts the process parameter settings based on the fault prediction data and the multi-objective optimization algorithm to achieve the best balance between the various objectives, outputs the optimal process parameter combination, and transmits it to the injection molding equipment to complete the closed-loop control; The injection molding equipment operating parameters include equipment operating voltage, melt temperature, injection speed, injection time, injection end point, clamping force, mold temperature and cooling time; The injection molding product parameters include the weight of the injection molding part and the quality appearance pictures; Furthermore, in step S1, the steps of using historical production parameter data to train the parameter prediction model and the defect prediction model in advance are detailed as follows: Taking into account the time series characteristics of injection molding process data, the random forest training process is improved: the historical production parameters with timestamps are divided into training sets and validation sets according to the production time sequence. The data in the first 80% time period is the training set, and the data in the last 20% time period is the validation set. This method can retain the temporal correlation between parameters and avoid the destruction of temporal causality due to random segmentation; the window sliding method is used to incrementally train the data, that is, I set a time window of fixed length, and then scroll the window range on the time axis in sequence to perform phased updates and iterative optimization on the model, and continue iterative training until the model with the best performance is obtained.

[0006] To address the misclassification issues inherent in traditional approaches due to nonlinear coupling of injection molding parameters (e.g., defects caused by the coordinated variation of multiple process parameters), I developed a multi-defect classification system based on an online incremental random forest classifier. This classification system possesses dynamic learning capabilities, continuously ingesting new samples from the actual production process and adjusting model parameters in real time as the data stream is continuously updated, thereby maintaining high adaptability to current operating conditions and high recognition accuracy. The specific implementation process involves first dividing a complete injection molding cycle into several key process stages: plasticization, injection, holding, and cooling. For each stage, representative statistical features, such as mean, range, and integral, are extracted from sensor data to capture the changing characteristics of the process state at each stage. Feature importance scores are calculated by permutation importance, and redundant features with low importance scores are removed. For highly important but highly fluctuating parameters (such as clamping pressure), a moving average filter is introduced to smooth the data.

[0007] Furthermore, in step S2, the sensors include but are not limited to temperature sensors, shaft speed sensors, pressure sensors, speed sensors, position sensors, etc.; Furthermore, in step S3, the data received by the sensor fluctuates greatly, requiring denoising and filtering. The algorithms used include clipping, recursive averaging, and median filtering. Clipping effectively removes the effects of abnormal waveforms in the data; recursive averaging takes a weighted average of continuously sampled data, making it suitable for online systems and yielding a smooth filtered output. Finally, the filtered data is normalized and standardized to improve data quality and the accuracy of subsequent analysis. The recursive averaging filter function is as follows:

[0008]

[0009] Furthermore, step S4 constructs a four-layer structure of "device-network-data-interface," enabling fully digital management and control of the injection molding workshop through wireless IoT technology. Core components include the injection molding machine, a wireless communication gateway, an SQLite database server, and a visual control center based on the Qt framework.

[0010] The communication module uses the TCP / IP protocol and utilizes Qt's built-in QTcpSocket and QTcpServer components for easy integration, ensuring real-time and reliable data transmission. The data storage component uses SQLite, which is directly supported by QT's QSQLite driver. The storage structure includes fields such as timestamps and parameter values, making it easy to store and retrieve historical data. In terms of the visual interface, Qt's QWidget and QCharts modules are used to draw line charts for multiple chart areas, and abnormal states are intuitively identified through color and shape. This solution fully leverages Qt's cross-platform advantages to achieve decoupled development of interface components and business logic, and combines edge computing to achieve microsecond response. Compared with traditional solutions, it significantly improves development efficiency and system stability, effectively reduces maintenance costs, and shortens operator training cycles.

[0011] Furthermore, in step S5, after extracting the data from the SQLite database, the data is first standardized and normalized to eliminate the influence of the differences in the dimensions of different parameters and their absolute values on the analysis results. Subsequently, dimensionality reduction is performed through principal component analysis (PCA), and an orthogonal transformation is used to project the original feature space into a low-dimensional subspace with maximum variance.

[0012] Furthermore, in step S6, the preprocessed data is input into the pre-trained model to output the prediction results of subsequent parameters and defect judgment.

[0013] Furthermore, in step S7, before production, a full-scale photograph of the standard injection molded parts is taken as a standard library, and feature points are extracted from each part of the parts; during the production process, the production parts are also photographed and compared with each part of the standard parts to calculate the similarity. By setting a similarity threshold to judge the quality of each part of the injection molded product, abnormal parts can be identified, providing a reference for solving injection molding anomalies. The comparison is divided into three levels: qualified, warning and unqualified. When the similarity is higher than the qualified threshold, it is judged as a qualified part; when the similarity is lower than the qualified threshold but higher than the unqualified threshold, it is judged as a warning part; when the similarity is lower than the unqualified threshold, it is judged as an unqualified part and the most dangerous warning is sent to the system. In the latter case, the system will not be able to automatically optimize and adjust the parameters. At this time, manual intervention is required to judge the problems of the injection molded product and equipment, which usually indicates that the equipment has an abnormality.

[0014] Furthermore, in step S8, the parameter optimization module derives predicted process parameters based on the predicted parameters and predicted defect data obtained in step 6 and in combination with a multi-objective optimization algorithm; and then adjusts the parameters of the injection molding machine based on these predicted process parameters to achieve injection molding production under a more optimal state.

[0015] It can be seen from the technical solution of this application that this application has the following advantages: The technical solution of this application is based on historical injection molding product data (including normal samples and defect samples), extracting process parameters (such as temperature, pressure, etc.) and defect characteristics. After classifying them by product category, the process parameter prediction model and defect classification model are simultaneously trained using the random forest algorithm. After the model is launched, real-time production data is collected and input into the prediction system to dynamically correct parameter deviations and identify potential defects. Based on the prediction results, the equipment operating parameters are automatically optimized or a graded shutdown protection mechanism is triggered, achieving automatic process parameter tuning, real-time defect warning, and intelligent equipment control, effectively improving production stability and fault response efficiency. Multi-angle cameras combined with deep learning algorithms enable intelligent location and diagnosis of product surface defects. Process personnel can quickly identify product anomalies and their causes by accessing defect location diagrams and the manufacturer's problem cause analysis database, accelerating maintenance response times while reducing technical requirements for process personnel and training costs. The process parameter adjustment module continuously optimizes process parameter settings based on actual production data, greatly reducing human intervention. The closed-loop adjustment mechanism ensures a more stable production process for the injection molding system and more consistent quality for products within the same batch. Remote monitoring, the ultimate safeguard for the entire system, allows process personnel to monitor the status of on-site injection molding equipment at any time via smart devices (such as smartphones and PC clients). Visualized line charts and data summaries provide clear and convenient product traceability. When the system encounters an anomaly that cannot be automatically handled, the remote monitoring module issues a three-level alarm and generates a problem report via the smart device, effectively avoiding the lag of manual monitoring. It also automatically adjusts the equipment's operating status based on the severity of the fault, enabling faster response times for process personnel and reducing downtime, significantly improving production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments described in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0017] Figure 1 It is a flow chart of each step of the present invention; Figure 2 It is a flow chart of the relationship between the modules of the present invention; Figure 3 This is the process of training the model in step S1 of the present invention; Figure 4 This is the logic flow of visual defect determination in step S7 of the present invention; DETAILED DESCRIPTION

[0018] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0019] First, a detailed description of the implementation of a self-optimizing injection molding prediction and monitoring method provided by this application is as follows: Figure 3 The present invention provides a method for training injection molding product parameter prediction and defect prediction, comprising: Step S101: Obtain historical production record data of the injection molding machine corresponding to the injection molded product; It should be noted that the historical production record data obtained includes the process setting parameters of the injection molding machine, the parameters output by each sensor, and the defect information of the molded products.

[0020] Step S102 performs denoising and standardization on the multi-source sensor features in the historical production data to generate training samples for the parameter prediction model; at the same time, classification features are extracted based on the process parameters and defect labels of the same batch of data to construct a training set for the defect prediction model; It should be noted that the above-mentioned denoising and standardization operations include PCA dimensionality reduction, which involves first denoising, standardizing, and normalizing the historical production record data, then calculating the eigenvalues of various parameters and selecting the principal components with a cumulative variance contribution rate greater than 90%. Defect information is used to measure the defect category of the injection molded product, and this step does not involve the evaluation of the appearance of the injection molded parts. Next, the injection molding machine setting parameters, sensor parameters and defect information are classified according to the timestamps to obtain several parameter groups with the timestamps as the main body.

[0021] Step S103 trains the model in advance so that the model does not need to be retrained in subsequent steps. The main advantage of using the sliding average method is that it can reduce the fluctuation of parameter updates, thereby improving the robustness of the model. It is particularly suitable for scenarios where the data distribution changes over time. The formula of the sliding average method is as follows:

[0022] In the above formula, α∈[0, 1). When α=0, it is equivalent to not using sliding average. θ_t is the value of the current time step.

[0023] In the defect prediction model, the 3σ principle (based on the normal distribution assumption) is first used to determine the first-level defects. The formula is as follows:

[0024] In the above formula, W_t is the current value of the input, μ_w is the mean of the sliding window, and σ_w is the standard deviation of the sliding window.

[0025] The second layer uses the random forest probability threshold method to judge defects. For the input data sample, each tree in the random forest will give a classification result. By counting the classification results of all trees, the predicted probability of the defective category is calculated. The formula is as follows:

[0026] In the above formula, I(T_i=c) indicates whether the i-th tree predicts category c; N represents the total number of trees in the forest.

[0027] The set probability threshold needs to be adjusted based on the training test results. The optimal threshold is analyzed using the ROC curve (Receiver Operating Characteristic Curve) to achieve the best balance between the false positive rate (FPR) and the true positive rate (TPR). The following is how to calculate FPR and TPR:

[0028] In the above formula, FP is a sample that is actually normal but is mistakenly judged as abnormal; TN is a sample that is correctly predicted to be normal; TP is a sample that is predicted to be abnormal; FN is a sample that is actually abnormal but is mistakenly judged to be normal. On the ROC curve, each point corresponds to a different probability threshold.

[0029] In the process of predicting defects, we hope to keep the FPR (false positive rate) low while maintaining a high TPR (true positive rate) so as to effectively reduce false positives while discovering more anomalies.

[0030] The Youden's Index is combined with the ROC curve to first determine the index value corresponding to the threshold, and then the index value is brought into the ROC curve to obtain the optimal threshold with the largest Youden's Index.

[0031]

[0032] In summary, this defect prediction model sets two levels of defect threshold judgment.

[0033] Step S104: This step primarily involves intelligently and adaptively adjusting process parameter settings based on defects and operating parameters. Because injection molding equipment has a wide variety of operating parameters, and these parameters must coordinate with each other to effectively optimize the injection molding process, the process adjustment module uses the NSGA-II (Non-Dominated Sorting Genetic Algorithm II) for real-time optimization.

[0034] First, the population is initialized, with its size and decision variable ranges (the ranges of various process parameters). The weighted defect rate is used as the objective function metric, requiring solution A to be no less effective than solution B on all objectives and superior on at least one objective. A fast non-dominated sort is then performed, requiring solution A to dominate solution B if the following conditions are met:

[0035] The dominance relationship of each solution is then calculated, and individuals that are not dominated by any solution (the number of dominated solutions is 0) are found. This serves as the first Pareto front. This process is repeated until all individuals have been stratified. The crowding degree of solutions on the same front is then calculated, and individuals with higher crowding degrees are prioritized to maintain solution diversity. Solutions are then selected for crossover mutation, following the principle of prioritizing the Pareto front; if the number of strata is equal, individuals with higher crowding degrees are selected.

[0036] Simulate binary crossover and polynomial mutation to generate new offspring populations.

[0037]

[0038] The parent and child populations are then continuously merged, and non-dominated sorting is repeated, selecting outstanding individuals to form the next generation population until the Pareto front converges (i.e., the change in the frontier solution is less than a preset threshold). Throughout the optimization process, since decision variables (i.e., adjustable process parameters) directly affect product quality and production efficiency, the range of decision variables must be appropriately set to ensure that a sufficient feasible solution space is explored while avoiding unreasonable parameter combinations.

[0039] The preceding section describes the entire process of training and validating the injection molding product operating parameter prediction model and the real-time injection molding defect prediction model. When the injection molding equipment is started, the data acquisition module periodically collects sensor parameters, including melt temperature, injection speed, injection time, injection endpoint, clamping force, mold temperature, and cooling time, and transmits these data to the data analysis module. Simultaneously, the appearance inspection module's camera captures the final product and transmits the image data to the data analysis module. After obtaining this process parameter information and product images, the data analysis module stores and organizes the data in a SQLite database for subsequent analysis and scheduling. After the process parameter data is processed by the prediction model, the output prediction results are then input into the real-time defect prediction model, thereby establishing a correlation model between process parameters and final product quality.

[0040] Figure 4 The new product image defect analysis method provided by the present invention includes: Step S701: Data collection: collect the final product image dataset corresponding to the historical injection molded products. It should be noted that the collected products include both defect-free products and products with multiple defect types. The collected image data is then preprocessed, including image scaling, cropping, and determining the ROI (region of interest) for detection. The main features of the final product image are then extracted using local binary patterns (LBP); these features include edges, textures, colors, and shapes.

[0041] Finally, the main features of the product image are used as input, and the corresponding defect type label is used as output, and a convolutional neural network is used for training.

[0042] The convolutional neural network constructed by the present invention consists of a convolutional layer, a pooling layer and a residual block.

[0043] The convolutional layer extracts different types of local features from the image and captures the key information of the input image. A ReLU activation function is added immediately after the convolutional layer to introduce nonlinear factors and enhance the expressive power of the model. Pooling layers are used to reduce computational complexity and enhance the model's translation invariance. Common pooling operations include max pooling and average pooling, which generate output by taking the maximum or average value within a neighborhood, respectively.

[0044] To address the vanishing gradient problem that often occurs in deep neural networks, we use residual blocks to ensure better information transmission in the network, thereby improving training stability and overall effectiveness.

[0045] Step S702 performs pre-processing and feature extraction on the image data in the same manner as step S701. After processing, each image is converted into a feature vector representation.

[0046] The core of the matching process in step S703 is to compare the similarity between the feature vector of the new product and the feature vectors of the reference product from each viewpoint. Common similarity metrics include cosine similarity and Euclidean distance. Because each product in the benchmark library contains multiple views, it is necessary to integrate the matching results from each viewpoint into an overall similarity score. Methods such as average similarity, maximum similarity, or weighted average similarity can be used.

[0047] In step S704, the new product is matched with the products or defect types in the reference library according to the overall similarity score, thereby identifying the category or defect of the new product.

[0048] Here are some common product defects, including direct cosmetic defects, and their possible causes: Color mixing: The masterbatch is not mixed evenly, and there are other masterbatches in the screw that were not cleaned out during the previous production.

[0049] Black spots: The injection molding material stays in the injection molding machine for too long or the injection temperature is too high, causing the material to decompose or burn.

[0050] Bubbles: The raw materials are not completely dried or the mold exhaust is abnormal, resulting in water vapor entering the injection molding process or the gas generated during the injection molding process cannot be discharged.

[0051] Shrinkage: The mold temperature is too low or the filling volume is insufficient, resulting in uneven shrinkage of the plastic when it cools and solidifies.

[0052] Crack: The cooling rate is too fast or the plastic toughness is insufficient, and the stress inside the product is released, resulting in cracks.

[0053] Flash / burrs: Mold wear, excessive clearance or excessive injection pressure cause the mold parting surface to not be completely closed.

[0054] Poor demoulding: The ejector pin is worn or the mold guide pin is poorly lubricated, resulting in insufficient ejection force or tilting of the mold.

[0055] Weld line: The material temperature is not high enough or the injection speed is too slow, causing the molten material flow fronts to meet but not completely fuse.

[0056] Short shot: Insufficient injection pressure or injection speed or poor material fluidity results in the mold cavity not being filled with material.

[0057] Furthermore, the defect types mentioned above include no defect, color mixing, black spots, bubbles, shrinkage, cracks, flash / burrs, poor demoulding, weld lines and short shots.

Claims

1. A self-optimizing injection molding prediction and monitoring method, characterized in that: The steps include: Step S1: This step is performed before actual production, using historical production record data to train the parameter prediction model and defect prediction model in advance; Step S2: The data acquisition module periodically collects parameters of the injection molding machine, plastic particles, and the environment through various sensors, and transmits the data to the data analysis module on the PC through a wireless network; Step S3: The data analysis module pre-processes the input data and sends the processed data to the remote monitoring module and the parameter prediction model respectively; Step S4: The remote monitoring module displays the received data in real time on the user side (such as a smartphone and a PC) in the form of a line chart; Step S5: The preprocessed data is converted into a set of independent principal components through linear transformation, which reduces the data dimension and redundant features to improve the model training efficiency; Step S6: Input the device parameters into the parameter prediction model to obtain predicted values of the device parameters. Input the predicted data into the trained defect prediction model to perform fault trend analysis and obtain fault trend data. Step S7: The appearance inspection module takes photos of the injection molded product for full inspection, determines the quality of each part of the injection molded part, and uploads images or reports any abnormalities; Step S8: The process parameter adjustment module adjusts the process parameter settings based on the fault prediction data and the multi-objective optimization algorithm to achieve the best balance between the various objectives, outputs the optimal process parameter combination, and transmits it to the injection molding equipment to complete the closed-loop control.

2. The training method for the parameter prediction model and defect prediction model of the self-optimizable injection molding prediction and monitoring method is characterized by: Step S101: Obtaining historical production record data of the injection molding machine corresponding to the injection molding product; Step S102: The historical production record data is divided into different features according to different sensors, and each feature data is subjected to operations such as denoising and standardization. Then, the setting parameter information and defect information corresponding to each injection molded product sample are extracted and classified to obtain several groups of sample data for the defect prediction model; Step S103: Using a window sliding algorithm for incremental training, the sensor sample data is input into the parameter prediction model to obtain the injection molding product operating parameter prediction results through the operation of the parameter prediction model; using an online incremental random forest classification model, by dynamically constructing time series features (lag / sliding statistics) and adaptive weight adjustment, setting a two-layer defect threshold judgment, and inputting the samples of the defect prediction model into the model training to achieve real-time prediction of injection molding defects; Step S104: The obtained prediction parameters, predicted defect probability and defect category are input, and a multi-objective optimization algorithm is used to find a set of optimal process parameters. The optimized parameters are then sent to the injection molding machine control system to dynamically adjust the production parameters and complete closed-loop control to ensure continuous optimal production.

3. The new product image defect analysis method of the self-optimizable injection molding prediction and monitoring method is characterized in that: The structure of the model includes: Convolutional layer, extracts different types of local features from the image and captures the key information of the input image; Pooling layer, which reduces computational complexity and enhances the translation invariance of the model; Residual blocks are used to solve the vanishing gradient problem that often occurs in deep neural networks and ensure better information transmission in the network.

Images