Intelligent tool changing decision-making method based on tool wear perception

Abstract

The invention relates to the technical field of machining automation control, and discloses an intelligent tool changing decision-making method based on tool wear perception, which comprises the following steps: acquiring real-time state data of a tool through vibration, acoustic emission, force and temperature sensors, and fusing features of a multi-modal graph neural network to obtain tool wear feature data. And inputting the parameters into a Bayesian decision network, optimizing a tool changing strategy by using a dynamic probabilistic reasoning structure, and generating decision optimization parameters. And a multi-target tool changing optimization model with the highest machining efficiency and the longest service life of the tool as targets is constructed, and an optimal tool changing strategy is determined by adopting an improved particle swarm algorithm. Based on this, a hierarchical decision control model is established and comprises a global evaluation layer, a dynamic adjustment layer and an execution control layer, and intelligent control of tool changing action is realized. In addition, a self-healing control module is embedded in the system to deal with abnormal wear of the cutter. The machining efficiency is improved, the service life of the cutter is prolonged, the machining quality is guaranteed, and intelligent development of machining is promoted.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic control of mechanical processing, in particular to an intelligent tool changing decision method based on tool wear perception. BACKGROUND

[0002] In the field of mechanical processing, as a key component, the tool wear directly affects the processing quality, production efficiency and production cost. In the traditional processing process, the tool changing decision mainly depends on the experience judgment of the operator or the fixed processing time setting. However, this method has many disadvantages.

[0003] The tool changing decision based on the experience of the operator is highly subjective and lacks scientific basis. Different operators have different standards for judging the degree of tool wear, which may lead to untimely tool changing or early tool changing. Untimely tool changing may cause excessive tool wear, which not only affects the processing precision, reduces the surface quality of the workpiece, increases the scrap rate, but also may damage the tool and the machine tool, increases the maintenance cost and downtime; while early tool changing may cause waste of tool resources, increase the production cost and reduce the production efficiency. For example, in the processing of precision parts, the dimensional deviation caused by tool wear may make the precision of the parts unable to meet the design requirements, resulting in the whole batch of parts being scrapped.

[0004] Setting the tool changing time according to the fixed processing time also has obvious shortcomings. During processing, the tool wear is affected by multiple factors, including cutting parameters (such as cutting speed, feed rate, cutting depth), workpiece material properties (hardness, toughness, etc.), cutting fluid usage, etc. Under different processing conditions, the tool wear rate is quite different. When processing high-hardness alloy materials, the tool wear rate is much faster than that of ordinary steel; improper selection of cutting parameters, such as too high cutting speed, will exacerbate tool wear. Therefore, the fixed tool changing time setting cannot adapt to the complex and variable processing conditions, and it is difficult to achieve reasonable utilization of tools.

[0005] With the development of manufacturing industry towards intelligentization and automation, the demand for precise control and optimized management of the processing process is increasingly urgent. In order to improve the processing efficiency, reduce the cost and ensure the stability of product quality, a technology is needed that can monitor the tool wear state in real time and make scientific and reasonable tool changing decision based on the wear condition. At present, although there are some researches on tool wear monitoring, most of them only stay at the monitoring level and fail to effectively convert the monitoring data into practical tool changing decision scheme. There are still technical bottlenecks in the accuracy of tool wear prediction, optimization of tool changing decision and collaborative control with the processing system, which need to be broken through. SUMMARY

[0006] The purpose of the present application is to provide an intelligent tool changing decision method based on tool wear perception to solve the problems raised in the background.

[0007] To achieve the above object, the present application provides the following technical solution: an intelligent tool changing decision method based on tool wear perception, the method comprising: Collecting real-time state data of the machining tool through a multi-source sensor, the multi-source sensor comprising a vibration sensor, an acoustic emission sensor, a force sensor, and a temperature sensor; performing feature fusion processing on the real-time state data based on a multi-modal graph neural network to obtain tool wear feature data; inputting the tool wear feature data into a pre-trained Bayesian decision network, the Bayesian decision network adopting a dynamic probability inference structure, iteratively optimizing a tool changing strategy based on a conditional probability distribution, and generating decision optimization parameters; Constructing a multi-objective tool changing optimization model according to the decision optimization parameters, the multi-objective tool changing optimization model taking the highest machining efficiency and the longest tool life as optimization objectives, and globally optimizing a tool changing time using an improved particle swarm algorithm, wherein the improved particle swarm algorithm introduces an adaptive inertia weight factor and a dynamic learning factor; outputting optimal tool changing strategy data based on the multi-objective tool changing optimization model; Establishing a hierarchical decision control model according to the optimal tool changing strategy data, the hierarchical decision control model comprising a global evaluation layer, a dynamic adjustment layer, and an execution control layer, wherein the global evaluation layer performs global tool changing planning based on the decision optimization parameters, the dynamic adjustment layer performs local parameter optimization based on the optimal tool changing strategy data, and the execution control layer implements time sequence scheduling of tool changing actions based on a fuzzy logic control algorithm; outputting tool changing control instructions through the hierarchical decision control model to realize intelligent tool changing decision of a machining system.

[0008] Preferably, inputting the tool wear feature data into a pre-trained Bayesian decision network, the Bayesian decision network adopting a dynamic probability inference structure, iteratively optimizing a tool changing strategy based on a conditional probability distribution, and generating decision optimization parameters comprises: Obtaining real-time state data, the real-time state data comprising a current tool wear amount, a cutting force amplitude, a vibration frequency spectrum peak value, a temperature change gradient, and a workpiece material parameter; constructing a multi-dimensional state space based on the real-time state data, and constructing an action space based on an executable tool changing delay time, a cutting parameter adjustment amount, and a tool compensation amount; Constructing a multi-objective joint probability model based on the state space and the action space, the multi-objective joint probability model comprising a wear prediction term, an efficiency evaluation term, and an energy consumption evaluation term, wherein the wear prediction term is calculated through a Bayesian posterior probability of tool wear rate and remaining life, the efficiency evaluation term is calculated through a joint distribution of unit time processing amount and tool changing frequency, and the energy consumption evaluation term is calculated through Markov chain transition probability of cutting power and tool changing energy consumption; A dynamic probabilistic inference structure is constructed, which includes a prior network and a posterior network, both of which comprise an input layer, a graph convolution layer, and an output layer, the dimension of the input layer matches the dimension of the state space, the graph convolution layer aggregates multi-modal features using an attention mechanism, and the output layer maps the action space distribution through a probability density function; The prior network and the posterior network are trained using a Monte Carlo sampling method, training samples are generated through an importance sampling strategy and stored in an experience pool, and samples are randomly selected from the experience pool for network parameter updating, wherein the prior network optimizes the probability distribution through variational inference, and the posterior network updates the conditional probability through Gibbs sampling; based on the trained network output decision optimization parameters, the parameters include network topology weight, state normalization matrix, and action distribution parameter.

[0009] Preferably, a multi-objective tool changing optimization model is constructed according to the decision optimization parameters, the multi-objective tool changing optimization model takes the highest processing efficiency and the longest tool life as the optimization target, and an improved particle swarm algorithm is used to globally optimize the tool changing time, and the optimal tool changing strategy data is output based on the multi-objective tool changing optimization model, including: A multi-objective function for tool changing optimization is constructed, which includes an efficiency optimization objective function and a life optimization objective function, wherein the efficiency optimization objective function is calculated by the ratio of the processing cycle to the tool changing time, and the life optimization objective function is calculated by the weighted sum of the cumulative amount of tool wear and the reciprocal of the remaining life; Based on the multi-objective function, a tool changing optimization constraint condition is constructed, which includes a maximum allowed wear amount, a cutting parameter threshold, a tool changing time window, and a processing continuity requirement; Time series encoding is used to discretize the tool changing strategy, each time node contains a wear prediction value and a cutting parameter adjustment amount, and the particle position and velocity are updated based on the particle fitness and the global optimal solution of the population; wherein the particle fitness is calculated by the weighted sum of the objective function value and the constraint violation degree; An adaptive inertia weight factor is introduced, which decreases linearly in segments with the increase of iteration number, and the global search and local development capabilities of the particle swarm are dynamically adjusted through the factor; A dynamic learning factor is introduced, which adaptively adjusts the weights of social learning and individual learning according to the particle distribution density, and optimizes the convergence trajectory of the particle swarm through the factor; based on the adaptive inertia weight factor and the dynamic learning factor, iterative optimization is performed, and the strategies generated by each iteration are screened through the Pareto front, and the non-dominated solutions are stored in the candidate strategy set; An optimal solution satisfying the efficiency-life trade-off is selected from the candidate strategy set as the tool change strategy, and the optimal strategy is subjected to time series interpolation smoothing processing to generate a tool change time series and corresponding cutting parameter adjustment instructions.

[0010] Preferably, the global evaluation layer performs global tool change planning based on the decision optimization parameters, including: The tool wear evolution trend is described using a non-uniform B-spline curve, and the trend is represented as a function of a time parameter covering the entire machining period. The tool vibration signal is subjected to multi-resolution analysis based on wavelet transform to extract energy distribution characteristics in different frequency bands, and a wear state evaluation index is constructed through feature fusion. Global optimization constraints are constructed, including the maximum number of machining interruptions, the minimum tool utilization rate, and the workpiece surface roughness limit. A multi-objective decision function is constructed, including a tool change cost term, a machining efficiency term, and a tool life term, and the constraint terms are coupled through Lagrange multipliers. The machining period is discretized into multiple time windows, and the decision function is subjected to piecewise linearization processing to construct a discretized optimization model. The discretized optimization model is solved using a branch and bound algorithm, and the invalid search space is reduced through pruning strategies to generate a global tool change planning sequence.

[0011] Preferably, the dynamic adjustment layer performs local parameter optimization based on the optimal tool change strategy data, including: A dynamic adjustment window is constructed based on the current tool wear, machining load, and remaining life, and the window size is scaled in real time through a load adaptive coefficient. An online learning mechanism is used to update the local environment model, and an incremental support vector machine is used to perform regression analysis on real-time sensor data to predict short-term wear trends. A local optimization cost function is constructed, including a tool change delay penalty term, a cutting parameter deviation term, and a machining quality fluctuation term, and the contribution proportion of each penalty term is adjusted through a dynamic weight matrix. The optimized parameters are subjected to feasibility verification, and if the verification fails, a backtracking mechanism is triggered to generate adjustment instructions again.

[0012] Preferably, the execution control layer implements time series scheduling of tool change actions based on a fuzzy logic control algorithm, including: A multi-degree-of-freedom kinematic model of the tool change mechanism is established, including translational axis velocity, rotational axis angle, and clamping force dynamic equations. The kinematic model is converted into a fuzzy rule base, and the input variables of the rule base include tool change time deviation, tool position error, and clamping force fluctuation, and the output variables are the control amounts of each axis. The input and output variables are fuzzified by using a Gaussian membership function, and precise output is realized by a weighted average method; A time sequence constraint condition is constructed, and the constraint condition comprises a maximum tool changing delay, a mechanical arm movement range and a torque limit; A tool changing action sequence is generated based on the fuzzy rule base and the constraint condition, and each actuator action is coordinated through a time stamp synchronization mechanism.

[0013] Preferably, the intelligent tool changing decision method based on tool wear perception further comprises: A self-recovery control module is embedded in the machining system, and when abnormal wear is detected, a compensation strategy is generated based on reinforcement learning to dynamically adjust the cutting parameters or switch to a backup tool.

[0014] Preferably, the present application further comprises an intelligent tool changing decision system based on tool wear perception for realizing the intelligent tool changing decision method based on tool wear perception, comprising: A data acquisition unit is used to acquire real-time state data of the machining tool through a multi-source sensor, and to perform feature fusion based on a multi-modal graph neural network; A decision optimization unit is used to generate tool changing strategy optimization parameters through a Bayesian decision network, and to perform global optimization using an improved particle swarm algorithm; A hierarchical control unit comprises a global evaluation module, a dynamic adjustment module and an execution control module, which are respectively used for global planning, local optimization and action scheduling; A tool changing execution unit is used to drive the tool changing mechanism to complete tool replacement and parameter adjustment according to the control instructions.

[0015] Compared with the prior art, the present application has the following advantages: The present application can accurately obtain tool wear characteristics by real-time acquisition of tool state data through a multi-source sensor and feature fusion using a multi-modal graph neural network. Based on this, the Bayesian decision network uses a dynamic probability reasoning structure to iteratively optimize the tool changing strategy in combination with various factors in the machining process, and the generated decision optimization parameters provide strong support for accurate judgment of the tool changing time. Taking machining of complex parts as an example, the traditional tool changing method may cause machining interruption and tool damage due to untimely tool changing, while the intelligent tool changing decision method of the present application can timely issue a tool changing instruction before the tool is excessively worn, avoiding machining stagnation caused by tool failure, greatly shortening the machining cycle and improving the machining efficiency.

[0016] The multi-objective tool changing optimization model takes the highest machining efficiency and the longest tool life as the optimization objectives, and globally optimizes the tool changing time by using an improved particle swarm algorithm. The adaptive inertia weight factor and the dynamic learning factor introduced in the algorithm enable the algorithm to quickly find the optimal solution in a complex search space. By reasonably planning the tool changing time, the excessive wear of the tool is avoided, and the service life of the tool is fully utilized. In batch production, the number of tool changes can be reduced, and the tool procurement cost can be reduced.

[0017] The global evaluation layer in the hierarchical decision control model globally plans the tool changing based on the decision optimization parameters, can grasp the whole machining process, and formulates a reasonable tool changing plan according to the tool wear evolution trend and the machining requirements; the dynamic adjustment layer locally optimizes the parameters based on the real-time tool wear, machining load and remaining life, and ensures that the machining process is always in the best state; and the execution control layer realizes the timing scheduling of the tool changing action based on the fuzzy logic control algorithm, and ensures the smooth and accurate tool changing process. A series of measures effectively avoid the decline of machining precision caused by tool wear, ensure the machining quality of the workpiece, reduce the scrap rate, and improve the market competitiveness of the product.

[0018] The self-healing control module is embedded in the machining system, when abnormal wear is detected, a compensation strategy is generated based on reinforcement learning, and the cutting parameters are dynamically adjusted or the standby tool is switched. This enables the machining system to self-repair and adjust when facing unexpected situations, enhances the adaptability and stability of the system, reduces the downtime caused by tool failure, and improves the continuity of production. For a long-time running production line, the maintenance cost can be effectively reduced, and the production benefit can be improved.

[0019] The present application deeply integrates sensor technology, artificial intelligence algorithm and machining system, realizes the whole-process intelligentization from tool state monitoring to tool changing decision, breaks away from the dependence on artificial experience, and improves the scientificity and accuracy of the decision. Meanwhile, the technology is easy to integrate into existing machining equipment, has good popularization and application prospect, and is helpful to promote the development of the whole mechanical machining industry towards intelligentization and automation. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 A working principle diagram of the intelligent tool changing decision method described in the present application; Figure 2 A flowchart for generating decision optimization parameters for a Bayesian decision network; Figure 3 A flowchart for constructing and solving a multi-objective tool changing optimization model Figure 4 A flowchart for the architecture of an intelligent tool changing decision system. DETAILED DESCRIPTION

[0021] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0022] Please refer to Figures 1-4 The present application provides an intelligent tool changing decision method based on tool wear perception, and the overall implementation scheme is as follows: Real-time state data of the machining tool is collected by using multi-source sensors such as vibration sensors, acoustic emission sensors, force sensors and temperature sensors. These sensors obtain information of the tool in the machining process from different dimensions, such as vibration sensors monitoring tool vibration, acoustic emission sensors capturing acoustic signals generated by tool wear, force sensors measuring cutting force, and temperature sensors monitoring tool temperature changes. Then, the collected real-time state data is processed by multi-modal graph neural network for feature fusion, and multi-source heterogeneous data is fused into tool wear feature data for subsequent analysis.

[0023] The tool wear feature data obtained by feature fusion is input into a pre-trained Bayesian decision network. The network adopts a dynamic probability inference structure, and iteratively optimizes the tool changing strategy based on conditional probability distribution. In this process, various factors are considered comprehensively, and finally decision optimization parameters are generated. These parameters provide key basis for subsequent tool changing strategy formulation.

[0024] According to the generated decision optimization parameters, a multi-objective tool changing optimization model is constructed with the optimization objectives of the highest machining efficiency and the longest tool life. In order to realize global optimization of the tool changing time, an improved particle swarm algorithm is used. The algorithm introduces adaptive inertia weight factor and dynamic learning factor to overcome the shortcomings of traditional particle swarm algorithm and improve search efficiency and accuracy. Through solving the model, the optimal tool changing strategy data is output.

[0025] According to the optimal tool changing strategy data, a hierarchical decision control model is established, which includes a global evaluation layer, a dynamic adjustment layer and an execution control layer. The global evaluation layer performs global tool changing planning based on the decision optimization parameters; the dynamic adjustment layer performs local parameter optimization based on the optimal tool changing strategy data; and the execution control layer realizes timing scheduling of tool changing action based on fuzzy logic control algorithm. Finally, the tool changing control instruction is output through the hierarchical decision control model, the tool changing mechanism is driven, and the intelligent tool changing decision of the machining system is realized.

[0026] The implementation of the present application will be further described below in conjunction with Examples 1 to 5. Examples

[0027] This embodiment is used to illustrate how the Bayesian decision network builds relevant models based on real-time state data and generates decision optimization parameters through training, providing core basis for subsequent tool changing strategy formulation. The specific method includes: Constructing state space and action space: Obtain real-time state data such as current tool wear, cutting force amplitude, vibration spectrum peak value, temperature change gradient, and processing workpiece material parameters. Based on these data, a multi-dimensional state space is constructed, where each dimension represents a state information. At the same time, based on the executable tool changing delay time, cutting parameter adjustment amount and tool compensation amount, the action space is constructed, which are the operations that can be taken in the tool changing decision process.

[0028] Constructing multi-objective joint probability model: Based on the constructed state space and action space, a multi-objective joint probability model is established. This model includes wear prediction term, efficiency evaluation term and energy consumption evaluation term.

[0029] Wear prediction term: calculated through the Bayesian posterior probability of tool wear rate and remaining life. Let the tool wear rate be , the remaining life be , and the Bayesian posterior probability formula be , where is the likelihood probability of observation data given the wear rate and remaining life, is the prior probability of wear rate and remaining life, is the probability of observation data.

[0030] Efficiency evaluation term: calculated through the joint distribution of unit time processing amount and tool changing frequency , the formula is , which reflects the relationship between processing efficiency and tool changing frequency.

[0031] Energy consumption evaluation term: calculated through the Markov chain transition probability of cutting power and tool changing energy . Let the Markov chain transition probability be , where represents the time step, and this probability describes the transition relationship between cutting power and tool changing energy between different time steps.

[0032] Constructing dynamic probability inference structure: constructing a dynamic probability inference structure containing prior network and posterior network. Both prior network and posterior network contain input layer, graph convolution layer and output layer. The dimension of input layer matches the dimension of state space, which is used to receive state data. The graph convolution layer adopts attention mechanism to aggregate multi-modal features, the formula is , where is the output feature of node , is a set of neighbor nodes of the node , is an attention coefficient, is a weight matrix, is an input feature of the node . The output layer maps the action space distribution through a probability density function.

[0033] Network training and decision optimization parameter output: the prior network and the posterior network are trained by using the Monte Carlo sampling method. Training samples are generated by using the importance sampling strategy and stored in the experience pool. Samples are randomly extracted from the experience pool for network parameter update. The prior network optimizes the probability distribution by using the variational inference. The variational inference formula is wherein is a variational distribution, is a posterior distribution. The posterior network updates the conditional probability by using the Gibbs sampling. After the training is completed, the decision optimization parameters, including the network topology weight, the state normalization matrix and the action distribution parameter, are output based on the trained network. Embodiment

[0034] This embodiment details how to build a multi-objective tool changing optimization model and solve it by using an improved particle swarm algorithm, so as to obtain optimal tool changing strategy data and realize the balanced optimization of machining efficiency and tool life. The specific method includes: Building a multi-objective function for tool changing optimization, including an efficiency optimization objective function and a life optimization objective function.

[0035] Efficiency optimization objective function: calculated by the ratio of the machining cycle to the tool changing time , the formula is The larger the function value is, the higher the machining efficiency is.

[0036] Life optimization objective function: calculated by the reciprocal weighted sum of the cumulative amount of tool wear and the remaining life , the formula is wherein and are weight coefficients, used to adjust the relative importance of the two factors in the life optimization objective function.

[0037] Building constraint conditions: based on the multi-objective function, the tool changing optimization constraint conditions are built, including the maximum allowed wear amount , the cutting parameter threshold (such as the cutting speed threshold , the feed amount threshold , etc.), the tool changing time window and the requirement of machining continuity. These constraint conditions ensure that the tool changing strategy is feasible and reasonable in actual machining.

[0038] The time series coding is used to discretize the tool changing strategy, and each time node contains the wear prediction value and the cutting parameter adjustment value. The particle position and velocity are updated based on the particle fitness and the optimal solution of the group. The particle fitness is calculated by the weighted sum of the objective function value and the constraint violation degree, and the formula is is the weight coefficient, is the constraint violation degree.

[0039] The adaptive inertia weight factor is introduced, which increases in a piecewise linear decreasing manner with the iteration number , and the formula is and are the maximum and minimum values of the inertia weight, is the maximum iteration number. Through this factor, the global search and local development capabilities of the particle swarm are dynamically adjusted. In the early stage of iteration, the global search is strengthened, and in the later stage, the local development is focused, improving the convergence speed and accuracy of the algorithm.

[0040] The dynamic learning factor is introduced to adaptively adjust the weights of social learning and individual learning according to the particle distribution density. Let the particle distribution density be , and the adjustment formula of the dynamic learning factor and is are the maximum and minimum values of the learning factor, and are the maximum and minimum values of the particle distribution density. Through this factor, the convergence trajectory of the particle swarm is optimized, avoiding the algorithm falling into local optimum.

[0041] The maximum value of the particle distribution density : refers to the highest distribution density value that the particle swarm may reach in the search space of the particle swarm algorithm. It represents the most dense state that the particle swarm may appear in the search process, which is usually pre-set according to the search space range and the number of particles of the optimization problem. For example, in the tool changing time optimization problem, if the highest degree of particle aggregation in the search space is 0.8 (assuming the density value range is 0-1), then can be set to 0.8.

[0042] The minimum value of the particle distribution density ​​​​​​​​​This refers to the lowest possible particle density in the search space of the particle swarm optimization algorithm. It reflects the sparsest state the particle swarm might exhibit during the search process and needs to be determined in conjunction with the characteristics of the specific optimization problem. For example, if the density of particles in the search space is 0.2, then... It can be set to 0.2.

[0043] Iterative optimization is performed based on adaptive inertia weighting factors and dynamic learning factors. The Pareto front is used to screen the strategies generated in each iteration, and non-dominated solutions are stored in a candidate strategy set. The optimal solution that satisfies the efficiency-lifetime tradeoff is selected from the candidate strategy set as the tool-changing strategy. The optimal strategy is then subjected to time-series interpolation smoothing to generate the tool-changing time series and corresponding cutting parameter adjustment instructions. Example

[0044] This embodiment details how the global evaluation layer performs global tool change planning based on decision optimization parameters, providing macro-level tool change guidance for the entire machining process and ensuring high efficiency and stability.

[0045] Describing the evolution trend of tool wear: Non-uniform B-spline curves are used to describe the evolution trend of tool wear, and the trend is expressed as a time parameter. function Time parameters Covering the entire processing cycle. The expression for a non-uniform B-spline curve is: ,in It controls the vertices. yes By adjusting the control vertices and basis functions, the B-spline basis function can accurately fit the change of tool wear over time.

[0046] Constructing a wear condition assessment index: Multi-resolution analysis of tool vibration signals based on wavelet transform is performed to extract energy distribution characteristics across different frequency bands. Let the vibration signal be... The wavelet transform formula is: in It is a scale factor. It is the translation factor. It is the conjugate of the wavelet function. A wear state assessment index is constructed through feature fusion. This indicator comprehensively reflects the wear condition of the cutting tool.

[0047] The wear state evaluation index is constructed. In the specific implementation process, after the multi-resolution analysis of the tool vibration signal by wavelet transform, the energy distribution characteristics of different frequency bands extracted are important features participating in fusion, and the tool current wear, cutting force amplitude, temperature change gradient and workpiece material parameters and other features collected by the multi-source sensor are combined for fusion processing. First, the features are preprocessed to remove noise interference and standardized, so that different features are in the same order of magnitude, eliminating the influence of dimensional differences. Then, based on the mechanism analysis of tool wear, the importance of each feature in reflecting the tool wear state is determined, and each feature is given a corresponding weight. The high-frequency energy feature of the vibration signal is given a higher weight because it is closely related to the tool edge micro-collapse and other serious wear states, and the tool current wear, as a direct reflection of the wear degree, is also given a higher weight, and other features are given corresponding weights according to their influence on the wear state. Then, the weighted sum is used to fuse all the features, and the result is the wear state evaluation index I, which integrates the feature information reflecting the tool wear from multiple aspects and can comprehensively and accurately reflect the actual wear state of the tool, providing a reliable basis for subsequent global tool changing planning.

[0048] Construct global optimization constraints: Construct global optimization constraints, including maximum number of processing interruptions , minimum tool utilization rate and workpiece surface roughness limit . These constraints limit the tool changing planning from different aspects to ensure the smooth progress of the machining process.

[0049] Construct multi-objective decision function: Construct multi-objective decision function, including tool changing cost term , machining efficiency term and tool life term , coupled with the constraint term by Lagrange multiplier , the formula is: Where is the constraint function, is the Lagrange multiplier.

[0050] Construct a discrete optimization model: Discretize the machining period into multiple time windows , segment the linearization of the decision function, and construct a discrete optimization model. In each time window, the decision function is approximated, and the continuous optimization problem is converted into a discrete optimization problem for easy solution.

[0051] Solving the discretized optimization model: Branch and bound algorithm is used to solve the discretized optimization model, and pruning strategy is used to reduce the invalid search space. In the search process, the branch and bound algorithm continuously decomposes the problem into sub-problems, and through the calculation of the boundary value of the sub-problems, the sub-problems that are not likely to contain the optimal solution are excluded, thereby improving the search efficiency, and finally generating the global tool change planning sequence. Embodiment

[0052] This embodiment mainly illustrates how the dynamic adjustment layer optimizes local parameters according to real-time conditions to adapt to changes in the machining process, further improving the accuracy and flexibility of tool change decision-making. The specific method includes: Based on the current tool wear , processing load and remaining life , a dynamic adjustment window is constructed, and the window size is scaled in real time through the load adaptive coefficient , the window size formula is , where is a function of tool wear, processing load and remaining life, which is determined according to the actual situation.

[0053] An online learning mechanism is used to update the local environment model, and an incremental support vector machine is used to perform regression analysis on real-time sensor data to predict short-term wear trends. The optimization objective of the incremental support vector machine is , and the constraint condition is , , where is the weight vector, is the bias, is the slack variable, is the penalty parameter, is the kernel function mapping. By continuously updating the model, the accuracy of short-term wear trend prediction is improved.

[0054] A local optimization cost function is constructed, including a tool change delay penalty term , a cutting parameter deviation term and a machining quality fluctuation term, and the contribution proportion of each penalty term is adjusted through a dynamic weight matrix , and the cost function formula is .

[0055] The optimized parameters are checked for feasibility, and if the check fails, a backtracking mechanism is triggered to generate adjustment instructions again. The checking process judges whether the optimized parameters are feasible according to the constraints in actual machining, such as cutting parameter range, tool life limit, etc. If not, adjust the parameters again from the previous step to ensure the effectiveness of local parameter optimization. Embodiment

[0056] This embodiment details how the execution control layer implements the timing scheduling of tool changing actions based on fuzzy logic control algorithms, and introduces the self-healing control function of the system, improving the reliability and stability of the machining system.

[0057] Establishing the kinematic model: Establish a multi-degree-of-freedom kinematic model of the tool changing mechanism, including the dynamic equations of translational axis speed , rotational axis angle and clamping force . Take the common mechanical arm type tool changing mechanism as an example, its kinematic model can be established by D-H parameter method, through describing the relative position and attitude relationship of each joint, the kinematic equation of the tool changing mechanism is obtained.

[0058] Convert to fuzzy rule base: Convert the kinematic model to a fuzzy rule base, the input variables of the rule base include tool changing time deviation , tool position error and clamping force fluctuation , and the output variable is the control amount of each axis . For example, the fuzzy rule can be "if the tool changing time deviation is large and the tool position error is small, and the clamping force fluctuation is within the allowed range, then increase the translational axis speed".

[0059] Fuzzification and precision processing: Gaussian membership function is used for fuzzification processing of input and output variables, the formula of Gaussian membership function is , where is the center of the membership function, is the standard deviation. The precise output is realized by weighted average method, the formula is , where is the fuzzy output value, is the corresponding membership degree.

[0060] Construct timing constraints: Construct timing constraints, including maximum tool changing delay , mechanical arm movement range and torque limit . These constraints ensure that the tool changing action is executed safely and reliably within the specified time and space range.

[0061] Generate tool changing action sequence: Generate tool changing action sequence based on fuzzy rule base and constraint conditions, and coordinate the actions of each execution mechanism through timestamp synchronization mechanism. The timestamp synchronization mechanism ensures that each execution mechanism works in accordance with the predetermined order and time node, realizing efficient and accurate tool changing operation.

[0062] A self-repairing control module is embedded in the machining system. When abnormal wear is detected, a compensation strategy is generated based on reinforcement learning to dynamically adjust the cutting parameters or switch to a backup tool. The goal of reinforcement learning is to maximize cumulative rewards by constantly interacting with the environment to learn the optimal strategy. For example, when abnormal wear of the tool is detected, the system selects to adjust the cutting parameters such as cutting speed, feed rate, or switch to a backup tool based on the current machining state and reward mechanism to ensure the continuity and quality of machining.

[0063] It should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one from another entity or action without necessarily requiring or implying any actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0064] While embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, changes, and variations can be made in the embodiments without departing from the spirit and scope of the present application, which is defined by the appended claims and their equivalents.

Claims

1. A tool change decision method based on tool wear perception, characterized in that, The application relates to a method for intelligent tool changing decision of a machining system. The method comprises the following steps: collecting real-time state data of a machining tool through a multi-source sensor, wherein the multi-source sensor comprises a vibration sensor, an acoustic emission sensor, a force sensor and a temperature sensor; performing feature fusion processing on the real-time state data based on a multi-modal graph neural network to obtain tool wear feature data; inputting the tool wear feature data into a pre-trained Bayesian decision network, wherein the Bayesian decision network adopts a dynamic probability inference structure, iteratively optimizes a tool changing strategy based on a conditional probability distribution, and generates decision optimization parameters; constructing a multi-objective tool changing optimization model according to the decision optimization parameters, wherein the multi-objective tool changing optimization model takes the highest machining efficiency and the longest tool life as optimization objectives, and globally optimizes a tool changing time through an improved particle swarm algorithm, wherein the improved particle swarm algorithm introduces an adaptive inertia weight factor and a dynamic learning factor; and outputting optimal tool changing strategy data based on the multi-objective tool changing optimization model; establishing a hierarchical decision control model according to the optimal tool changing strategy data, wherein the hierarchical decision control model comprises a global evaluation layer, a dynamic adjustment layer and an execution control layer, the global evaluation layer performs global tool changing planning based on the decision optimization parameters, the dynamic adjustment layer performs local parameter optimization based on the optimal tool changing strategy data, and the execution control layer realizes time sequence scheduling of a tool changing action based on a fuzzy logic control algorithm; and outputting a tool changing control instruction through the hierarchical decision control model to realize intelligent tool changing decision of the machining system.

2. The method of claim 1, wherein, The method comprises the following steps: inputting the tool wear feature data into a pre-trained Bayesian decision network, wherein the Bayesian decision network adopts a dynamic probability inference structure, iteratively optimizes a tool changing strategy based on a conditional probability distribution, and generates decision optimization parameters. acquiring real-time state data, wherein the real-time state data comprises a current tool wear amount, a cutting force amplitude, a vibration frequency spectrum peak value, a temperature change gradient and a machining workpiece material parameter; constructing a multi-dimensional state space based on the real-time state data, and constructing an action space based on an executable tool changing delay time, a cutting parameter adjustment amount and a tool compensation amount; constructing a multi-objective joint probability model based on the state space and the action space, wherein the multi-objective joint probability model comprises a wear prediction item, an efficiency evaluation item and an energy consumption evaluation item, the wear prediction item is calculated through a Bayesian posterior probability of a tool wear rate and a remaining life, the efficiency evaluation item is calculated through a joint distribution of a unit time machining amount and a tool changing frequency, and the energy consumption evaluation item is calculated through Markov chain transition probability of a cutting power and a tool changing energy consumption; constructing a dynamic probability inference structure, wherein the structure comprises a prior network and a posterior network, the prior network and the posterior network both comprise an input layer, a graph convolution layer and an output layer, the dimension of the input layer matches the dimension of the state space, the graph convolution layer aggregates multi-modal features through an attention mechanism, and the output layer maps an action space distribution through a probability density function; and The prior network and the posterior network are trained by using a Monte Carlo sampling method, training samples are generated by an importance sampling strategy and stored in an experience pool, and samples are randomly selected from the experience pool to update network parameters, wherein the prior network optimizes a probability distribution by variational inference, and the posterior network updates a conditional probability by Gibbs sampling; and decision optimization parameters including network topology weights, state normalization matrices, and action distribution parameters are output based on the trained networks.

3. The method of claim 1, wherein, A multi-objective tool changing optimization model is constructed based on the decision optimization parameters, the model takes the highest machining efficiency and the longest tool life as optimization objectives, an improved particle swarm optimization algorithm is used to globally optimize tool changing time, and optimal tool changing strategy data is output based on the multi-objective tool changing optimization model, including: A multi-objective function for tool changing optimization is constructed, the function includes an efficiency optimization objective function and a life optimization objective function, wherein the efficiency optimization objective function is calculated by the ratio of a machining cycle to a tool changing time, and the life optimization objective function is calculated by the weighted sum of a tool wear cumulative amount and the reciprocal of a remaining life; Tool changing optimization constraints are constructed based on the multi-objective function, the constraints include a maximum allowable wear amount, a cutting parameter threshold, a tool changing time window, and a machining continuity requirement; Time series encoding is used to discretize the tool changing strategy, each time node contains a wear prediction value and a cutting parameter adjustment amount, and particle position and velocity are updated based on particle fitness and the best solution in the population; wherein the particle fitness is calculated by the weighted sum of the objective function value and the constraint violation degree; An adaptive inertia weight factor is introduced, the factor decreases linearly in segments as the number of iterations increases, and the global search and local development capabilities of the particle swarm are dynamically adjusted through the factor; A dynamic learning factor is introduced, the factor adaptively adjusts the weights of social learning and individual learning according to the particle distribution density, and the particle swarm convergence trajectory is optimized through the factor; based on the adaptive inertia weight factor and the dynamic learning factor, iterative optimization is performed, and the strategies generated at each iteration are screened by the Pareto frontier, and non-dominated solutions are stored in a candidate strategy set; An optimal solution that meets the efficiency-life trade-off is selected from the candidate strategy set as the tool changing strategy, and the optimal strategy is smoothed by time series interpolation to generate a tool changing time series and corresponding cutting parameter adjustment instructions.

4. The method of claim 1, wherein, The global evaluation layer performs global tool changing planning based on the decision optimization parameters, including: A non-uniform B-spline curve is used to describe the tool wear evolution trend, which is represented as a function of time parameter, and the time parameter covers the entire machining cycle; A wavelet transform is used to perform multi-resolution analysis on the tool vibration signal to extract energy distribution features in different frequency bands, and a wear state evaluation index is constructed by feature fusion; Global optimization constraints are constructed, including the maximum number of machining interruptions, the minimum tool utilization rate, and the workpiece surface roughness limit; A multi-objective decision function is constructed, including a tool changing cost term, a machining efficiency term, and a tool life term, and the constraint terms are coupled by Lagrange multipliers. Discretize the machining cycle into multiple time windows, piecewise linearize the decision function, and construct a discretized optimization model; Solve the discretized optimization model using a branch and bound algorithm, reduce invalid search space through pruning strategies, and generate a global tool change planning sequence.

5. The method of claim 1, wherein, The dynamic adjustment layer includes: Based on the current tool wear, processing load and remaining life, a dynamic adjustment window is constructed, and the window size is scaled in real time through a load adaptive coefficient; An online learning mechanism is used to update the local environment model, and incremental support vector machines are used to perform regression analysis on real-time sensor data to predict short-term wear trends; A local optimization cost function is constructed, which includes a tool change delay penalty term, a cutting parameter deviation term, and a processing quality fluctuation term, and the contribution proportion of each penalty term is adjusted through a dynamic weight matrix; If the feasibility check fails, a backtracking mechanism is triggered to generate adjustment instructions again.

6. The method of claim 1, wherein, The execution control layer includes: A multi-degree-of-freedom kinematics model of the tool changing mechanism is established, which includes translational axis velocity, rotational axis angle, and clamping force dynamic equation; The kinematics model is converted into a fuzzy rule base, the input variables of the rule base include tool change time deviation, tool position error and clamping force fluctuation, and the output variables are the control amounts of each axis; Gaussian membership functions are used to fuzz the input and output variables, and weighted average method is used to realize accurate output; Temporal constraints are constructed, including maximum tool change delay, mechanical arm movement range and torque limit; Based on the fuzzy rule base and the constraint conditions, a tool change action sequence is generated, and a time stamp synchronization mechanism is used to coordinate the actions of each execution mechanism.

7. The method of claim 1, wherein, Further comprising: Embed a self-healing control module in the machining system, when abnormal wear is detected, generate a compensation strategy based on reinforcement learning, and dynamically adjust the cutting parameters or switch to a backup tool.

8. An intelligent tool change decision system based on tool wear perception for implementing the method of any one of claims 1-7, characterized in that, Comprise: A data acquisition unit for acquiring real-time state data of the machining tool through multiple source sensors and performing feature fusion based on a multi-modal graph neural network; A decision optimization unit for generating tool change strategy optimization parameters through a Bayesian decision network and performing global optimization using an improved particle swarm algorithm; A hierarchical control unit including a global evaluation module, a dynamic adjustment module, and an execution control module for global planning, local optimization, and action scheduling, respectively; A tool change execution unit for driving the tool changing mechanism to complete tool replacement and parameter adjustment according to the control instructions.

Images