A laser processing quality prediction method based on an improved starfish optimization algorithm

Abstract

The application discloses a laser processing quality prediction method based on an improved starfish optimization algorithm and relates to the technical field of neural network models. The method comprises the following steps: first, improving the starfish optimization algorithm based on the Levy flight principle to obtain an improved starfish optimization algorithm; second, optimizing the weights and biases of a back propagation neural network based on the improved starfish algorithm, and inhibiting overfitting of the back propagation neural network based on an L1 regularization method to obtain a target back propagation neural network; third, acquiring process parameters of silicon carbide laser processing; and finally, inputting the process parameters into the pre-trained target back propagation neural network to obtain the processing quality parameters of the silicon carbide laser processing. The starfish optimization algorithm is improved by introducing the Levy flight principle, and high-precision prediction of the silicon carbide picosecond laser processing quality is realized.

Description

Technical Field

[0001] This invention relates to the field of neural network model technology, and more specifically to a laser processing quality prediction method based on an improved starfish optimization algorithm. Background Technology

[0002] Silicon carbide (SiC) ceramics are widely used in semiconductor manufacturing due to their high thermal conductivity, low coefficient of thermal expansion, and excellent mechanical stability, especially as the core material for bump chucks. The micropillar array structure on the surface of the bump chuck is crucial for achieving stable wafer adsorption and precise handling. Its geometrical consistency, ablation depth control precision, and surface quality directly affect the reliability of wafer adsorption and the requirement for damage-free handling.

[0003] Picosecond laser processing technology offers advantages such as non-contact processing, a small heat-affected zone, and high processing precision, making it particularly suitable for the precision machining of microstructures in hard and brittle materials like silicon carbide. It has gradually become the primary processing method for micropillar array structures. However, in actual processing, multiple process parameters, including laser power, laser frequency, number of processing passes, scan overlap rate (LOR), and spot overlap rate (SOR), interact and form complex nonlinear coupling relationships with key quality indicators such as ablation depth, surface roughness, and boss top circle diameter. This significantly increases the difficulty of predicting processing quality and optimizing parameters.

[0004] Existing process modeling and optimization methods are insufficient to effectively address the aforementioned issues. For example, orthogonal experimental design is limited by the range of preset parameter levels, making it difficult to cover optimal parameter combinations and limiting its ability to characterize complex nonlinear relationships; response surface methodology typically only constructs first- or second-order models, resulting in insufficient fitting accuracy for scenarios with strong coupling of multiple parameters; swarm intelligence algorithms such as the Standard Starfish Optimization Algorithm (SFOA) generally suffer from slow convergence speed and a tendency to get trapped in local optima during application, thus requiring further improvement in the accuracy and stability of processing quality prediction models built upon them.

[0005] Therefore, how to achieve high-precision prediction of silicon carbide picosecond laser processing quality under the condition of strong nonlinear coupling of multiple process parameters has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] The purpose of this invention is to provide a laser processing quality prediction method, device, equipment, and readable storage medium based on an improved starfish optimization algorithm. By introducing the Lévy flight principle to improve the starfish optimization algorithm, the global exploration capability during parameter search is enhanced, effectively reducing the slow convergence speed and susceptibility to local optima problems of the traditional starfish optimization algorithm. Furthermore, the improved starfish optimization algorithm is used to optimize the weights and biases of the backpropagation neural network, and L1 regularization is combined to suppress network overfitting. This improves the generalization ability and prediction stability of the neural network model under conditions of strong nonlinear coupling of multiple process parameters. By using key process parameters such as laser processing frequency, laser processing power, number of processing cycles, scan overlap rate, and spot overlap rate as inputs, accurate prediction of processing quality parameters such as laser ablation depth, ablation surface roughness, and boss top circle diameter is achieved. This provides a reliable basis for optimizing and adjusting process parameters during silicon carbide laser processing, improving the consistency and controllability of microstructure processing quality.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] In a first aspect, the present invention provides a laser processing quality prediction method based on an improved starfish optimization algorithm, the method comprising:

[0009] The starfish optimization algorithm is improved based on the Levy flight principle, resulting in the improved starfish optimization algorithm;

[0010] The weights and biases of the backpropagation neural network are optimized based on the improved starfish algorithm, and the overfitting of the backpropagation neural network is suppressed based on the L1 regularization method to obtain the target backpropagation neural network.

[0011] Obtain the process parameters for silicon carbide laser processing; these parameters include laser processing frequency, laser processing power, number of processing cycles, scanning overlap rate, and spot overlap rate.

[0012] The process parameters are input into a pre-trained target backpropagation neural network to obtain the processing quality parameters of silicon carbide laser processing; the processing quality parameters include laser ablation depth, laser ablation surface roughness, and boss top circle diameter.

[0013] In some embodiments, the starfish optimization algorithm is improved based on the Lévy flight principle to obtain an improved starfish optimization algorithm, including:

[0014] Levy flight is introduced into the breeding behavior of the starfish optimization algorithm to optimize the global exploration formula of the starfish optimization algorithm;

[0015] Levy flight perturbation is introduced into the predation behavior of the starfish optimization algorithm;

[0016] An improved starfish optimization algorithm is obtained based on the adaptive adjustment of the step size of the Levy flight step size.

[0017] In some embodiments, the network parameters of the backpropagation neural network are optimized based on the improved starfish algorithm, and the overfitting of the backpropagation neural network is suppressed based on the L1 regularization method to obtain the target backpropagation neural network, including:

[0018] The fitness function of the backpropagation neural network is determined based on the L1 regularization method;

[0019] The optimal parameters of the backpropagation neural network are determined based on the improved starfish algorithm, and the target backpropagation neural network is obtained.

[0020] In some embodiments, the optimal parameters of the backpropagation neural network are determined based on the improved starfish algorithm to obtain the target backpropagation neural network, including:

[0021] The improved starfish algorithm is used to iterate on each individual to obtain the fitness of each individual;

[0022] The individual with the highest fitness is retained, and some individuals are selected to be retained from the remaining individuals. This process is repeated until the preset conditions are met, and the optimal parameters of the backpropagation neural network are determined.

[0023] Substituting the optimal parameters into the backpropagation neural network yields the target backpropagation neural network.

[0024] In some embodiments, obtaining the process parameters for silicon carbide laser processing includes:

[0025] The original parameters for silicon carbide laser processing were obtained, and the original parameters were cleaned to obtain candidate parameters.

[0026] The candidate parameters are normalized to obtain the process parameters.

[0027] In some embodiments, the training method for the target backpropagation neural network includes:

[0028] Obtain sample process parameters for silicon carbide laser processing;

[0029] The sample process parameters are input into the target backpropagation neural network to obtain the sample processing quality parameters;

[0030] Update the control weights of the target backpropagation neural network until the training stopping condition is met.

[0031] Secondly, the present invention also provides a laser processing quality prediction device based on an improved starfish optimization algorithm, the device comprising:

[0032] The algorithm optimization module is used to improve the starfish optimization algorithm based on the Levy flight principle, resulting in an improved starfish optimization algorithm.

[0033] The network improvement module is used to optimize the weights and biases of the backpropagation neural network based on the improved starfish algorithm, and to suppress overfitting of the backpropagation neural network based on the L1 regularization method, so as to obtain the target backpropagation neural network.

[0034] The parameter acquisition module is used to acquire the process parameters for silicon carbide laser processing; the process parameters include laser processing frequency, laser processing power, number of processing cycles, scanning overlap rate, and spot overlap rate.

[0035] The quality prediction module is used to input process parameters into a pre-trained target backpropagation neural network to obtain the processing quality parameters of silicon carbide laser processing. The processing quality parameters include laser ablation depth, laser ablation surface roughness, and boss top circle diameter.

[0036] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the laser processing quality prediction method based on the improved starfish optimization algorithm provided in the first aspect.

[0037] Fourthly, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the laser processing quality prediction method based on the improved starfish optimization algorithm provided in the first aspect.

[0038] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the laser processing quality prediction method based on the improved starfish optimization algorithm provided in the first aspect.

[0039] The beneficial effects of this invention are as follows: First, the starfish optimization algorithm is improved based on the Levy flight principle to obtain the improved starfish optimization algorithm; then, the weights and biases of the backpropagation neural network are optimized based on the improved starfish algorithm, and the overfitting of the backpropagation neural network is suppressed based on the L1 regularization method to obtain the target backpropagation neural network; then, the process parameters of silicon carbide laser processing are obtained; the process parameters include laser processing frequency, laser processing power, number of processing times, scanning overlap rate, and spot overlap rate; finally, the process parameters are input into the pre-trained target backpropagation neural network to obtain the processing quality parameters of silicon carbide laser processing; the processing quality parameters include laser ablation depth, laser ablation surface roughness, and boss top circle diameter. By introducing the Levy principle to improve the starfish optimization algorithm, the algorithm's global exploration capability during parameter search is enhanced, effectively reducing the slow convergence speed and susceptibility to local optima problems of the traditional starfish optimization algorithm. Based on this, the weights and biases of the backpropagation neural network are optimized using the improved starfish optimization algorithm, and L1 regularization is combined to suppress network overfitting. This improves the generalization ability and prediction stability of the neural network model under conditions of strong nonlinear coupling of multiple process parameters. By using key process parameters such as laser processing frequency, laser processing power, number of processing cycles, scanning overlap rate, and spot overlap rate as inputs, accurate prediction of processing quality parameters such as laser ablation depth, ablation surface roughness, and boss top circle diameter is achieved. This provides a reliable basis for optimizing and adjusting process parameters during silicon carbide laser processing, improving the consistency and controllability of microstructure processing quality.

[0040] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0041] Figure 1 This is a flowchart illustrating a laser processing quality prediction method based on an improved starfish optimization algorithm, according to an embodiment of the present invention.

[0042] Figure 2 This is a schematic diagram of a laser processing quality prediction device based on an improved starfish optimization algorithm, according to an embodiment of the present invention.

[0043] Figure 3 This is a schematic diagram of an electronic device structure provided in an embodiment of this application. Detailed Implementation

[0044] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0045] It should be noted that references to "an embodiment," "embodiment," "example embodiment," etc., in this specification refer to the described embodiment including specific features, structures, or characteristics; however, not every embodiment must include these specific features, structures, or characteristics. Furthermore, such expressions do not refer to the same embodiment. Moreover, when describing specific features, structures, or characteristics in conjunction with embodiments, whether or not explicitly described, it is indicated that incorporating such features, structures, or characteristics into other embodiments is within the knowledge of those skilled in the art.

[0046] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0047] In some embodiments, such as Figure 1 As shown, Figure 1 A flowchart illustrating a laser processing quality prediction method based on an improved starfish optimization algorithm is provided. The specific method includes:

[0048] S101, based on the Levy flight principle, improves the starfish optimization algorithm to obtain the improved starfish optimization algorithm.

[0049] Specifically, in order to better illustrate this embodiment, the starfish optimization algorithm and the Levy flight principle will be explained first.

[0050] Among them, the Starfish Optimization Algorithm (SFOA) is a swarm intelligence optimization algorithm inspired by the foraging and reproductive behaviors of starfish in nature. It simulates the process by which starfish search for food in the marine environment (optimizing the optimal solution of the objective function) and reproduce asexually through division (maintaining population diversity and exploring new areas).

[0051] SFOA comprises two main phases: exploration and development. The exploration phase employs a hybrid search model combining five-dimensional and one-dimensional approaches to simulate the starfish's exploratory behavior, improving computational efficiency and ensuring search capability. The development phase simulates the starfish's predation and regeneration behavior, employing a bidirectional search strategy and unique movement patterns to guarantee convergence in the development process. Its core principle lies in achieving efficient exploration and development of a complex solution space through a dimension-adaptive hybrid search strategy and a bidirectional guided individual update mechanism.

[0052] The population initialization algorithm first randomly generates a uniformly distributed initial population in the solution space, and the position of each individual (starfish) is represented as:

[0053] ;

[0054] in, For the first The individual in the first The position of the dimension and These are the upper and lower bounds of that dimension, respectively. The value is a random number within the range [0, 1]. After initialization, the fitness value of each individual is calculated, and the global optimal solution is recorded. The exploration phase employs a hybrid search mechanism that simulates the behavior of starfish using their multi-arm structure for wide-area exploration. A dimension-adaptive hybrid search strategy is used: when the dimension... At 5 o'clock, a five-dimensional search mode was adopted, in which only five dimensions were randomly selected for updating in each iteration, simulating the coordinated detection of the five arms of a starfish:

[0055] ;

[0056] in, For random coefficients, The angle parameter decays with iteration. Represents a sine or cosine function to control the search direction. When the dimension... At 5 o'clock, a one-dimensional search mode is adopted, updating only one dimension at a time, and using group information for guidance:

[0057] ;

[0058] in, The individual energy coefficient decreases with iteration; , The weights are random. The individual index is randomly selected. This design significantly improves the search efficiency of the algorithm in high-dimensional indivisible problems. During the development phase, predation and regeneration behaviors are simulated, mimicking the starfish's predation strategies and regeneration capabilities. A parallel bidirectional search mechanism enhances local exploitation and global escape capabilities.

[0059] Predatory behavior: An individual updates its position based on the global optimum and the directions of two randomly selected neighboring individuals.

[0060] ;

[0061] in, , This is the deviation vector from the global optimal solution. , These are random coefficients.

[0062] Regeneration behavior: The regeneration operation is performed only on the last individual in the population, simulating the starfish's self-repair ability after being damaged.

[0063] ;

[0064] This mechanism effectively enhances the global convergence of the algorithm with minimal computational overhead. Boundary handling and iteration termination: All updated positions undergo boundary constraint processing; if a position exceeds the feasible region, its original position is retained. After each iteration, the fitness is re-evaluated and the global optimum is updated. The algorithm terminates after reaching a preset maximum number of iterations, outputting the historical optima and the convergence trajectory.

[0065] The Levy flight principle is a special type of random walk whose step size follows a heavy-tailed Levy stable distribution. This strategy simulates the movement trajectories of many organisms in nature (such as albatrosses, fruit flies, and bees) during resource searching. Its core characteristic is that it involves mostly short-distance local exploration, occasionally interspersed with extremely long jumps. This search strategy is widely used in metaheuristic optimization algorithms, effectively balancing global exploration and local exploitation capabilities. Levy flight is a non-Gaussian random process, and its step size distribution follows a Levy stable distribution. The probability density function of the Levy stable distribution is relatively complex and is usually defined by its characteristic function. For the one-dimensional case, the characteristic function of the Levy stable distribution is:

[0066] ;

[0067] Where α is the stability index (0 < α ≤ 2) and β is the skewness parameter (-1 ≤ β ≤ 1). It is a scale parameter When α=2, the Lévy distribution degenerates into a normal distribution; when α=1 and β=0, it degenerates into a Cauchy distribution. In optimization algorithms, α=1.5 is usually chosen, at which point Lévy flight exhibits good global and local search capabilities. The step size of Lévy flight has the following characteristics: most of the time it involves short-distance local searches, with occasional long-distance jumps. This characteristic allows the algorithm to explore a large search space while also performing fine-grained exploration in local regions, which helps avoid the algorithm getting trapped in local optima.

[0068] Optionally, the starfish optimization algorithm can be improved based on the Levy flight principle to obtain an improved starfish optimization algorithm, including: introducing Levy flight into the breeding behavior of the starfish optimization algorithm to optimize the global exploration formula of the starfish optimization algorithm; introducing Levy flight perturbation into the predation behavior of the starfish optimization algorithm; and adaptively adjusting the step size based on the Levy flight step size to obtain the improved starfish optimization algorithm.

[0069] Specifically, the starfish algorithm has shown some potential in solving complex optimization problems, but it also has shortcomings such as slow convergence speed and easy getting trapped in local optima. Therefore, the Levy flight mechanism is introduced to optimize and improve the starfish optimization algorithm.

[0070] In LF-SFOA, Levy's flight is introduced into the global exploration (breeding behavior) of SFOA to enhance the algorithm's global search capability. Specific improvements are as follows:

[0071] First, Levi's flight is introduced into the breeding behavior, modifying the original global exploration formula to:

[0072] ;

[0073] in, Levy represents element-wise multiplication. It is the Levi flight step length vector.

[0074] Introducing the Lévy flight perturbation into the predation behavior, with a certain probability of adding the Lévy flight perturbation after a local search, avoids getting trapped in local optima:

[0075] ;

[0076] Adaptive Lévy flight step size, with the step size scaling factor α adaptively adjusted with the number of iterations:

[0077] ;

[0078] in, It is the initial step size. It is the attenuation coefficient. It represents the maximum number of iterations.

[0079] The specific process of the Lever Flight Enhanced Starfish Optimization Algorithm (LF-SOFA) is as follows:

[0080] Initialize the population, with a population size of N and the position of each starfish individual as follows: (D represents the problem dimension).

[0081] Randomly initialize within the search space:

[0082] ;

[0083] in , For the dimensional boundary, These are uniformly random numbers.

[0084] Secondly, there is the behavioral selection mechanism.

[0085] Starfish individuals choose the following behaviors based on probability ( (For leader threshold)

[0086] Leader behavior (Levi transfer): If Perform Levi's flight behavior (global exploration);

[0087] Follower behavior (cooperative predation): Otherwise, move toward the best individual (local exploitation);

[0088] Random movement (injury recovery): based on probability Randomly reset the location (preserving diversity).

[0089] Then, the location update strategy is discussed, including:

[0090] Levi's Flight: A Simulation of Long-Distance Starfish Migration

[0091] ;

[0092] in: This is the step size scaling factor; This represents element-wise multiplication; Generated using the Mantegna algorithm :

[0093] ;

[0094] ;

[0095] Towards the globally optimal individual move:

[0096] ;

[0097] It is a random number; The system simulates group collaboration by randomly selecting individuals.

[0098] Global random exploration:

[0099] ;

[0100] Use a mirror reflection strategy to prevent out-of-bounds access:

[0101] ;

[0102] Calculate the updated population fitness ;

[0103] Update the global optimal solution.

[0104] Repeat the above process until the following condition is met:

[0105] Maximum number of iterations or the rate of change of the optimal solution Output and .

[0106] Lévy Flight provides powerful global exploration capabilities, achieving global region coverage through long-distance jumps with heavy-tailed distributions; while SFOA provides development capabilities for fine-grained local search, achieving rapid convergence through elite guidance. The combination of the two (LF-SFOA) achieves a dynamic balance between exploration and development, thereby improving the overall performance of the algorithm.

[0107] S102, the weights and biases of the backpropagation neural network are optimized based on the improved starfish algorithm, and the overfitting of the backpropagation neural network is suppressed based on the L1 regularization method to obtain the target backpropagation neural network.

[0108] Specifically, let's first explain the backpropagation neural network (BP neural network). The backpropagation neural network (BP neural network) is a multi-layer feedforward neural network trained using an error backpropagation algorithm, and it is one of the most widely used neural network models. A BP neural network consists of an input layer, hidden layers (which can be multiple), and an output layer. Each layer consists of multiple neurons connected by weights. Information propagates unidirectionally from the input layer to the output layer, making it a feedforward network. The BP neural network employs a typical three-layer feedforward structure: Input layer: receives external feature vectors. The number of nodes equals the feature dimension; Hidden layers: perform non-linear feature transformations, the number of layers and nodes depends on the problem complexity; Output layer: produces the final prediction result. The number of nodes corresponds to the output dimension, and adjacent layers use a fully connected approach. Layer The net input to each neuron is:

[0109] ;

[0110] in, For connection weights, For bias terms, This activates the output of the previous layer.

[0111] The working process of a BP neural network consists of two stages: forward propagation: information is passed from the input layer through the hidden layers to the output layer, layer by layer. Layer The activation output of each neuron is:

[0112] ;

[0113] Common activation functions include the Sigmoid function, Tanh function, and ReLU function. Backpropagation: If there is an error between the actual output of the output layer and the expected output, the error is backpropagated to adjust the weights and biases of neurons in each layer. The number of hidden layers is positively correlated with prediction accuracy; increasing the number of hidden layers can improve prediction accuracy. However, the number of hidden layers is also positively correlated with training time. Increasing the number of hidden layers prolongs the network's training time and reduces the training speed, increasing time costs in practical applications. Since this experimental data includes three output quantities—processing depth, surface roughness, and boss top circle diameter—a single-layer hidden layer BP neural network can fit the nonlinear functional relationship between laser processing parameters and the silicon carbide microstructure.

[0114] Optionally, the fitness function of the backpropagation neural network can be determined based on the L1 regularization method; the optimal parameters of the backpropagation neural network can be determined based on the improved starfish algorithm to obtain the target backpropagation neural network.

[0115] The process of determining the optimal parameters of the backpropagation neural network based on the improved starfish algorithm to obtain the target backpropagation neural network includes: iterating each individual based on the improved starfish algorithm to obtain the fitness of each individual; retaining the individual with the highest fitness and selecting some individuals to retain from the remaining individuals until the iteration reaches the preset condition to determine the optimal parameters of the backpropagation neural network; and substituting the optimal parameters into the backpropagation neural network to obtain the target backpropagation neural network.

[0116] Specifically, algorithm initialization: Set the topology of the BP neural network to nmk (input layer - hidden layer - output layer), and the starfish population size to [value missing]. =30, maximum number of iterations is 50 (controlling the iteration convergence boundary), Lévy flight step size factor. =1.5 (balancing random walk characteristics and stability), the network parameters for the position vector encoding of individual starfish are:

[0117] ;

[0118] in, denoted as the total dimension of the parameters; P is the weight matrix from the input layer to the hidden layer; T is the weight matrix from the hidden layer to the output layer; For hidden layer bias; Set the output layer bias. Initialize the population:

[0119] ;

[0120] Where L and U are the lower and upper bounds of the search space, used to limit the range of values ​​of the optimization variables. Let L = -4 and U = 4.

[0121] Fitness function design: The fitness function is used to quantitatively evaluate the performance of the network parameters represented by each starfish individual. For an individual First, its position vector is decoded into specific network parameters. Initialize the BP network with this set of parameters and calculate its prediction mean square error on the training set:

[0122] ;

[0123] Where N is the number of training samples, For the first The true output vector of each unit. Let be the network's predicted output vector. To prevent overfitting and enhance the model's generalization ability, the fitness function is defined as the reciprocal of the MSE including the L1 regularization term: The fitness function is defined as:

[0124] ;

[0125] Where λ is the L1 regularization coefficient (λ=0.01 in this paper), the larger the fitness value, the better the network performance corresponding to this set of parameters.

[0126] The starfish-like position update mechanism of the Levi flight: To balance global exploration and local development capabilities, LF-SFOA employs the following four strategies for position updates, each executed with a specific probability:

[0127] An individual performs a Levi flight with probability P1=0.3, utilizing its long-step random jump characteristic to explore new regions of the solution space.

[0128] ;

[0129] The step size scaling factor α adopts an exponential decay strategy, with α=0.1 being the initial step size; β is the Lévy flight random step size vector, generated by the Mantegna algorithm.

[0130] ;

[0131] ;

[0132] In the formula, Γ ( ) is the Gamma function.

[0133] The starfish moves in random directions, simulating a global search:

[0134] ;

[0135] in, This is the decay step size; It is a uniform random vector.

[0136] Starfish tend to move closer to the optimal individual, improving their local search capabilities.

[0137] ;

[0138] in, For learning efficiency, It is a Gaussian random vector.

[0139] If an individual's fitness does not improve for G=5 consecutive generations and remains below the population average fitness, then the dynamic mutation probability is used. Randomly reset some of its dimensions and simulate regeneration to escape local optima.

[0140] ;

[0141] For individuals The dimension:

[0142] ;

[0143] Elite Preservation and Selection Strategies: Employing a selection strategy combining elite preservation and roulette wheel selection to maintain population diversity and evolutionary direction.

[0144] Elite retention: The individual with the highest fitness in the current generation is retained and directly enters the next generation to ensure that the historical optimal solution is not destroyed.

[0145] ;

[0146] Roulette wheel selection: Except for elite individuals, the remaining individuals in the next generation population are selected probabilistically based on their fitness values. The probability of an individual being selected is:

[0147] ;

[0148] Termination condition: Iteration stops when any of the following conditions are met:

[0149] or ;

[0150] in, This is the convergence threshold.

[0151] Optimizing the BP neural network: After obtaining the optimal parameters, the forward propagation formula is:

[0152] ;

[0153] in, The network parameters are fine-tuned using the Sigmoid activation function via backpropagation.

[0154] S103, obtain the process parameters for silicon carbide laser processing.

[0155] The process parameters include laser processing frequency, laser processing power, number of processing cycles, scanning overlap rate, and spot overlap rate.

[0156] Optionally, the process parameters for silicon carbide laser processing are obtained, including: obtaining the original parameters for silicon carbide laser processing, cleaning the original parameters to obtain candidate parameters, and normalizing the candidate parameters to obtain the process parameters.

[0157] Specifically, the original parameters for silicon carbide laser processing are obtained, and outliers in each feature dimension are removed using the 3σ criterion. For missing values, mean imputation (for continuous data) or nearest neighbor imputation (for time-series data) is selected based on the characteristics of the dataset. Simultaneously, input features strongly correlated with the output target are screened using the Pearson correlation coefficient, and redundant features are removed to simplify the network structure, resulting in candidate parameters. The Min-Max normalization method is used to map all data to the [0, 1] interval to eliminate weight update bias caused by dimensional differences. Let the input sample matrix be P, containing five process parameters: laser processing frequency, laser processing power, number of processing operations, LOR (scan overlap rate), and SOR (spot overlap rate); the target output matrix is ​​T, containing three quality indicators: laser ablation surface roughness, laser ablation depth, and boss top circle diameter. To eliminate the influence of dimensions, the input and output data are subjected to minimax normalization to the [0, 1] interval.

[0158] .

[0159] S104: Input the process parameters into the pre-trained target backpropagation neural network to obtain the processing quality parameters of silicon carbide laser processing.

[0160] The processing quality parameters include laser ablation depth, laser ablation surface roughness, and boss top circle diameter.

[0161] Specifically, the process parameters are input into the pre-trained target backpropagation neural network, which can then output the laser ablation depth, laser ablation surface roughness, and boss top circle diameter.

[0162] It should be noted that the training method of the target backpropagation neural network includes: obtaining sample process parameters of silicon carbide laser processing; inputting the sample process parameters into the target backpropagation neural network to obtain sample processing quality parameters; updating the control weights of the target backpropagation neural network until the training stopping condition is reached.

[0163] Specifically, a 3-layer feedforward neural network is created. The input data matrix of the training set (normalized) and the target output matrix of the training set (normalized) obtained above are imported into the network. Let m=5 and n=3 be the number of nodes in the input layer and output layer of the BP network, respectively. According to Kolmogorov's theorem, the formula for calculating the number of hidden layer nodes is: hiddennum=2×inputnum+1.

[0164] P different samples , ,in , representing the magnitudes of the five feature values ​​(laser power, laser frequency, number of scans, SOR, LOR) of the i-th sample, respectively;

[0165] , representing the magnitudes of the three sample values ​​(processing depth, surface roughness, and top circle diameter) of the i-th sample, respectively.

[0166] The maximum number of training iterations is set to 1000; the learning rate, controlling the step size of weight updates, is set to 0.1; the target error is set to 0.00001, which is one of the conditions for stopping training, stopping when the error is less than 0.00001; the display frequency is set to 25, displaying training results once every 25 training iterations; the momentum factor is set to 0.01 to accelerate convergence and avoid getting trapped in local optima; the minimum performance gradient is set to 1×10⁻⁶, used to determine whether to stop training.

[0167] The maximum number of validation failures is set to 6. If the validation error does not improve in 6 consecutive iterations, training is stopped. The hidden layer uses the non-linear activation function logsig, the output layer uses the purelin function, and training uses the trainlm function.

[0168] After the model is built, according to the machine learning algorithm process, its performance needs to be evaluated. If the performance is unsatisfactory, the parameters need to be readjusted or a second modeling process needs to be performed. The metrics for regression problems are Root Mean Square Error (RMSE), Coefficient of Determination (R²), Mean Absolute Error (MAE), and Mean Bias Error (MBE). RMSE expresses the deviation between the predicted and actual values, with units consistent with the target variable. It intuitively reflects the sensitivity of the average error to outliers. The formula for calculating RMSE is as follows:

[0169] ;

[0170] The coefficient of determination (R²) is the proportion of the variance in the target variable that a model can explain. It typically ranges from 0 to 1; the closer to 1, the stronger the model's explanatory power. The formula for calculating R² is as follows:

[0171] ;

[0172] Mean Absolute Error (MAE) is the average of the absolute differences between predicted and actual values. It is robust to outliers and reflects the mean absolute error. The formula for calculating MAE is as follows:

[0173] ;

[0174] The mean deviation error (MBE) is the average deviation between the predicted and actual values. The absolute value should be as small as possible to avoid systematic bias. The formula for calculating MBE is as follows:

[0175] ;

[0176] in, Represents the actual value. This represents the predicted value, and N is the number of experiments run.

[0177] The laser processing quality prediction method based on the improved starfish optimization algorithm in the above embodiments improves the starfish optimization algorithm by introducing the Lévy flight principle, enhancing the algorithm's global exploration capability during parameter search and effectively reducing the problems of slow convergence speed and easy getting trapped in local optima in the traditional starfish optimization algorithm. Based on this, the improved starfish optimization algorithm is used to optimize the weights and biases of the backpropagation neural network, and L1 regularization is combined to suppress network overfitting. This improves the generalization ability and prediction stability of the neural network model under the condition of strong nonlinear coupling of multiple process parameters. By using key process parameters such as laser processing frequency, laser processing power, number of processing cycles, scanning overlap rate, and spot overlap rate as inputs, accurate prediction of processing quality parameters such as laser ablation depth, ablation surface roughness, and boss top circle diameter is achieved. This provides a reliable basis for optimizing and adjusting process parameters during silicon carbide laser processing, improving the consistency and controllability of microstructure processing quality.

[0178] In another embodiment, to fully test the performance of the proposed LF-SFOA algorithm, four representative single-peak test functions and four multi-peak test functions were selected for experiments, and Particle Swarm Optimization (PSO), Starfish Optimization (SFOA), and Grey Wolf Optimization (GWO) algorithms were selected for comparative testing. Benchmark functions F1 to F8 were used to compare the performance of PSO, LF-SFOA, SFOA, and GWO, where F1 to F4 are single-peak test functions and F5 to F8 are multi-peak test functions.

[0179] Single-peaked test functions are a class of standard functions with specific properties. Their significant characteristic is that only one global optimum exists within the domain. Single-peaked test functions are suitable for testing the search speed and convergence performance of algorithms. Multi-peaked test functions are a class of optimization problems with complex characteristics, containing multiple local optima but only one global optimum. They are mainly used to test the algorithm's ability to escape local optima and its global search capability. To ensure the reliability of the algorithm performance test results, the problem dimension of the four optimization algorithms was set to 30, the population size to 50, the maximum number of iterations to 500, and the number of independent runs to 10. This paper selects the average, standard deviation, optimal value, and worst value of the 10 test results as evaluation metrics for algorithm performance. The optimal value and average value are used to evaluate the convergence of the algorithm; the smaller the optimal value and average value, the better the convergence. The worst value is used to evaluate the robustness of the algorithm; the smaller the worst value and the closer it is to the average value, the better the robustness. The standard deviation is used to evaluate the stability of the algorithm; the smaller the standard deviation, the better the stability. Convergence curves can intuitively reflect the convergence speed, stability, and global search capability of an algorithm. The performance of the LF-SFOA algorithm, along with SFOA, PSO, and GWO algorithms, was tested using the test functions in Tables 1 and 2, and the results were compared and analyzed.

[0180] The results of solving the single-peak test function using the four algorithms are shown in Table 1:

[0181] Table 1. Results of solving the single-peak test function using four algorithms.

[0182]

[0183] The results of solving the multi-peak test function using the four algorithms are shown in Table 2:

[0184] Table 2 Results of solving the multi-peak test function using four algorithms

[0185]

[0186] According to the test results of each algorithm in Tables 1 and 2, the LF-SFOA algorithm has an advantage in finding the optimal value compared to the other three algorithms, and performs well in terms of convergence accuracy, convergence speed and optimization stability. Among the F1 to F8 test functions, the LF-SFOA algorithm has the smallest standard deviation and the best stability, and has certain advantages compared to other algorithms, showing good performance in terms of stability and applicability.

[0187] In another embodiment, to verify the performance of the LF-SFOA-BP multi-objective prediction model, a comparative experiment was conducted with SFOA-BP (original starfish algorithm), PSO-BP (particle swarm optimization algorithm), GWO-BP (grey wolf optimization algorithm), and the traditional BP model. In the experiment, the previously acquired dataset was first input into each of the above models, and model training was completed following a unified training process; subsequently, the root mean square error (RMSE) and coefficient of determination (R²) of each model were calculated. 2 The specific values ​​for these two core evaluation indicators.

[0188] The test results show that LF-SFOA-BP improves the machining depth prediction accuracy by approximately 5.6% compared to the original SFOA-BP and by approximately 52.1% compared to the traditional BP, indicating its strong fitting ability when handling nonlinear mapping relationships. The most significant improvement is in roughness prediction accuracy, with LF-SFOA-BP improving by approximately 55.6% compared to BP. This demonstrates the model's advantage in handling small-scale, high-precision indicators. While LF-SFOA-BP still performs well on this indicator, the differences between the various optimization algorithms are relatively small, suggesting that this indicator may have low sensitivity to optimization algorithms.

[0189] Based on the same inventive concept, this application also provides a laser processing quality prediction device based on the improved starfish optimization algorithm for implementing the laser processing quality prediction method based on the improved starfish optimization algorithm described above. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the laser processing quality prediction device based on the improved starfish optimization algorithm provided below can be found in the limitations of the laser processing quality prediction method based on the improved starfish optimization algorithm described above, and will not be repeated here.

[0190] In one embodiment, such as Figure 2 As shown, a laser processing quality prediction device based on an improved starfish optimization algorithm is provided. The device includes:

[0191] The algorithm optimization module 30 is used to improve the starfish optimization algorithm based on the Levy flight principle, resulting in an improved starfish optimization algorithm.

[0192] The network improvement module 31 is used to optimize the weights and biases of the backpropagation neural network based on the improved starfish algorithm, and to suppress overfitting of the backpropagation neural network based on the L1 regularization method, so as to obtain the target backpropagation neural network.

[0193] The parameter acquisition module 32 is used to acquire the process parameters of silicon carbide laser processing; the process parameters include laser processing frequency, laser processing power, number of processing times, scanning overlap rate and spot overlap rate;

[0194] The quality prediction module 33 is used to input process parameters into a pre-trained target backpropagation neural network to obtain the processing quality parameters of silicon carbide laser processing; the processing quality parameters include laser ablation depth, laser ablation surface roughness, and boss top circle diameter.

[0195] This application also provides an electronic device, in some embodiments, referring to... Figure 3 As shown, the electronic device 700 includes an input unit 710, a memory 720, a processor 730, and an output unit 740. The memory 720 stores program instructions that can be executed on the processor 730. The processor 730 can execute the laser processing quality prediction method and / or technical solution based on the improved starfish optimization algorithm in the foregoing embodiments by calling the program instructions. The electronic device 700 can be a mobile terminal device such as a mobile phone or a computer.

[0196] Furthermore, embodiments of this application also provide a computer-readable storage medium for storing a computer program that executes a laser processing quality prediction method based on an improved starfish optimization algorithm. For example, computer program instructions, when executed by a computer, can invoke or provide the methods and / or technical solutions according to this application through the operation of the computer. The program instructions that invoke the methods of this application may be stored in a fixed or removable storage medium, and / or transmitted via data streams in broadcast or other signal carrying media, and / or stored in a storage medium that operates according to the program instructions.

[0197] Obviously, those skilled in the art should understand that the modules or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device, or fabricating them separately as individual integrated circuit modules, or fabricating multiple modules or steps as a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0198] The technical features of the above embodiments can be arbitrarily integrated. For the sake of brevity, not all possible integrations of the technical features in the above embodiments are described. However, as long as the integration of these technical features does not contradict each other, they should be considered to be within the scope of this specification.

[0199] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A laser processing quality prediction method based on an improved starfish optimization algorithm, characterized in that, The method includes: An improved starfish optimization algorithm is obtained by improving the starfish optimization algorithm based on the Levy flight principle. Specifically, the algorithm includes: introducing Levy flight into the breeding behavior of the starfish optimization algorithm to optimize the global exploration formula of the starfish optimization algorithm; introducing Levy flight perturbation into the predation behavior of the starfish optimization algorithm; and adaptively adjusting the step size based on the Levy flight step size. The target backpropagation neural network is obtained by optimizing the weights and biases of the backpropagation neural network using an improved starfish algorithm and suppressing overfitting using L1 regularization. Specifically, this involves: determining the fitness function of the backpropagation neural network using L1 regularization; and determining the optimal parameters of the backpropagation neural network using the improved starfish algorithm. The determination of the optimal parameters using the improved starfish algorithm includes: iterating through each individual using the improved starfish algorithm to obtain the fitness of each individual; retaining the individual with the highest fitness and selecting some individuals from the remaining individuals until the iteration reaches a preset condition, thereby determining the optimal parameters of the backpropagation neural network; and substituting the optimal parameters into the backpropagation neural network to obtain the target backpropagation neural network. Obtaining the process parameters for silicon carbide laser processing specifically includes: obtaining the original parameters for silicon carbide laser processing; using the 3σ criterion to remove outliers in each feature dimension; for missing values, selecting mean filling or nearest neighbor filling based on the characteristics of the dataset; simultaneously, using the Pearson correlation coefficient to screen input features strongly correlated with the output target, removing redundant features to simplify the network structure, and obtaining candidate parameters; using the Min-Max normalization method to map the candidate parameters to the [0, 1] interval, obtaining the process parameters; the process parameters include laser processing frequency, laser processing power, number of processing operations, scanning overlap rate, and spot overlap rate; the formula for the Min-Max normalization method is: ； Where P is a candidate parameter. These are process parameters; The process parameters are input into a pre-trained target backpropagation neural network to obtain the processing quality parameters of the silicon carbide laser processing; wherein the training method of the target backpropagation neural network includes: obtaining sample process parameters of silicon carbide laser processing; inputting the sample process parameters into the target backpropagation neural network to obtain sample processing quality parameters; updating the control weights of the target backpropagation neural network until the training stop condition is reached; the processing quality parameters include laser ablation depth, laser ablation surface roughness, and boss top circle diameter.

2. The laser processing quality prediction method based on the improved starfish optimization algorithm as described in claim 1, characterized in that, Obtain the process parameters for silicon carbide laser processing, including: The original parameters for silicon carbide laser processing are obtained, and the original parameters are cleaned to obtain candidate parameters. The candidate parameters are normalized to obtain the process parameters.

3. The laser processing quality prediction method based on the improved starfish optimization algorithm as described in claim 1, characterized in that, The training method for the target backpropagation neural network includes: Obtain sample process parameters for silicon carbide laser processing; The sample process parameters are input into the target backpropagation neural network to obtain the sample processing quality parameters; Update the control weights of the target backpropagation neural network until the training stopping condition is met.

4. A laser processing quality prediction device based on an improved starfish optimization algorithm, characterized in that, The device includes: The algorithm optimization module is used to improve the starfish optimization algorithm based on the Levy flight principle to obtain an improved starfish optimization algorithm. Specifically, it includes: introducing Levy flight into the breeding behavior of the starfish optimization algorithm to optimize the global exploration formula of the starfish optimization algorithm; introducing Levy flight perturbation into the predation behavior of the starfish optimization algorithm; and adaptively adjusting the step size based on the Levy flight step size. The network improvement module is used to optimize the weights and biases of the backpropagation neural network based on the improved starfish algorithm, and to suppress overfitting of the backpropagation neural network based on the L1 regularization method to obtain the target backpropagation neural network. Specifically, it includes: determining the fitness function of the backpropagation neural network based on the L1 regularization method; determining the optimal parameters of the backpropagation neural network based on the improved starfish algorithm to obtain the target backpropagation neural network; wherein, determining the optimal parameters of the backpropagation neural network based on the improved starfish algorithm includes: iterating through each individual using the improved starfish algorithm to obtain the fitness of each individual; retaining the individual with the highest fitness, and selecting some individuals to retain from the remaining individuals, until the iteration reaches a preset condition to determine the optimal parameters of the backpropagation neural network; and substituting the optimal parameters into the backpropagation neural network to obtain the target backpropagation neural network. The parameter acquisition module is used to acquire the process parameters of silicon carbide laser processing. Specifically, it includes: acquiring the original parameters of silicon carbide laser processing; using the 3σ criterion to remove outliers in each feature dimension; for missing values, selecting mean-filling or nearest-neighbor filling based on dataset characteristics; simultaneously filtering input features strongly correlated with the output target using the Pearson correlation coefficient; removing redundant features to simplify the network structure and obtaining candidate parameters; and using the Min-Max normalization method to map the candidate parameters to the [0, 1] interval to obtain the process parameters. The process parameters include laser processing frequency, laser processing power, number of processing passes, scanning overlap rate, and spot overlap rate. The formula for the Min-Max normalization method is: ； Where P is a candidate parameter. These are process parameters; A quality prediction module is used to input the process parameters into a pre-trained target backpropagation neural network to obtain the processing quality parameters of the silicon carbide laser processing; wherein the training method of the target backpropagation neural network includes: obtaining sample process parameters of silicon carbide laser processing; inputting the sample process parameters into the target backpropagation neural network to obtain sample processing quality parameters; updating the control weights of the target backpropagation neural network until the training stopping condition is reached; the processing quality parameters include laser ablation depth, laser ablation surface roughness, and boss top circle diameter.

5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the laser processing quality prediction method based on the improved starfish optimization algorithm as described in any one of claims 1 to 3.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the laser processing quality prediction method based on the improved starfish optimization algorithm as described in any one of claims 1 to 3.

7. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the laser processing quality prediction method based on the improved starfish optimization algorithm as described in any one of claims 1 to 3.

Images