Intelligent Control System for Shot Peening Process Parameters Based on Deep Reinforcement Learning

The intelligent control system for shot peening process parameters, developed through deep reinforcement learning, solves the data misalignment problem caused by fixture resonance and fluid time delay in the shot peening control system, thereby improving the surface quality of the workpiece and ensuring stable system operation.

CN123113777APending Publication Date: 2026-11-17SHANGHAI PEENTECH EQUIP TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611530374.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-30
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing shot peening control systems suffer from data alignment misalignment caused by mechanical resonance of the fixture and fluid time delay, leading to microcracks on the workpiece surface and interruption of control system operation.

Method used

An intelligent control system for shot peening process parameters based on deep reinforcement learning is adopted. Through frequency domain filtering unit, timing alignment unit, action probability generation unit and action masking unit, acoustic signals, fluid response time delay and internal cavity three-dimensional mesh topology model are processed respectively to achieve timing alignment of action commands and overload protection.

Benefits of technology

It effectively filters out misidentification of mechanical vibration signals, compensates for fluid transmission delays, prevents dense particle rebound and control system interruption, and ensures workpiece surface quality.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123113777A_ABST
    Figure CN123113777A_ABST
Patent Text Reader

Abstract

This invention relates to the field of metal surface treatment and intelligent control technology, and discloses an intelligent control system for shot peening process parameters based on deep reinforcement learning. The system uses a frequency domain filtering unit to perform band-stop filtering on the surface acoustic signal according to the workpiece's resonant frequency band; a timing alignment unit extracts the historical three-dimensional pose using the fluid response time delay constant and concatenates it with acoustic features to generate a synchronization state matrix; an action probability generation unit outputs initial probabilities through a policy network; and an action masking unit performs ray tracing within the three-dimensional mesh topology model of the cavity. When the ray density exponent exceeds a threshold, it generates a mask weight to attenuate the probability of the target action command and resamples it. This invention eliminates mechanical vibration interference and fluid time delay misalignment, cuts off the overload action of dense particle rebound in the cavity, and avoids the initiation of microcracks in the workpiece and system interruption.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of metal surface treatment and intelligent control technology, and in particular to an intelligent control system for shot peening process parameters based on deep reinforcement learning. Background Technology

[0002] In shot peening surface strengthening processes for thin-walled workpieces such as cup-shaped flexible wheels, surface acoustic signals are typically used to monitor the workpiece's plastic deformation state. In actual working scenarios, the mechanical motion of multi-axis robots and tooling fixtures generate low-frequency mechanical vibrations. These vibrations are transmitted to the workpiece through physical contact and interfere with the acoustic sensors. Because these vibrations are not physically isolated, existing control systems, when acquiring acoustic data, easily misidentify the mechanical resonance phenomenon caused by fixture constraints as actual microscopic plastic deformation of the workpiece surface, leading to perceptual bias in the extracted initial acoustic features.

[0003] The aforementioned biased initial acoustic features are further affected by fluid physical delays when transmitted to the control system for state alignment. Due to the inertia of compressed gas transmission in the pneumatic pipeline, there is a physical time difference between the establishment of shot peening pressure and acoustic feedback. Existing control systems typically rely on transient clocks for data stitching, directly aligning the acoustic features acquired at the current moment with the execution action issued at the current moment, resulting in a temporal and causal misalignment of the state data pairs input to the reinforcement learning network.

[0004] Data pairs with perceptual bias and timing misalignment mask the physical overload risk within enclosed spaces when the driving action strategy network generates control commands. When the spray gun enters the inner cavity or inner radius region of the workpiece, the sprayed particles will experience dense secondary and multiple rebounds within the confined space. Due to the preceding data misalignment, the strategy network cannot predict such local overloads in time, and the system cannot cut off or reduce the shot peening pressure in time before the deformation exceeds the limit, leading to the initiation of microcracks on the surface of thin-walled workpieces. At the same time, under extreme overload conditions, if the control algorithm lacks underlying deadlock prevention logic, the sharp fluctuations in the probability distribution can easily cause division by zero errors, leading to controller interruption. Summary of the Invention

[0005] This invention provides an intelligent control system for shot peening process parameters based on deep reinforcement learning, which solves the technical problem in existing shot peening control systems where data alignment misalignment is caused by mechanical resonance of the fixture and fluid time delay, resulting in the inability of the action strategy network to intervene in the dense rebound of particles in the inner cavity of the workpiece in a timely manner, ultimately causing microcracks to appear on the workpiece surface and interrupting the operation of the control system.

[0006] In view of the above problems, the present invention provides an intelligent control system for shot peening process parameters based on deep reinforcement learning, the system comprising: The frequency domain filtering unit is configured to acquire the first resonant frequency band of the target workpiece under constrained conditions, use the first resonant frequency band as the cutoff frequency band parameter of the band-stop filter, perform frequency domain filtering on the acquired original surface acoustic signal sequence, and generate the target acoustic feature vector. The timing alignment unit is configured to acquire the response time delay constant of the fluid jet terminal, determine the target historical moment based on the response time delay constant, extract the spatial three-dimensional pose of the target historical moment from the historical motion trajectory cache of the spatial attitude actuator, and concatenate the spatial three-dimensional pose with the target acoustic feature vector to generate a synchronization state space matrix. The action probability generation unit is configured to input the synchronization state space matrix into a preset policy network and output an initial action probability distribution tensor, wherein the initial action probability distribution tensor includes spatial motion feature parameters and hydrodynamic field parameters. The action masking unit is configured to acquire the internal three-dimensional mesh topology model of the target workpiece, extract the target action command whose probability value is greater than a preset first threshold from the initial action probability distribution tensor; perform ray tracing calculation in the internal three-dimensional mesh topology model according to the parameters of the target action command to generate a ray clustering density index; if the ray clustering density index is greater than a second threshold, generate mask weights based on the ray clustering density index, update the probability value of the target action command in the initial action probability distribution tensor through the mask weights, and resample and output the updated action command sequence.

[0007] The technical solution provided in this application has at least the following technical effects: By extracting the first N natural resonant frequencies of the target workpiece under constrained conditions as the cutoff frequency parameters of the band-stop filter, the original surface acoustic signal sequence is frequency-domain filtered to remove the low-frequency resonance amplitudes transmitted to the workpiece by the clamping of the fixture and the movement of the robotic arm, thus avoiding misidentification of mechanical vibration signals as plastic deformation features of the workpiece surface.

[0008] By obtaining the response time delay constant of the fluid injection terminal, the difference between the current timestamp and the constant is calculated to obtain the target historical moment. The spatial three-dimensional pose of the historical moment is then tensor-stitched with the current target acoustic feature vector to compensate for the transmission time difference of compressed gas in the pneumatic pipeline, thus achieving time alignment between the actuator's motion trajectory and the surface acoustic feedback data.

[0009] By performing ray tracing calculations within the three-dimensional mesh topology model of the inner cavity, when the ray clustering density index is greater than the second threshold, a mask weight is generated using an exponential formula containing a minimum probability floor constant to attenuate the probability of the target action command. A zero-prevention constant is set in the denominator of the tensor normalization formula. This probability attenuation cuts off overload operation actions in the inner cavity region that are prone to particle dense rebound. At the same time, the zero-prevention constant in the denominator prevents the interruption of the control system operation caused by the weight returning to zero. Attached Figure Description

[0010] Figure 1 A flowchart of an intelligent control method for shot peening process parameters based on deep reinforcement learning provided in an embodiment of the present invention; Figure 2 The diagram shows the structure of the intelligent control system for shot peening process parameters based on deep reinforcement learning, as provided in an embodiment of the present invention. Detailed Implementation

[0011] The above technical solutions will now be described in detail with reference to the accompanying drawings and specific embodiments to provide a better understanding of them. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. It should be understood that the present invention is not limited to the exemplary embodiments used only to explain the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. Furthermore, it should be noted that, for ease of description, only the parts related to the present invention are shown in the drawings, not all of them.

[0012] Example: Please refer to Figure 1 and Figure 2 The intelligent control system for shot peening process parameters based on deep reinforcement learning constructs a decoupled operating topology using an execution layer and an edge control layer. The execution layer includes a fixture module, a spatial attitude actuator, and a fluid injection terminal. The fixture module supports and fixes the target workpiece, which is instantiated as a cup-shaped flexible wheel. The fixture module includes a tooling base and an inner hole pressure cap. The spatial attitude actuator is instantiated as a multi-axis robot. The edge control layer deploys an industrial control computer and a programmable logic controller. The industrial control computer is connected to the execution layer via a high-speed industrial Ethernet bus. Upon initiation of the shot peening operation, the industrial control computer sends a Network Time Protocol (NTP) clock synchronization signal. The acoustic sensor acquisition card receives the clock synchronization signal and initiates analog-to-digital conversion. Simultaneously, the servo driver of the spatial attitude actuator receives the clock synchronization signal, locks the position loop refresh cycle, and establishes the data interaction clock tree between the acoustic sensor acquisition card and the servo driver.

[0013] After establishing the data interaction clock tree, the control system sequentially triggers the execution pipeline of four major units. The first step is executed by the frequency domain filtering unit, which receives the original surface acoustic signal sequence and performs feature separation based on the first resonant frequency band, shielding the low-frequency amplitude superposition constraint caused by the target workpiece structural resonance. The second step is executed by the timing alignment unit, which extracts the hydrodynamic response time delay constant of the fluid jet terminal, compensates for the time offset constraint between the spatial trajectory and the flow field, and generates a synchronized state space matrix. The third step is executed by the action probability generation unit, which processes the synchronized state space matrix and outputs the initial action probability distribution tensor. The fourth step is executed by the action masking unit, which uses ray tracing to calculate the ray clustering density index, masks the initial action probability distribution tensor, and truncates the action command sequence that triggers microcracks.

[0014] The specific process for the frequency domain filtering unit to obtain the first resonant frequency band of the target workpiece under constrained conditions involves the frequency domain filtering unit reading the material stiffness matrix containing Young's modulus and Poisson's ratio parameters, and simultaneously extracting the material density physical quantity of the target workpiece. Based on the material density physical quantity and the volume fraction of the three-dimensional geometric topology model, the frequency domain filtering unit generates a mass matrix that participates in the calculation of the dynamic partial differential equations. The frequency domain filtering unit extracts the set of contact node coordinates in the three-dimensional geometric topology model that spatially coincide with the tooling base and the inner hole pressure cap, and applies zero-degree-of-freedom fixed constraint boundary conditions to the contact node coordinate set.

[0015] After the zero-DOF fixed constraint boundary conditions are applied, the frequency domain filtering unit calls the Lanzos algorithm to solve the dynamic partial differential equations based on the target workpiece's mass matrix and material stiffness matrix. The Lanzos algorithm executes the eigenvalue buckling extraction operator, outputting the first N natural resonant frequencies of the target workpiece under the clamping force boundary conditions, where N is a positive integer. The frequency domain filtering unit aggregates the first N natural resonant frequencies into frequency bands, generating a continuous frequency range with a lower limit of 0Hz and an upper limit covering the highest N natural resonant frequencies. The frequency domain filtering unit writes this continuous frequency range into memory and sets it as the first resonant frequency band.

[0016] As the first resonant frequency band is generated and written into memory, the frequency domain filtering unit configures the discrete transfer function of the band-stop filter based on the first resonant frequency band. The band-stop filter adopts a Butterworth digital notch filter architecture. The lower and upper limits of the first resonant frequency band are input into the cutoff frequency parameters of the band-stop filter, and the quality factor of the band-stop filter is set to a positive real number that controls the attenuation steepness of the stopband. When the original surface acoustic signal sequence is input into the frequency domain filtering unit, the unit opens a sliding window of a preset length to extract the original surface acoustic signal sequence and performs a Fast Fourier Transform (FFT) on the original surface acoustic signal sequence within the sliding window. The FFT outputs a complex sequence in the frequency domain. The band-stop filter applies the discrete transfer function to perform a frequency domain multiplication operation with the complex sequence to eliminate the resonance peak amplitude located within the first resonant frequency band. The frequency domain filtering unit performs an inverse FFT on the result of the multiplication operation to restore the time-domain signal. A root mean square integral is performed on the restored time-domain signal, and a single scalar value is output. The frequency domain filtering unit pushes the single scalar values ​​output in multiple consecutive control clock cycles into the shift register in chronological order according to the timestamps, and combines them to generate a target acoustic feature vector with a fixed dimension.

[0017] Along with the generation of the target acoustic feature vector, the timing alignment unit performs the operation of extracting the response time delay constant of the fluid jet terminal. The calibration process of the response time delay constant is triggered by the timing alignment unit sending a pressure step control signal to the pneumatic proportional valve controlling the fluid jet terminal. The timing alignment unit uses the pressure sensor at the front end of the pipeline to measure the time interval consumed by the fluid medium to reach the fluid jet terminal and establish a stable flow field pressure. The time interval value is stored in the non-volatile memory by the timing alignment unit as the response time delay constant.

[0018] The timing alignment unit reads the instantaneous time value of the signal synchronization trigger from the local clock controller and determines it as the current timestamp of the current control cycle. The timing alignment unit obtains the single target acoustic feature vector generated at the current timestamp. The timing alignment unit subtracts the response delay constant value from the current timestamp to generate the target's historical moment occurring in the past. The historical motion trajectory cache of the spatial attitude actuator is configured as a first-in-first-out circular queue arranged in memory. The timing alignment unit uses the target's historical moment as a memory address pointer to read the multiple servo axis joint angles recorded corresponding to the target's historical moment from the historical motion trajectory cache. The timing alignment unit extracts the link length parameters and link offset constants of the spatial attitude actuator and constructs a forward kinematics transfer matrix by combining them with a preset DH parameter model. The timing alignment unit substitutes the multiple servo axis joint angles into the forward kinematics transfer matrix for matrix multiplication calculation and outputs the spatial three-dimensional pose corresponding to the target's historical moment. The spatial three-dimensional pose includes three-dimensional Cartesian coordinates defining the jet space coordinates and Euler angles defining the jet attitude.

[0019] After the spatial 3D pose is output, the temporal alignment unit combines the spatial 3D pose with the target acoustic feature vector through tensor splicing operations. During the execution cycle of the tensor splicing operation instruction, the temporal alignment unit appends and arranges the spatial 3D pose sequence containing 6 floating-point dimensions to the end of the target acoustic feature vector containing a preset number of dimensions at the feature dimension level, generating a 1-dimensional single-row synchronization state space matrix in memory, and completing the phase mapping binding between the spatial execution trajectory and the surface acoustic sampling data.

[0020] The policy network nodes deployed within the action probability generation unit are configured as a multilayer perceptron architecture comprising an input layer, hidden layers, and an output layer. The synchronized state space matrix is ​​input as a feature vector to the input layer. Inside the hidden layer, a dot product operation is performed with the pre-trained weight matrix, and a bias term constant is added. The result is then passed through a linear rectified activation function to output activation values. These activation values ​​are passed to the output layer, which calculates and outputs discrete probability distribution vectors corresponding to each independent action within a predefined discrete action space based on a fully connected mapping operator and a normalized exponential function. The predefined discrete action space is formatted in memory as a key-value pair mapping dictionary. This dictionary encodes and binds each discrete action index to a specific set of spatial motion feature parameters and hydrodynamic field parameters. The action probability generation unit formats the discrete probability distribution vectors in memory into an initial action probability distribution tensor.

[0021] The action masking unit performs a tensor polling comparison operation to extract target action commands whose floating-point probability values ​​are greater than a preset first threshold from the initial action probability distribution tensor. Based on the index of the target action command, the action masking unit performs an address lookup in the mapping dictionary to extract the corresponding spatial motion feature parameters and hydrodynamic field parameters. The spatial motion feature parameters include three-dimensional spatial coordinates, incident elevation angle, and yaw angle. The action masking unit assigns the three-dimensional spatial coordinates as the origin vector of the ray projection, representing the discrete particle. The action masking unit substitutes the incident elevation angle and yaw angle into the mathematical transformation matrix from spherical to Cartesian coordinates, performs spatial trigonometric function calculations, and outputs the ray projection direction vector in the three-dimensional spatial coordinate system. The ray projection origin vector and the ray projection direction vector are aggregated in the buffer to form the initial emission parameters.

[0022] After initial launch parameter initialization, the action masking unit loads the internal 3D mesh topology model of the target workpiece into the dynamic random access memory. Using the origin vector of the ray projection as the starting point of spatial coordinates, the action masking unit performs a forward ray projection simulation within the internal 3D mesh topology model along the ray projection direction vector. The action masking unit calls the Möller-Trumbore cross-detection algorithm operator to iteratively calculate the collision intersection coordinates between the ray equation of the projected ray and the equations of each spatial triangular facet within the internal 3D mesh topology model. The action masking unit reads the unit normal vector of the colliding spatial triangular facet at the collision intersection coordinates. The action masking unit extracts the ray projection direction vector and the unit normal vector, performs a dot product operation on them, multiplies the dot product value by a scaling factor of -2, and then performs a vector addition operation with the original ray projection direction vector, outputting a spatial reflection trajectory representing the local ejection trajectory of the microscopic particle.

[0023] Based on the coordinate distribution set of spatial reflection trajectories, the action masking unit discretizes the continuous three-dimensional volume space contained in the internal cavity three-dimensional mesh topology model into several three-dimensional voxel meshes with preset volumes. To compress the memory indexing overhead of the three-dimensional voxel meshes, the spatial nesting data structure of the three-dimensional voxel meshes is set as an octree structure. The action masking unit uses an octree node depth-first traversal algorithm to scan the spatial reflection trajectory distribution set and counts the number of intersections of spatial reflection trajectories that collide two or more times within the boundary of each three-dimensional voxel mesh. The action masking unit divides the number of intersections within a single three-dimensional voxel mesh by the preset volume value of the corresponding three-dimensional voxel mesh to generate a ray clustering density index with dimensions per cubic millimeter.

[0024] With the generation of the ray clustering density index, the control pipeline enters the parameter calibration and boundary defense stage. The action mask unit retrieves the second threshold pre-stored in memory and performs a numerical comparison operation. The second threshold is pre-calibrated by the external material analysis module and has a value greater than 0. The external material analysis module, in conjunction with the Mises yield criterion, reverse-engineers the limit value of the cumulative plastic work done before the target workpiece's lattice fractures, converts the limit value of the cumulative plastic work done into the equivalent collision frequency value of the dense collisions of discrete particles, and assigns it to the second threshold.

[0025] During the calculation of the generated mask weights, the action mask unit reads the attenuation steepness constant to control the gradient of the penalty formula. When the target workpiece is manufactured using a material with fracture toughness parameters lower than the engineering preset benchmark, the external configuration software writes an attenuation steepness constant of 20 into the register of the action mask unit. A constant of 20 is greater than 0, causing the mask weights to attenuate towards zero after exceeding the limit.

[0026] Before performing tensor calculations, the action masking unit constructs a full-one matrix in memory with the same dimensions as the initial action probability distribution tensor, serving as the initial skeleton of the penalty coefficient tensor. For action commands that do not trigger ray tracing calculations due to probability values ​​being less than or equal to the first threshold, the action masking unit directly locks the corresponding penalty coefficient to a constant floating-point number of 1 in the full-one matrix. For target action commands that trigger ray tracing calculations, if the generated ray clustering density index is less than or equal to the second threshold, the action masking unit determines that the fluid jetting terminal is currently in an open area of ​​the target workpiece topology, and similarly locks the corresponding penalty coefficient to a constant floating-point number of 1.

[0027] When the fluid jetting terminal is controlled to move into the enclosed inner cavity of the target workpiece, it triggers dense rebounds, driving a surge in the ray clustering density index. The action mask unit substitutes the ray clustering density index exceeding the second threshold into the penalty formula. The formula is as follows:

[0028] in, The penalty coefficient is... The ray clustering density index, This is the lowest probability support constant; The second threshold is greater than 0; The attenuation steepness constant is preset and is greater than 0; the squared term amplifies as the input value increases, and combined with the externally nested negative natural exponent operator, the penalty coefficient converges to approach the minimum probability floor constant. The action mask unit, based on the position index of the target action command, overwrites the calculated penalty coefficient into the initial skeleton at fixed points, generating complete tensor mask weights. This prevents gradient vanishing caused by the penalty coefficient returning to zero.

[0029] Subsequently, the action mask unit extracts the zero constant and injects it into the denominator matrix of the tensor normalized probability distribution formula, as follows:

[0030] in, For the updated action probability distribution tensor, Let be the initial action probability distribution tensor. This represents the element-wise multiplication operation of tensors. To prevent division by zero constants that are greater than 0, under overload conditions, the presence of these constants in the denominator prevents division-by-zero errors that could cause system interruptions. After the system interruption is prevented, the algebraic division operator amplifies the update probability value of the non-target action command representing the tool retraction action. For the resampled output process, the action mask unit is based on the updated action probability distribution tensor. The multinomial distribution sampling operator is invoked to extract discrete action indices, and resampling is used to generate an instruction sequence containing safety action instructions, which is then sent to the space attitude actuator to drive the space attitude actuator to execute a trajectory to escape the danger zone.

[0031] When the control system is deployed on a computing node with limited computing power, the action masking unit triggers a linearly truncated control path with bottom line protection. The action masking unit retrieves a preset linear truncation rate constant and a minimum probability bottoming constant from non-volatile memory. It calculates the algebraic difference between the ray clustering density index and a second threshold, filtering out negative differences. The action masking unit divides the algebraic difference by the second threshold, multiplies it by the linear truncation rate constant to obtain an attenuation deduction term, and subtracts the attenuation deduction term from the constant floating-point number 1 to obtain the linear attenuation value. The action masking unit then executes a maximum value comparison operation instruction, extracting the maximum value between the minimum probability bottoming constant and the linear attenuation value and assigning it to the mask weights. The presence of the minimum probability bottoming constant in the underlying registers prevents the backpropagation gradient vanishing fault caused by the mask weights returning to zero.

[0032] The motion masking unit executes a voxel heatmap-based graphical substitution path. An external cloud server pre-extracts the 3D mesh topology model of the target workpiece's internal cavity and traverses the possible initial parameters of the emission sources of the spatial attitude actuator, performing offline global ray tracing calculations. The external cloud server binds the ray clustering density index output from the offline calculation to the corresponding initial parameters of the emission sources using key-value pairs, generating a pose-density 3D lookup table. This pose-density 3D lookup table is burned into the read-only memory (ROM) of the edge computing node's motherboard in binary file format. During the runtime cycle, the motion masking unit parses the target motion command to extract the initial parameters of the emission sources. Using these initial parameters as memory address pointers, the motion masking unit executes low-level lookup instructions in the ROM, reading the ray clustering density index within the corresponding address segment, completing the data flow that replaces computation time with storage space.

[0033] The action masking unit extracts a logistic function with a safe offset as an equivalent replacement operator for the penalty control logic. When the input ray clustering density exponent exceeds the second threshold, the logistic function with a safe offset directly maps to the smoothly transitioning mask weight floating-point numbers at the output. The logistic function with a safe offset maintains the global gradient's continuous differentiability across the entire domain. This continuous differentiability ensures that the action probability generation unit successfully performs calculus backpropagation when receiving feedback rewards, avoiding the interruption of the update data stream of the policy network neuron weight matrix at non-differentiable breakpoints.

[0034] The intelligent control system for shot peening process parameters based on deep reinforcement learning operates on a hardware base containing silicon-based electronic components. The hardware base is equipped with a central processing unit (CPU), a dedicated graphics coprocessor unit for parallel execution of ray tracing simulation and policy network tensor multiplication operations, and a high-speed dynamic random access memory (DRAM) providing data throughput buffer space. The mainboard of the hardware base connects to an industrial Ethernet communication interface card via an external device interconnect standard expansion slot. The industrial Ethernet communication interface card is connected to the control cabinet base plate of the spatial attitude actuator via shielded twisted-pair cables and synchronously connected to the pressure proportional valve of the fluid injection terminal. The sequence of action commands is converted into differential voltage signals by the chip of the industrial Ethernet communication interface card, and the differential voltage signals are transmitted and received from the main control unit to the actuators via shielded twisted-pair cables.

[0035] The frequency domain filtering unit, timing alignment unit, action probability generation unit, and action masking unit configured within the system are instantiated at the hardware level as a machine-readable instruction set embedded in non-volatile memory. When the machine-readable instruction set is loaded from non-volatile memory to high-speed dynamic random access memory (DRAM) by the CPU and activated for execution, the assembly-level machine code contained in the machine-readable instruction set drives transistor-level gate state transitions in the arithmetic logic units within the CPU. These gate state transitions, synchronized with the system clock ticks, continuously rewrite the charge polarity distribution of a specified memory block in DRAM, completing the transformation from data tensor to machine memory state.

[0036] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deep reinforcement learning based intelligent control system for shot peening process parameters, characterized in that, The system includes: The frequency domain filtering unit is configured to acquire the first resonant frequency band of the target workpiece under constrained conditions, use the first resonant frequency band as the cutoff frequency band parameter of the band-stop filter, perform frequency domain filtering on the acquired original surface acoustic signal sequence, and generate the target acoustic feature vector. The timing alignment unit is configured to acquire the response time delay constant of the fluid jet terminal, determine the target historical moment based on the response time delay constant, extract the spatial three-dimensional pose of the target historical moment from the historical motion trajectory cache of the spatial attitude actuator, and concatenate the spatial three-dimensional pose with the target acoustic feature vector to generate a synchronization state space matrix. The action probability generation unit is configured to input the synchronization state space matrix into a preset policy network and output an initial action probability distribution tensor, wherein the initial action probability distribution tensor includes spatial motion feature parameters and hydrodynamic field parameters. The action masking unit is configured to acquire the internal three-dimensional mesh topology model of the target workpiece, extract the target action command whose probability value is greater than a preset first threshold from the initial action probability distribution tensor; perform ray tracing calculation in the internal three-dimensional mesh topology model according to the parameters of the target action command to generate a ray clustering density index; if the ray clustering density index is greater than a second threshold, generate mask weights based on the ray clustering density index, update the probability value of the target action command in the initial action probability distribution tensor through the mask weights, and resample and output the updated action command sequence.

2. The deep reinforcement learning based intelligent shot process parameter control system of claim 1, wherein, The specific configuration of the first resonant frequency band of the target workpiece under constrained state obtained by the frequency domain filtering unit includes: Obtain the three-dimensional geometric topology model and material stiffness matrix of the target workpiece; Calculate the first N natural resonant frequencies of the target workpiece under the preset constraint boundary conditions, where N is a positive integer; The first resonant frequency band is generated based on the first N natural resonant frequencies.

3. The deep reinforcement learning based intelligent shot process parameter control system of claim 1, wherein, The specific configuration for the timing alignment unit to generate the synchronization state space matrix includes: Determine the current timestamp of the current control cycle, and obtain the target acoustic feature vector generated under the current timestamp; Subtract the response delay constant from the current timestamp to obtain the target historical moment; Retrieve the spatial three-dimensional pose recorded at the target's historical moment from the historical motion trajectory cache; The synchronization state space matrix is ​​generated by combining the spatial three-dimensional pose with the target acoustic feature vector through tensor splicing operations.

4. The deep reinforcement learning based intelligent shot process parameter control system of claim 1, wherein, The specific configuration for the action mask unit to perform ray tracing calculations and generate ray clustering density indices includes: The target action command is analyzed, and the corresponding spatial motion feature parameters and hydrodynamic field parameters are extracted as the initial emission parameters of discrete particles; The spatial reflection trajectory of the discrete particle is calculated within the three-dimensional mesh topology model of the inner cavity based on the initial emission parameters; The internal cavity three-dimensional mesh topology model is divided into three-dimensional voxel meshes of a preset volume, and the number of intersections of the spatial reflection trajectories that undergo two or more reflections within each three-dimensional voxel mesh is counted. The ratio of the number of intersections to the preset volume is used as the ray clustering density index of the corresponding three-dimensional voxel mesh.

5. The deep reinforcement learning based shot peening process parameter intelligent control system of claim 4, wherein, The specific configuration of the action masking unit for generating mask weights based on the ray clustering density index includes: When the ray clustering density index is less than or equal to the second threshold, the mask weight is set to a constant preservation coefficient; When the ray clustering density index is greater than the second threshold, the difference between the ray clustering density index and the second threshold is obtained, a monotonically decreasing penalty coefficient is calculated based on the difference, and the penalty coefficient is used as the mask weight; wherein, the larger the difference, the closer the penalty coefficient is to the preset minimum probability support constant, and the minimum probability support constant is greater than or equal to 0 and less than the constant preservation coefficient.

6. The deep reinforcement learning based intelligent shot process parameter control system of claim 5, wherein, The specific configuration for calculating the monotonically decreasing penalty coefficient by the action mask unit is as follows: The penalty coefficient is calculated according to the following formula : ; wherein, is the penalty coefficient, is the ray cluster density index, is the lowest probability floor constant; is the second threshold value and the second threshold value is greater than 0; is a preset attenuation steepness constant and the attenuation steepness constant is greater than 0; the action mask unit updates the specific configuration of the probability value of the target action instruction in the initial action probability distribution tensor through the mask weight as follows: According to the formula ; computing an updated action probability distribution tensor ; wherein, is the initial action probability distribution tensor, denotes a tensor element-wise multiplication operation, is a constant preventing zero and greater than 0.

7. The deep reinforcement learning based intelligent shot process parameter control system of claim 5, wherein, The specific configuration for calculating the monotonically decreasing penalty coefficient in the action mask unit is replaced as follows: Calculate the difference between the ray clustering density index and the second threshold, and multiply the ratio of the difference to the second threshold by a preset linear cutoff rate constant to obtain the attenuation deduction term; Subtracting the attenuation deduction term from the constant retention coefficient yields the linear attenuation value; Extract the maximum value between the linear decay value and the minimum probability bottoming constant, and assign it as the penalty coefficient.