Intelligent tool scheduling control system for a foundry robot

Abstract

The application discloses a kind of casting parts cleaning robot intelligent tool scheduling control system, adopt point cloud processing module to obtain the workpiece error of casting workpiece and transmission to tool scheduling module obtains the processing tool sequence required in casting process;Process parameter selection module matches process parameter according to processing tool sequence;And adopt motion planning module to obtain processing track path;And processing track path is transmitted to mechanical arm control module;And through the vibration analysis unit of mechanical arm control module, vibration data in processing are generated feedback signal, process parameter selection module decides process parameter or calls tool scheduling module according to feedback signal;Tool scheduling module is based on risk assessment and decision model to feedback signal evaluation, and according to the tool of processing tool sequence or calling point cloud processing module reacquires workpiece error according to evaluation result.Call frequency of camera is greatly reduced by continuously updating tool scheduling module by self-learning, and the production efficiency of casting part and the service life of camera are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent tool scheduling technology, and in particular to an intelligent tool scheduling and control system for a casting cleaning robot. Background Technology

[0002] With technological advancements and rising labor costs, traditional manufacturing sectors are increasingly using robots to replace manual processing methods. In grinding tasks, castings are complex. Due to process limitations, the positions and dimensions of risers, flash, and fixing nails on castings vary, making it difficult to complete grinding tasks quickly and with high quality using a single grinding tool or a fixed process. Because the cutting and high-intensity grinding processes generate a large amount of sparks, dust, and residue that impact the surrounding environment, binocular cameras cannot be continuously mounted on the robotic arm. Typically, after each stage, the camera needs to identify the casting's processing status to make the next decision. Repeatedly calling the camera significantly reduces grinding efficiency and damages the camera, while the price of industrial cameras far exceeds that of the robotic arm and the casting itself. Summary of the Invention

[0003] This invention provides an intelligent tool scheduling and control system for a casting cleaning robot to overcome the above-mentioned technical problems.

[0004] To achieve the above objectives, the technical solution of the present invention is as follows:

[0005] A smart tool scheduling and control system for a casting cleaning robot includes a point cloud processing module, a tool scheduling module, a process parameter selection module, a motion planning module, and a robotic arm control module.

[0006] The point cloud processing module is used to obtain the workpiece error model of the casting workpiece, and transmit the data of the workpiece error model to the process parameter selection module, motion planning module and tool scheduling module respectively.

[0007] The tool scheduling module is used to obtain the sequence of machining tools required for the casting process based on the data of the workpiece error model; and transmit the sequence of machining tools to the process parameter selection module.

[0008] The processing tool sequence includes one or more tool combinations such as an air planer, a high-power grinder, an angle grinder, a high-power belt sander, a fine belt sander, and a telescopic file.

[0009] The process parameter selection module is used to match the process parameters of the corresponding machining tool with the data of the workpiece error model and the machining tool sequence; and transmit the process parameters to the motion planning module.

[0010] The process parameters include initial tool force, tool feed rate, tool entry angle, and tool spindle speed;

[0011] The motion planning module is used to obtain the machining trajectory path based on the workpiece error model data, machining tool sequence, and process parameters; and transmit the machining trajectory path to the robotic arm control module.

[0012] The robotic arm control module includes a robotic arm drive module, a robotic arm end force control unit, and a vibration analysis unit.

[0013] The robotic arm drive module is used to drive the robotic arm to move the machining tool according to the machining trajectory path.

[0014] The end effector force control unit of the robotic arm is used to collect vibration data of the end tool during the operation of the robotic arm through a six-dimensional force control sensor, and transmit the vibration data to the vibration analysis unit;

[0015] The vibration analysis unit generates a feedback signal from the vibration data and transmits it to the robotic arm drive module and the process parameter selection module. The robotic arm drive module controls the start and stop of the robotic arm based on the feedback signal. The process parameter selection module determines the process parameters based on the feedback signal or calls the tool scheduling module to evaluate the machining tool sequence. The tool scheduling module evaluates the feedback signal based on the constructed risk assessment and decision-making model to obtain the evaluation result, and calls the tool of the machining tool sequence or calls the point cloud processing module to re-obtain the workpiece error based on the evaluation result.

[0016] Furthermore, the specific strategy for acquiring the data of the workpiece error model of the cast workpiece is as follows:

[0017] The casting workpiece is scanned using a 3D binocular camera to obtain point cloud data, which is then transmitted to a point cloud processing module. The point cloud processing module generates a model of the casting workpiece to be processed based on the point cloud data.

[0018] The model of the casting workpiece to be processed is divided into regions according to the processing technology of the casting workpiece, and multiple model regions are obtained.

[0019] The spatial offset of each model region is obtained based on the model region and the set standard model, and the data of the workpiece error model of the casting workpiece is obtained based on the spatial offset.

[0020] Furthermore, the tool scheduling module uses a strategy to obtain the sequence of machining tools required for the casting process based on the workpiece error model data.

[0021] The error feature data of the workpiece error model is extracted, and the error feature data includes the maximum error, the average error, and the error distribution; and the error feature data is combined into a state vector as the current state of the workpiece error model.

[0022] Based on the deep reinforcement learning algorithm network DQN, the state space of the workpiece error model, the action space of the tool sequence, and the reward function are defined according to the state of the workpiece error model.

[0023] The state space of the workpiece error model includes different states of the workpiece error model; and based on the experience of casting part processing, a tool sequence table corresponding to each workpiece error model state is set as the action space of the tool sequence.

[0024] The reward function is used to obtain the tool sequence with the largest reward value in the action space of the corresponding tool sequence as the optimal tool sequence based on the current workpiece error model state.

[0025] The expression for the reward function is:

[0026] R(s,a,s')=-E(s')+γ0·E(s)-C(a)

[0027] In the formula: E(s) and E(s') represent the workpiece error value in the current state s and the workpiece error value in the new state s', respectively; γ0 is a discount factor between 0 and 1; C(a) represents the cost of taking action a; s represents the current state, i.e., the workpiece error model of the current casting; a represents the action taken, i.e., the selected tool sequence; s' represents the new state obtained after performing action a; R(s,a,s') represents the reward obtained by taking action a from the current state s and transitioning to the new state s'.

[0028] The state of the workpiece error model and the action of the tool sequence in the action space are used as inputs to the deep reinforcement learning algorithm network DQN. Based on the greedy algorithm, new actions are output according to the reward function and a four-tuple data (state, action, reward, next state) is established.

[0029] Create an experience replay storage pool, store the quadruple data (state, action, reward, nextstate) in the experience replay storage pool, and randomly extract quadruple data from the experience replay storage pool as a subset of training data;

[0030] The Deep Reinforcement Learning Network (DQN) is trained using the training data subset to obtain predicted tool sequence action values. The network parameter weights of the DQN are then updated using the backpropagation method, with the mean squared error between the predicted tool sequence action values ​​and the target tool sequence action values ​​as the loss function.

[0031] The optimal network parameter weights are determined by finding the weights of the Deep Reinforcement Learning (DQN) network that minimize the mean squared error between the predicted tool sequence action values ​​and the target tool sequence action values.

[0032] The optimal deep reinforcement learning algorithm network DQN is obtained based on the optimal network parameter weights.

[0033] The target tool sequence action value Q target The formula for calculating (s,a) is:

[0034]

[0035] In the formula: reward represents the reward obtained after selecting a certain processing tool; s' represents the state; γ0 represents the discount factor; This represents the maximum value of the actions in the tool sequence list under the new state s';

[0036] The sequence of machining tools required for the casting process is obtained based on the optimal deep reinforcement learning algorithm network DQN.

[0037] Furthermore, the process parameter selection module is used to match the process parameters of the corresponding machining tool according to the data of the workpiece error model and the machining tool sequence.

[0038] Based on the workpiece error model data, machining tools, and existing machining process library information, multiple machining process parameter vectors are defined; and a tool process parameter database is obtained based on the machining process parameter vectors.

[0039] The expression for the processing parameter vector is as follows:

[0040] M i =(T,E,F) initial V feed ,θ cut ,R spindle )

[0041] In the formula: T represents the tool type; E represents the workpiece error model size; F initial V represents the initial force; feed Indicates the feed rate; θ cut Indicates the tool's entry angle; R spindle Indicates the spindle speed of the tool;

[0042] Define the desired vector of machining process parameters as S=(E',T',F') initial ,V' feed ,θ' cut ,R' spindle Based on the Euclidean distance calculation function, the machining process parameter vector with the smallest Euclidean distance to the expected vector of the machining process parameters is obtained from the tool process parameter database and used as the current machining process parameter vector to match the corresponding machining tool.

[0043] The Euclidean distance calculation function D(S,M) i The expression for ) is

[0044]

[0045] In the formula: T' represents the expected instrument type; E' represents the expected instrument error model size; F' initial V' represents the desired initial force. feed Indicates the desired feed rate; θ' cut R' represents the desired angle of entry for the tool. spindle Indicates the desired tool spindle speed; Templates represent the parameter vectors in the tool process parameter database; M * This represents the machining parameter vector that has the smallest Euclidean distance to the expected vector of the machining parameters.

[0046] Furthermore, the process parameter selection module is used to determine process parameters based on the feedback signal or to invoke the tool scheduling module to evaluate the feedback signal according to the following decision rules:

[0047] The time-domain / frequency-domain data of the feedback signal is obtained through the process parameter selection module, and it is determined whether the time-domain / frequency-domain data is within a preset decision threshold.

[0048] If the time-domain / frequency-domain data is within the preset decision threshold range, then the current process parameters are confirmed to meet the standard, and the current process parameters are maintained.

[0049] If the time-domain / frequency-domain data exceeds the preset decision threshold range, it is confirmed that the current process parameters do not meet the standard, and the feedback signal is fed back to the tool scheduling module.

[0050] Furthermore, the strategy employed by the tool scheduling module to evaluate the feedback signal and obtain the evaluation result based on the constructed risk assessment and decision-making model is as follows:

[0051] The specific process of constructing a risk assessment and decision-making model is as follows:

[0052] Historical vibration signal data classified by risk labels are collected, and the FFT transformation algorithm is used to obtain the feature values ​​of the vibration signal data corresponding to the risk label classification; and a feature sample training set is obtained based on the feature values.

[0053] The characteristic values ​​include vibration amplitude and frequency components;

[0054] An optimization model for SVM based on radial basis function neural network (RBF) is constructed; the optimization model is as follows:

[0055]

[0056] subject to:

[0057]

[0058] K(x i ,x j )=exp(-γ||x i -x j || 2 )

[0059]

[0060] In the formula: N represents the number of samples in the feature sample training set; y i The sample label represents the feature sample training set, with a value of 1 or -1; α represents the Lagrange multiplier; C represents the regularization parameter; K(x) i ,x j ) represents the kernel function of the support vector machine; x i With x j The data points represent the vibration amplitude and frequency components, respectively; γ represents the kernel parameter of the radial basis function neural network (RBF); x s With y s All represent data points of the preset support vectors; b represents the bias.

[0061] The SMO optimization algorithm is used to solve for the Lagrange multipliers α and bias b of the optimization model; and the classification decision function is obtained based on the Lagrange multipliers α and bias b; the expression of the classification decision function is as follows:

[0062]

[0063] In the formula: f(x) represents the classification prediction result of vibration signal feature x;

[0064] Based on the classification decision function, vibration data features, including the vibration amplitude and frequency components of the vibration signal, are obtained according to the feedback signal output by the vibration analysis unit.

[0065] F(t) = [f1(t), f2(t), ..., f n (t)]

[0066] In the formula: F(t) represents the characteristic value of the vibration data; f i (t) represents the i-th vibration data feature at time t, where i = 1, 2, 3, ..., n.

[0067] The trained optimization model is used to obtain the risk label corresponding to the current vibration data features, and the current vibration data features are compared with the vibration data features f corresponding to the tools stored in the set process library. lib，iIf the Euclidean distance of the vibration feature is less than or equal to a preset vibration threshold, then it is confirmed that the feedback signal does not pose a risk, and the current tool sequence continues to be used.

[0068] If the Euclidean distance of the vibration feature is greater than the preset vibration threshold, it is confirmed that the feedback signal is at risk, and the point cloud processing module is called again to update the workpiece error.

[0069] The formula for calculating the vibration characteristic Euclidean distance D(t) is as follows:

[0070]

[0071] In the formula: f i (t) represents the i-th vibration data feature at time t; f lib，i This indicates the vibration data characteristics corresponding to the tools stored in the set process library.

[0072] This invention discloses an intelligent tool scheduling and control system for a casting cleaning robot. Based on a constructed risk assessment and decision-making model, the tool scheduling module obtains the sequence of processing tools required for the casting process according to the workpiece error model. Through self-learning, the tool scheduling module continuously updates, enabling the rational scheduling of various grinding and cutting tools. The process parameter selection module matches the process parameters of the corresponding processing tools based on the workpiece error and the processing tool sequence. A motion planning module optimizes and obtains the processing trajectory path. Furthermore, the system determines process parameters based on feedback signals transmitted from the vibration analysis module or calls the tool scheduling module to evaluate the processing tool sequence and obtain evaluation results. Based on the evaluation results, the system calls the tools in the processing tool sequence or calls the point cloud processing module to re-obtain the workpiece error, significantly reducing the number of camera calls and improving the production efficiency of castings and the lifespan of the camera. Attached Figure Description

[0073] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0074] Figure 1 This is a schematic diagram of an intelligent tool scheduling and control system for a casting cleaning robot according to the present invention. Detailed Implementation

[0075] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0076] This embodiment provides an intelligent tool scheduling and control system for a casting cleaning robot, such as... Figure 1 As shown, it includes a point cloud processing module, a tool scheduling module, a process parameter selection module, a motion planning module, and a robotic arm control module;

[0077] The point cloud processing module is used to obtain the workpiece error model of the casting workpiece, and transmit the data of the workpiece error model to the process parameter selection module, motion planning module and tool scheduling module respectively.

[0078] The tool scheduling module is used to obtain the sequence of machining tools required for the casting process based on the data of the workpiece error model; and transmit the sequence of machining tools to the process parameter selection module.

[0079] The processing tool sequence includes one or more tool combinations such as an air planer, a high-power grinder, an angle grinder, a high-power belt sander, a fine belt sander, and a telescopic file.

[0080] The process parameter selection module is used to match the process parameters of the corresponding machining tool with the data of the workpiece error model and the machining tool sequence; and transmit the process parameters to the motion planning module.

[0081] The process parameters include initial tool force, tool feed rate, tool entry angle, and tool spindle speed;

[0082] The motion planning module is used to obtain the machining trajectory path based on the workpiece error model data, machining tool sequence, and process parameters; and transmit the machining trajectory path to the robotic arm control module.

[0083] The robotic arm control module includes a robotic arm drive module, a robotic arm end force control unit, and a vibration analysis unit.

[0084] The robotic arm drive module is used to drive the robotic arm to move the machining tool according to the machining trajectory path.

[0085] The end effector force control unit of the robotic arm is used to collect vibration data of the end tool during the operation of the robotic arm through a six-dimensional force control sensor, and transmit the vibration data to the vibration analysis unit;

[0086] The vibration analysis unit generates a feedback signal from the vibration data and transmits it to the robotic arm drive module and the process parameter selection module. The robotic arm drive module controls the start and stop of the robotic arm based on the feedback signal. The process parameter selection module determines the process parameters based on the feedback signal or calls the tool scheduling module to evaluate the feedback signal. The tool scheduling module evaluates the feedback signal based on the constructed risk assessment and decision-making model to obtain the evaluation result, and calls the tool of the processing tool sequence or calls the point cloud processing module to re-obtain the workpiece error based on the evaluation result.

[0087] In this embodiment, a point cloud processing module acquires the workpiece error model of the casting workpiece. A tool scheduling module obtains the sequence of machining tools required for the casting process based on the data from the workpiece error model. A process parameter selection module matches the process parameters of the corresponding machining tools with the workpiece error model data and the machining tool sequence. A motion planning module acquires the machining trajectory path and transmits it to the robotic arm control module. The robotic arm control module performs machining using the tool, and a vibration analysis unit acquires vibration data during machining to generate feedback signals, which are transmitted to the robotic arm drive module and the process parameter selection module. The robotic arm drive module controls the start and stop of the robotic arm based on the feedback signals. The process parameter selection module decides on process parameters or calls the tool scheduling module based on the feedback signals. The tool scheduling module evaluates the feedback signals based on a constructed risk assessment and decision-making model, obtains the evaluation results, and calls the tools in the machining tool sequence or calls the point cloud processing module to reacquire the workpiece error model based on the evaluation results. This allows for the reasonable scheduling of various grinding and cutting tools, adjustment of process parameters and machining trajectories, and continuous updating of the tool scheduling module through self-learning, significantly reducing the number of camera calls and improving the production efficiency of castings and the lifespan of the camera.

[0088] In a specific embodiment, the strategy for obtaining the workpiece error model data of the cast workpiece is as follows:

[0089] The casting workpiece is scanned using a 3D binocular camera to obtain point cloud data, which is then transmitted to a point cloud processing module. The point cloud processing module generates a model of the casting workpiece to be processed based on the point cloud data.

[0090] The model of the casting workpiece to be processed is divided into regions according to the processing technology of the casting workpiece, and multiple model regions are obtained.

[0091] The spatial offset of each model region is obtained based on the model region and the set standard model, and the data of the workpiece error model of the casting workpiece is obtained based on the spatial offset; the data of the workpiece error model specifically includes the maximum error, the average error, and the error distribution data.

[0092] Specifically, comparing the model of the casting workpiece to be processed with a standard CAD model to generate an error model includes the following steps:

[0093] Converting a standard CAD model: The first step is to convert the standard CAD model into a point cloud representation, which can typically be achieved by uniformly sampling the model's surface.

[0094] P CAD =SampleSurface(CAD)

[0095] Among them, P CAD This represents a point cloud obtained by sampling from a standard CAD model;

[0096] Point cloud denoising: Point cloud data obtained from a 3D stereo camera may contain noise, which is denoised using radial basis function (RBF) filtering;

[0097] P clean =Denoise(P raw )

[0098] Among them, P raw This represents point cloud data obtained through scanning with a 3D binocular camera; P clean This represents the denoised point cloud data;

[0099] Point cloud registration: To compare two point clouds, they need to be in the same coordinate system. The Iterative Closest Point (ICP) algorithm is a commonly used registration technique.

[0100] (R, t) = ICP(P) clean P CAD )

[0101] Where R is the rotation matrix and t is the translation vector, used to transform P clean Align to P CAD ;

[0102] Calculating workpiece error: Workpiece error can be calculated for each point at P. clean China and its P CAD The distance between the nearest points in the model is the data of the workpiece error model.

[0103] E i =dist(P clean,i ,nearest(P clean,i ,P CAD ))

[0104] Among them, E i It is point P clean Workpiece error;

[0105] Generate workpiece error model: Calculate the workpiece error E at each point. iMapping to color or other visual representations generates a workpiece error model that can provide visualized error information for each point;

[0106] Determine if the workpiece error meets the requirements.

[0107] ifmax(E) <Threshold then End Process

[0108] Where Threshold represents the preset error threshold; max(E) represents the maximum workpiece error space offset.

[0109] In a specific embodiment, the tool scheduling module uses a strategy to obtain the sequence of machining tools required for the casting process based on the workpiece error model data.

[0110] The error feature data of the workpiece error model is extracted, and the error feature data includes the maximum error, the average error, and the error distribution; and the error feature data is combined into a state vector as the current state of the workpiece error model.

[0111] Based on the deep reinforcement learning algorithm network DQN, the state space of the workpiece error model, the action space of the tool sequence, and the reward function are defined according to the state of the workpiece error model. The state space of the error model contains all possible states, and each state represents a possible error model. The action space of the tool sequence includes all possible tool sequences, such as [air planer, high-power grinder], [angle grinder, fine sander belt], etc.

[0112] The state space of the workpiece error model includes different states of the workpiece error model; and based on the experience of casting part processing, a tool sequence table corresponding to each workpiece error model state is set as the action space of the tool sequence; a reward function is defined to evaluate the quality of each action. If the tool sequence can reduce the casting part error, then a positive reward is given; otherwise, a negative reward is given.

[0113] The reward function is used to obtain the tool sequence with the largest reward value in the action space of the corresponding tool sequence as the optimal tool sequence based on the current workpiece error model state.

[0114] The expression for the reward function is:

[0115] R(s,a,s')=-E(s')+γ0·E(s)-C(a)

[0116] In the formula: E(s) and E(s') represent the workpiece error value in the current state s and the workpiece error value in the new state s', respectively; γ0 is a discount factor between 0 and 1; C(a) represents the cost of taking action a; s represents the current state, i.e., the workpiece error model of the current casting; a represents the action taken, i.e., the selected tool sequence; s' represents the new state obtained after performing action a; R(s,a,s') represents the reward obtained by taking action a from the current state s and transitioning to the new state s'.

[0117] Specifically, γ0 is a discount factor between 0 and 1, representing the importance of the current state error to the reward relative to the new state error. If γ0 is close to 1, the model is more concerned with reducing the current error; if γ0 is close to 0, the model is more concerned with reducing future errors. C(a) is the cost of taking action a, which can include tool lifespan, time consumption, energy consumption, etc. We subtract this cost from the reward because we want the model to choose actions with lower costs. Assuming that in a certain state s, the error is 10, a tool sequence a is chosen with a cost of 2, the error of the new state s' after execution is 7, and γ0 = 0.8, then the reward is calculated as follows:

[0118] R(s,a,s')=-7+0.8×10-2=-1

[0119] This reward value indicates that taking the action resulted in a reduction in error after taking into account the cost of the action, and is therefore a positive result. By appropriately defining the reward function, reinforcement learning models can focus on minimizing error while taking into account the cost of the action, choosing an appropriate value of γ0, and accurately estimating C(a).

[0120] The state of the error model's state space and the action of the tool sequence's action space are used as inputs to the deep reinforcement learning algorithm network DQN. Based on the greedy algorithm, a new action is output according to the reward function, and a four-tuple data (state, action, reward function, next state) is established.

[0121] Create an experience replay storage pool, store the quadruple data (state, action, reward, nextstate) in the experience replay storage pool, and randomly extract quadruple data from the experience replay storage pool as a subset of training data;

[0122] Specifically, the ε-greedy algorithm is a strategy frequently used in reinforcement learning to achieve the exploration-exploitation trade-off. In reinforcement learning problems, the algorithm (agent) needs to find a balance between utilizing its current knowledge (making the best choice based on existing knowledge) and exploring new, unknown strategies or actions. The implementation of the ε-greedy strategy is as follows: In the ε-greedy algorithm, ε is a value between 0 and 1. At each time step, the agent randomly selects an action (exploration) with probability ε, and selects the action with the highest estimated Q-value with probability 1-ε. For example, assuming ε = 0.1, the agent will choose the action it believes will yield the maximum reward 90% of the time, and randomly select an action 10% of the time. The ε-greedy algorithm ensures that the agent does not completely ignore exploration, thus having the opportunity to discover better strategies. Over time, to better utilize known strategies, the ε value is usually gradually decreased, causing the agent to exploit more rather than explore.

[0123] In this embodiment, experience replay allows the Deep Reinforcement Learning (DQN) network to randomly sample from past experiences, thereby breaking the correlation between data and making training more stable. Simultaneously, to achieve a balance between exploration and exploitation, an ε-greedy strategy is used to select tool sequences. This means there is an ε probability of selecting a random working sequence, and a 1-ε probability of selecting a tool sequence corresponding to the maximum output value of the DQN network. As training progresses, ε gradually decreases, and the ε-greedy strategy selects a tool sequence action and executes it in a real-world environment. The deep reinforcement learning network includes a DQN environment, a task, and a reward function. The environment specifically refers to the error model and tools of the casting; the task specifically refers to selecting a suitable tool sequence to correct the casting error; the reward function specifically refers to the amount of error actually corrected after tool selection and the cost or efficiency of tool use. The DQN environment returns a new error model (new state) and a reward, calculated based on the actual error correction by the selected tool and the cost of tool use. The current state, the selected tool sequence, the obtained reward, and the new state are stored in the experience replay storage.

[0124] The Deep Reinforcement Learning Network (DQN) is trained using the training data subset to obtain predicted tool sequence action values. The network parameter weights of the DQN are then updated using the backpropagation method, with the mean squared error between the predicted tool sequence action values ​​and the target tool sequence action values ​​as the loss function.

[0125] The optimal network parameter weights are determined by finding the weights of the Deep Reinforcement Learning (DQN) network that minimize the mean squared error between the predicted tool sequence action values ​​and the target tool sequence action values.

[0126] The optimal deep reinforcement learning algorithm network DQN is obtained based on the optimal network parameter weights.

[0127] The target tool sequence action value Q target The formula for calculating (s,a) is:

[0128]

[0129] In the formula: reward represents the reward obtained after selecting a certain processing tool; s' represents the state; γ0 represents the discount factor; This represents the maximum value of the actions in the tool sequence list under the new state s';

[0130] The optimal deep reinforcement learning algorithm network (DQN) is used to obtain the sequence of machining tools required for the casting process. The DQN algorithm will be able to learn how to select the best tool sequence for a given state, thereby minimizing the error.

[0131] In a specific embodiment, the process parameter selection module is used to match the process parameters of the corresponding machining tool according to the data of the workpiece error model and the machining tool sequence.

[0132] Based on the workpiece error model data, machining tools, and existing machining process library information, multiple machining process parameter vectors are defined; and the tool process parameter database is obtained according to the machining process parameter vectors; the process parameter template needs to select a template that matches the input tool sequence from the process library, and then determine the corresponding process parameters. The Euclidean distance metric method is used for matching. Each template should have a clear structure, including tool type, tool error model size, and a set of corresponding process parameters (initial force, feed rate, tool entry angle, tool spindle speed, etc.).

[0133] A template for the processing parameter vector can be represented as:

[0134] M i =(T,E,F) initial V feed ,θ cut ,R spindle )

[0135] In the formula: T represents the tool type; E represents the error model size; Final represents the initial force; V feed Indicates the feed rate; θ cut Indicates the tool's entry angle; R spindle Indicates the spindle speed of the tool;

[0136] Define the desired vector of machining process parameters as S=(E',T',F') initial ,V' feed ,θ' cut ,R' spindle Based on the Euclidean distance calculation function, the machining process parameter vector with the smallest Euclidean distance to the expected vector of the machining process parameters is obtained from the tool process parameter database and used as the current machining process parameter vector to match the corresponding machining tool.

[0137] The Euclidean distance calculation function D(S,M) i The expression for ) is

[0138]

[0139]

[0140] In the formula: T' represents the expected instrument type; E' represents the expected error model size; F' intiial V' represents the desired initial force. feed Indicates the desired feed rate; θ' cut R' represents the desired angle of entry for the tool. spindle Indicates the desired tool spindle speed; Templates represent the parameter vectors in the tool process parameter database; M * This represents the machining parameter vector that has the smallest Euclidean distance to the expected vector of the machining parameters.

[0141] This also includes selecting the nearest neighbor algorithm with respect to D(S,M). i The nearest N neighbors (processing parameters) in Euclidean space are identified, and the average process parameter of the N neighbors (processing parameters) is obtained and added to the tool process parameter database to update the tool process parameter database; and the nearest neighbor algorithm is a known existing technology and is not the inventive point of this application, so it will not be described in detail here.

[0142] In a specific embodiment, the decision rule used by the process parameter selection module to determine process parameters based on the feedback signal or to call the tool scheduling module to evaluate the feedback signal is as follows:

[0143] The time-domain / frequency-domain data of the feedback signal is obtained through the process parameter selection module, and it is determined whether the time-domain / frequency-domain data is within a preset decision threshold.

[0144] If the time-domain / frequency-domain data is within the preset decision threshold range, then the current process parameters are confirmed to meet the standard, and the current process parameters are maintained.

[0145] If the time-domain / frequency-domain data exceeds the preset decision threshold range, it is confirmed that the current process parameters do not meet the standard, and the feedback signal is fed back to the tool scheduling module.

[0146] In a specific embodiment, the strategy employed by the tool scheduling module to evaluate the feedback signal and obtain the evaluation result based on the constructed risk assessment and decision-making model is as follows:

[0147] The process of building a risk assessment and decision-making model is specifically as follows:

[0148] Historical vibration signal data classified by risk labels are collected, and the FFT transformation algorithm is used to obtain the feature values ​​of the vibration signal data corresponding to the risk label classification; and a feature sample training set is obtained based on the feature values.

[0149] The characteristic values ​​include vibration amplitude and frequency components;

[0150] An optimization model for SVM based on radial basis function neural network (RBF) is constructed; the optimization model is as follows:

[0151]

[0152] subject to:

[0153]

[0154] K(x i ,x j )=exp(-γ||x i -x j || 2 )

[0155]

[0156] In the formula: N represents the number of samples in the feature sample training set; y i The sample label represents the feature sample training set, with a value of 1 or -1; α represents the Lagrange multiplier; C represents the regularization parameter; K(x) i ,x j ) represents the kernel function of the support vector machine; x i With x j The data points represent the vibration amplitude and frequency components, respectively; γ represents the kernel parameter of the radial basis function neural network (RBF); x s With y s All represent data points of the preset support vectors; b represents the bias.

[0157] The SMO optimization algorithm is used to solve for the Lagrange multipliers α and bias b of the optimization model; and the classification decision function is obtained based on the Lagrange multipliers α and bias b; the expression of the classification decision function is as follows:

[0158]

[0159] In the formula: f(x) represents the classification prediction result of vibration signal feature x;

[0160] The features obtained from the vibration analysis unit are vectors containing multiple frequency and time domain features of the signal; based on the classification decision function, vibration data features including the vibration amplitude and frequency components of the vibration signal are obtained according to the feedback signal output by the vibration analysis unit.

[0161] F(t) = [f1(t), f2(t), ..., f n (t)]

[0162] In the formula: F(t) represents the characteristic value of the vibration data; f i (t) represents the i-th vibration data feature at time t, where i = 1, 2, 3, ..., n;

[0163] The trained optimization model is used to obtain the risk label corresponding to the current vibration data features, and the current vibration data features are compared with the vibration data features f corresponding to the tools stored in the set process library. lib,i If the Euclidean distance of the vibration feature is less than or equal to a preset vibration threshold, then it is confirmed that the feedback signal does not pose a risk, and the current tool sequence continues to be used.

[0164] If the Euclidean distance of the vibration feature is greater than the preset vibration threshold, it is confirmed that the feedback signal is at risk, and the point cloud processing module is called again to update the workpiece error.

[0165] The formula for calculating the vibration characteristic Euclidean distance D(t) is as follows:

[0166]

[0167] In the formula: f i (t) represents the i-th vibration data feature at time t; f lib,i This indicates the vibration data characteristics corresponding to the tools stored in the set process library.

[0168] In this embodiment, a point cloud processing module performs 3D point cloud imaging processing on the visual signals transmitted by the 3D binocular camera. The imaging is then compared with a standard model to generate a workpiece error model, which serves as the processing target. When the processing target is less than a set threshold, the cutting and grinding process is completed. If the processing target does not meet the completion conditions, it is sent to the tool scheduling module. The tool scheduling module generates a tool sequence based on the error model. This tool sequence contains one or more tools that may be needed, including a planer, a high-powered grinder, an angle grinder, a high-powered belt sander, a fine-powered belt sander, and a telescopic file. The first tool is selected from the tool sequence, and its information is transmitted to the process parameter selection module. The process parameter selection module simultaneously receives the tool information and the workpiece error model, using the workpiece error model as a reference to match suitable process parameters. These process parameters include, but are not limited to, initial force, feed rate, tool entry angle, and tool spindle speed. The process information is transmitted to the subsequent motion planning module. The motion planning module also receives the workpiece error model and tool sequence information. Combining the tool sequence information, it provides the tool path and optimizes the trajectory. For example, when a cutting tool is identified, a path fitting method for the cutting process is provided to fit the point cloud information. Then, trajectory optimization is performed using the minimization of the robotic arm joint movement amplitude and the reduction of the distance between the robotic arm end-points as optimization conditions. This method or means of trajectory optimization using the minimization of the robotic arm joint movement amplitude and the reduction of the distance between the robotic arm end-points as optimization conditions is a known existing technology and is not the inventive point of this application, so it will not be elaborated here. Then, the robotic arm control module sends the robotic arm drive information to the robotic arm drive module according to the processing trajectory. The robotic arm drive module is responsible for communicating with the PLC robot control cabinet, transmitting the given generated information to the PLC via the Modbus TCP protocol. The PLC drives each robotic arm joint of the robot to complete the operation. The robotic arm end-point force control unit collects the vibration data of the robot's end-point tool during operation through a six-dimensional force control sensor and transmits the data to the subsequent vibration analysis unit. The vibration analysis unit receives vibration signals from the end effector of the robotic arm during a specific process and performs statistical analysis on the signals in both the time and frequency domains. When a significant difference occurs between the vibration signal and the workpiece model, indicating an abnormal vibration, a stop command is sent to the robotic arm drive module, and the abnormal information is fed back to the process parameter selection module. If the vibration signal is within the preset normal range of the workpiece model, the information indicating the completion of the machining process is fed back to the process parameter selection module. Upon receiving the feedback, the process parameter selection module determines whether the time and frequency domain parameters are within thresholds. If they are, the process parameters are considered good, and operation continues with those parameters. If the threshold is reached, the process parameters need to be adjusted. If the parameters are too low or too high, it may indicate that the workpiece has completed the cutting and grinding process or that the tool is no longer suitable for use; in this case, the process parameter selection module feeds back the information to the tool scheduling module.The tool scheduling module receives feedback from the process parameter selection module and sends the feedback information to the risk assessment and decision-making model. If the risk assessment is within the normal range, the next tool is called according to the tool sequence. If the risk assessment value is too high, the subsequent steps are abandoned, and the feedback information is sent to the point cloud processing module to regenerate the point cloud image and train the model. The constructed risk assessment and decision-making model has self-learning capabilities, allowing it to readjust the risk assessment value in the next risk assessment. Based on the re-acquired workpiece error model, the subsequent tool sequence is regenerated based on the risk assessment and decision-making model, and the forward process continues. The tool scheduling module and the process selection module can self-learn from the feedback, enhancing their ability to schedule tools and process parameters, improving model matching probability, and further reducing the use of the 3D stereo camera.

[0169] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A smart tool scheduling and control system for a casting cleaning robot, characterized in that, It includes a point cloud processing module, a tool scheduling module, a process parameter selection module, a motion planning module, and a robotic arm control module; The point cloud processing module is used to obtain the workpiece error model of the casting workpiece, and transmit the data of the workpiece error model to the process parameter selection module, motion planning module and tool scheduling module respectively. The tool scheduling module is used to obtain the sequence of machining tools required for the casting process based on the data of the workpiece error model; and transmit the sequence of machining tools to the process parameter selection module. The processing tool sequence includes one or more tool combinations such as an air planer, a high-power grinder, an angle grinder, a high-power belt sander, a fine belt sander, and a telescopic file. The process parameter selection module is used to match the process parameters of the corresponding machining tool with the data of the workpiece error model and the machining tool sequence; and transmit the process parameters to the motion planning module. The process parameters include initial tool force, tool feed rate, tool entry angle, and tool spindle speed; The motion planning module is used to obtain the machining trajectory path based on the workpiece error model data, machining tool sequence, and process parameters; and transmit the machining trajectory path to the robotic arm control module. The robotic arm control module includes a robotic arm drive module, a robotic arm end force control unit, and a vibration analysis unit. The robotic arm drive module is used to drive the robotic arm to move the machining tool according to the machining trajectory path. The end effector force control unit of the robotic arm is used to collect vibration data of the end tool during the operation of the robotic arm through a six-dimensional force control sensor, and transmit the vibration data to the vibration analysis unit; The vibration analysis unit is used to generate feedback signals from the vibration data and transmit them to the robotic arm drive module and the process parameter selection module. The robotic arm drive module controls the start and stop of the robotic arm according to the feedback signal; The process parameter selection module is used to decide on process parameters based on the feedback signal or to call the tool scheduling module to evaluate the feedback signal; the tool scheduling module evaluates the feedback signal based on the constructed risk assessment and decision model to obtain the evaluation result, and calls the tool of the processing tool sequence or calls the point cloud processing module to re-obtain the workpiece error based on the evaluation result.

2. The intelligent tool scheduling and control system for a casting cleaning robot according to claim 1, characterized in that, The specific strategy for acquiring the data of the workpiece error model of the cast workpiece is as follows: The casting workpiece is scanned using a 3D binocular camera to obtain point cloud data, which is then transmitted to a point cloud processing module. The point cloud processing module generates a model of the casting workpiece to be processed based on the point cloud data. The model of the casting workpiece to be processed is divided into regions according to the processing technology of the casting workpiece, and multiple model regions are obtained. The spatial offset of each model region is obtained based on the model region and the set standard model, and the data of the workpiece error model of the casting workpiece is obtained based on the spatial offset.

3. The intelligent tool scheduling and control system for a casting cleaning robot according to claim 1, characterized in that, The tool scheduling module uses a strategy to obtain the sequence of machining tools required for the casting process based on the workpiece error model data. The error feature data of the workpiece error model is extracted, and the error feature data includes the maximum error, the average error, and the error distribution; and the error feature data is combined into a state vector as the current state of the workpiece error model. Based on the deep reinforcement learning algorithm network DQN, the state space of the workpiece error model, the action space of the tool sequence, and the reward function are defined according to the state of the workpiece error model. The state space of the workpiece error model includes different states of the workpiece error model; and based on the experience of casting part processing, a tool sequence table corresponding to each workpiece error model state is set as the action space of the tool sequence. The reward function is used to obtain the tool sequence with the largest reward value in the action space of the corresponding tool sequence as the optimal tool sequence based on the current workpiece error model state. The expression for the reward function is: R(s,a,s')=-E(s')+γ0·E(s)-C(a) In the formula: E(s) and E(s') represent the workpiece error value in the current state s and the workpiece error value in the new state s', respectively; γ0 is a discount factor between 0 and 1; C(a) represents the cost of taking action a; s represents the current state, i.e., the workpiece error model of the current casting; a represents the action taken, i.e., the selected tool sequence; s' represents the new state obtained after performing action a; R(s,a,s') represents the reward obtained by taking action a from the current state s and transitioning to the new state s'. The state of the workpiece error model and the action of the tool sequence in the action space are used as inputs to the deep reinforcement learning algorithm network DQN. Based on the greedy algorithm, new actions are output according to the reward function and a four-tuple data (state, action, reward, next state) is established. Create an experience replay storage pool, store the quadruple data (state, action, reward, next state) in the experience replay storage pool, and randomly extract quadruple data from the experience replay storage pool as a subset of training data; The Deep Reinforcement Learning Network (DQN) is trained using the training data subset to obtain predicted tool sequence action values. The network parameter weights of the DQN are then updated using the backpropagation method, with the mean squared error between the predicted tool sequence action values ​​and the target tool sequence action values ​​as the loss function. The optimal network parameter weights are determined by finding the weights of the Deep Reinforcement Learning (DQN) network that minimize the mean squared error between the predicted tool sequence action values ​​and the target tool sequence action values. The optimal deep reinforcement learning algorithm network DQN is obtained based on the optimal network parameter weights. The target tool sequence action value Q target The formula for calculating (s,a) is: In the formula: reward represents the reward obtained after selecting a certain processing tool; s' represents the state; γ0 represents the discount factor; This represents the maximum value of the actions in the tool sequence list under the new state s'; The sequence of machining tools required for the casting process is obtained based on the optimal deep reinforcement learning algorithm network DQN.

4. The intelligent tool scheduling and control system for a casting cleaning robot according to claim 1, characterized in that, The process parameter selection module is used to match the process parameters of the corresponding machining tool based on the data of the workpiece error model and the machining tool sequence. Multiple machining process parameter vectors are defined based on workpiece error model data, machining tools, and existing machining process library information; and The tool process parameter database is obtained based on the aforementioned processing process parameter vector; The expression for the processing parameter vector is as follows: M i =(T,E,F initial ,V feed ,θ cut ,R spindle ) In the formula: T represents the tool type; E represents the workpiece error model size; F initial V represents the initial force; feed Indicates the feed rate; θ cut Indicates the tool's entry angle; R spindle Indicates the spindle speed of the tool; Define the desired vector of machining process parameters as S=(E',T',F') initial ,V' feed ,θ' cut ,R' spindle Based on the Euclidean distance calculation function, the machining process parameter vector with the smallest Euclidean distance to the expected vector of the machining process parameters is obtained from the tool process parameter database and used as the current machining process parameter vector to match the corresponding machining tool. The Euclidean distance calculation function D(S,M) i The expression for ) is In the formula: T' represents the expected instrument type; E' represents the expected error model size; F' initial Indicates the expected initial force; V' feed Indicates the desired feed rate; θ' cut R' represents the desired angle of entry for the tool. spindle Indicates the desired tool spindle speed; Templates represent the parameter vectors in the tool process parameter database; M * This represents the machining parameter vector that has the smallest Euclidean distance to the expected vector of the machining parameters.

5. The intelligent tool scheduling and control system for a casting cleaning robot according to claim 1, characterized in that, The process parameter selection module is used to determine process parameters based on the feedback signal or to call the tool scheduling module to evaluate the feedback signal. The decision rules are as follows: The time-domain / frequency-domain data of the feedback signal is obtained through the process parameter selection module, and it is determined whether the time-domain / frequency-domain data is within a preset decision threshold. If the time-domain / frequency-domain data is within the preset decision threshold range, then the current process parameters are confirmed to meet the standard, and the current process parameters are maintained. If the time-domain / frequency-domain data exceeds the preset decision threshold range, it is confirmed that the current process parameters do not meet the standard, and the feedback signal is fed back to the tool scheduling module.

6. The intelligent tool scheduling and control system for a casting cleaning robot according to claim 1, characterized in that, The tool scheduling module evaluates the feedback signals based on the constructed risk assessment and decision-making model to obtain the evaluation results. The specific process of constructing a risk assessment and decision-making model is as follows: Historical vibration signal data classified by risk labels are collected, and the FFT transformation algorithm is used to obtain the feature values ​​of the vibration signal data corresponding to the risk label classification; and a feature sample training set is obtained based on the feature values. The characteristic values ​​include vibration amplitude and frequency components; An optimization model for SVM based on radial basis function neural network (RBF) is constructed; the optimization model is as follows: subject to: K(x i ,x j )=exp(-γ||x i -x j || 2 ) In the formula: N represents the number of samples in the feature sample training set; y i The sample label represents the feature sample training set, with a value of 1 or -1; α represents the Lagrange multiplier; C represents the regularization parameter; K(x) i ,x j ) represents the kernel function of the support vector machine; x i With x j The data points represent the vibration amplitude and frequency components, respectively; γ represents the kernel parameter of the radial basis function neural network (RBF); x s With y s All represent data points of the preset support vectors; b represents the bias. The SMO optimization algorithm is used to solve for the Lagrange multipliers α and bias b of the optimization model; and the classification decision function is obtained based on the Lagrange multipliers α and bias b; the expression of the classification decision function is as follows: In the formula: f(x) represents the classification prediction result of vibration signal feature x; Based on the classification decision function, vibration data features, including the vibration amplitude and frequency components of the vibration signal, are obtained according to the feedback signal output by the vibration analysis unit. F(t)=[f1(t),f2(t),...,f n (t)] In the formula: F(t) represents the characteristic value of the vibration data; f i (t) represents the i-th vibration data feature at time t, where i = 1, 2, 3, ..., n; The risk label corresponding to the current vibration data feature is obtained through the trained optimization model, and the current vibration data feature is compared with the vibration data feature f corresponding to the tool stored in the set process library. lib,i If the Euclidean distance of the vibration feature is less than or equal to a preset vibration threshold, then it is confirmed that the feedback signal does not pose a risk, and the current tool sequence continues to be used. If the Euclidean distance of the vibration feature is greater than the preset vibration threshold, it is confirmed that the feedback signal is at risk, and the point cloud processing module is called again to update the workpiece error. The formula for calculating the vibration characteristic Euclidean distance D(t) is as follows: In the formula: f i (t) represents the i-th vibration data feature at time t; f lib,i This indicates the vibration data characteristics corresponding to the tools stored in the set process library.

Images