Intelligent control method, system and equipment of intelligent stamping machine and medium

Abstract

The invention relates to an intelligent control method, system and equipment of an intelligent stamping machine and a medium. The method comprises the following steps: acquiring environment state data and target boundary data of a current stamping task; obtaining a state feature vector based on the environment state data and the target boundary data; the state feature vectors are input into a pre-trained reinforcement learning decision model, a plurality of adjustable parameters of an infrared sensing system on the stamping machine are subjected to joint optimization decision making through the reinforcement learning decision model, and a current optimal parameter combination is output; configuring working parameters of an infrared sensing system according to the current optimal parameter combination, and driving the infrared sensing system to scan and detect the boundary of the paper to obtain actual boundary information of the paper; and obtaining a stamping instruction based on the actual boundary information and the target boundary data. By adopting the method, the problems of poor environmental adaptability, insufficient boundary detection precision and potential safety hazards of stamping in an unauthorized area caused by parameter solidification of a traditional stamping machine can be solved, and the accuracy and safety of stamping are improved.

Description

Technical Field

[0001] This invention belongs to the field of stamping technology, and in particular relates to an intelligent control method, system, device and medium for an intelligent stamping machine. Background Technology

[0002] With the continuous development of office automation technology, intelligent stamping devices are playing an increasingly important role in improving stamping efficiency and security. These devices typically use infrared sensors around the stamping area to detect the paper's position. Stamping is only allowed when the paper is completely within the preset stamping area, thus preventing accidental stamping in unauthorized areas.

[0003] In traditional technology, to adapt to different paper sizes (such as A3, A4, and A5), a common solution is to use infrared sensors with fixed positions or parameters. When using the device, the user selects the paper size option, and the device will call up a set of preset infrared sensor parameters to match the corresponding printing range.

[0004] However, this current approach has significant limitations. Because the parameters of the infrared sensor (such as emission power) are pre-defined, they cannot be dynamically adjusted based on real-time ambient light, paper characteristics (such as thickness and transparency), or minor performance drift of the sensor itself. This makes the sensor signal susceptible to interference in practical applications, potentially leading to misjudgments. For example, in boundary areas, signal coupling or attenuation might cause compliant paper to be mistakenly identified as exceeding the boundary and stamped, or a slight paper offset might be missed, resulting in stamping in an unexpected location, posing a security risk. Summary of the Invention

[0005] Therefore, it is necessary to provide an intelligent control method, system, device, and medium for an intelligent stamping machine to address the aforementioned technical problems.

[0006] Firstly, this application provides an intelligent control method for an intelligent stamping machine, including:

[0007] S1. Obtain the environmental status data and target boundary data of the current stamping task;

[0008] S2. Based on environmental state data and target boundary data, obtain the state feature vector; the state feature vector is used to characterize the current working condition of the stamping machine;

[0009] S3. Input the state feature vector into the pre-trained reinforcement learning decision model, and use the reinforcement learning decision model to jointly optimize multiple adjustable parameters of the infrared sensing system on the stamping machine, and output the current optimal parameter combination.

[0010] S4. Configure the working parameters of the infrared sensing system according to the current optimal parameter combination, drive the infrared sensing system to scan and detect the paper boundary, and obtain the actual boundary information of the paper.

[0011] S5. Based on the actual boundary information and target boundary data, obtain the stamping instruction; the stamping instruction is used to instruct the stamping machine to perform the stamping operation and obtain the stamping result.

[0012] In one embodiment, acquiring the environmental state data and target boundary data of the current stamping task includes:

[0013] S11. Control all infrared transmitters in the infrared sensing system to transmit signals at a reference power, and synchronously acquire the raw signal values ​​of all infrared receivers to obtain the reference signal vector.

[0014] S12. Based on the reference signal vector and the pre-calibrated no-load environment reference signal vector, calculate the signal attenuation coefficient vector and background noise intensity vector under the current environment;

[0015] S13. Read the historical task log of the stamping machine, obtain the final decision results and boundary error statistics of historical tasks similar to the current target boundary data, and use them as historical performance reference data.

[0016] S14. Combine the signal attenuation coefficient vector, background noise intensity vector, and historical performance reference data to form environmental status data.

[0017] In one embodiment, a state feature vector is obtained based on environmental state data and target boundary data, including:

[0018] S21. Normalize the environmental state data to obtain a normalized environmental vector;

[0019] S22. Encode and transform the target boundary data to obtain the boundary encoding vector;

[0020] S23. Based on the normalized environment vector and the boundary coding vector, feature fusion is performed to obtain the fused feature vector;

[0021] S24. The dimensionality of the fused feature vector is transformed by a fully connected layer to obtain the state feature vector.

[0022] In one embodiment, the pre-training process of the reinforcement learning decision model includes:

[0023] S31. Construct the state space of the reinforcement learning decision model. The state vector of the state space includes the state feature vector, the boundary detection error of the previous cycle in real time feedback, and the signal-to-noise ratio of the sensor signal.

[0024] S32. Construct the action space of the reinforcement learning decision model. The action vector of the action space is defined as the adjustment amount of multiple adjustable parameters in the infrared sensing system. The adjustable parameters include the transmission power of each infrared transmitter, the gain threshold of each infrared receiver, and the signal sampling frequency.

[0025] S33. Define the reward function of the reinforcement learning decision model. The reward function calculates the instant reward based on the degree of agreement between the actual boundary information generated after the action is executed and the target boundary data, as well as the sensor resources consumed to achieve the agreement.

[0026] S34. Use the proximal policy optimization algorithm to train the reinforcement learning decision model offline until the model can stably output parameter adjustment actions that maximize the reward function in response to simulated state inputs, thus completing the pre-training.

[0027] In one embodiment, the operating parameters of the infrared sensing system are configured according to the current optimal parameter combination, driving the infrared sensing system to scan and detect the paper boundary to obtain the actual boundary information of the paper, including:

[0028] S41. Based on the current optimal parameter combination, set the modulation frequency and phase of each infrared transmitter in the infrared sensing system to transmit pulse signals in a time-division manner according to a preset time sequence;

[0029] S42. During each transmission pulse cycle, the response signal strength and transit time of the corresponding infrared receiver are synchronously acquired to obtain multi-channel signal strength and time series data.

[0030] S43. Filter and perform time-frequency analysis on the multi-channel signal strength and time series data to calculate the occlusion position of the paper edge on each independent infrared detection path and obtain the occlusion calculation result.

[0031] S44. By fusing the occlusion calculation results of all infrared detection paths, the complete actual boundary contour coordinate sequence of the paper is reconstructed through geometric calculation, and the actual boundary information of the paper is obtained.

[0032] In one embodiment, filtering and time-frequency analysis are performed on the multi-channel signal intensity and time-series data to calculate the occlusion position of the paper edge on each independent infrared detection path, obtaining the occlusion calculation result, including:

[0033] S51. Filter the received signal of each channel to obtain the filtered signal;

[0034] S52. Extract the peak intensity and the time to reach the peak value of the filtered signal envelope, and calculate the flight time of the signal by combining the known signal transmission start time.

[0035] S53. Construct an obstruction location calculation model based on the filtered signal and the signal's time of flight. The obstruction location calculation model is used to map the received signal features to the physical obstruction distance to obtain the obstruction location.

[0036] S54. A Kalman filter is used to filter and track the occlusion positions of each channel calculated within a continuous sampling period, and the real-time occlusion position sequence is output to obtain the occlusion calculation result.

[0037] In one embodiment, the stamping instruction is obtained based on the actual boundary information and the target boundary data, including:

[0038] S61. Calculate the overlap between the contour reconstructed from the actual boundary information and the region defined by the target boundary data, and the set of shortest distances from each point of the actual contour to the boundary of the target region.

[0039] S62. When the overlap is 100% and all values ​​in the shortest distance set are greater than the first security threshold, generate a stamping permission instruction.

[0040] S63. When the overlap is less than 100% and greater than the minimum allowable threshold, and the maximum value in the shortest distance set is less than zero but the absolute value is less than the second fault tolerance threshold, a fine-tuning instruction is generated; the fine-tuning instruction is used to control the micro-motion of the stamping platform to push the paper to the target area.

[0041] S64. When the overlap is less than the minimum allowable threshold, or the minimum value in the shortest distance set is less than a negative locking threshold, generate an immediate lock command.

[0042] S65. Generate a stamping instruction based on the stamping permission instruction, fine-tuning instruction, and immediate stamping lock instruction.

[0043] Secondly, this application also provides an intelligent control system for an intelligent stamping machine, comprising:

[0044] The data acquisition module is used to acquire environmental status data and target boundary data for the current stamping task;

[0045] The stamping machine status module is used to obtain a status feature vector based on environmental status data and target boundary data; the status feature vector is used to characterize the current working condition of the stamping machine.

[0046] The sensor parameter optimization module is used to input the state feature vector into the pre-trained reinforcement learning decision model. The reinforcement learning decision model performs joint optimization decisions on multiple adjustable parameters of the infrared sensing system on the stamping machine and outputs the current optimal parameter combination.

[0047] The paper actual boundary recognition module is used to configure the working parameters of the infrared sensing system according to the current optimal parameter combination, drive the infrared sensing system to scan and detect the paper boundary, and obtain the actual boundary information of the paper.

[0048] The stamping permission execution module is used to obtain stamping instructions based on actual boundary information and target boundary data; the stamping instructions are used to instruct the stamping machine to perform the stamping operation and obtain the stamping result.

[0049] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect.

[0050] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect.

[0051] The aforementioned intelligent control method, system, device, and medium for an intelligent stamping machine first acquires environmental state data and target boundary data for the current stamping task. Then, based on this data, a state feature vector characterizing the current operating condition of the stamping machine is obtained. This state feature vector is then input into a pre-trained reinforcement learning decision model to jointly optimize multiple adjustable parameters of the infrared sensing system, outputting the current optimal parameter combination. Next, the operating parameters of the infrared sensing system are configured according to the optimal parameter combination, and the system is driven to scan and detect the paper boundary to obtain actual boundary information. Finally, a stamping command is generated based on the actual boundary information and target boundary data, and the stamping operation is executed. This achieves dynamic optimization of the infrared sensing system parameters, effectively improving the accuracy of actual paper boundary recognition, ensuring compliance of stamping operations, and thus solving problems such as poor environmental adaptability, insufficient boundary detection accuracy, and safety hazards associated with traditional stamping machines. This improves the accuracy, security, and intelligence level of stamping. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is a flowchart illustrating an intelligent control method for an intelligent stamping machine in one embodiment;

[0054] Figure 2 This is a schematic diagram of the structure of an intelligent stamping machine in one embodiment;

[0055] Figure 3 This is a schematic diagram of the structure of an intelligent stamping machine for stamping in one embodiment;

[0056] Figure 4 This is a schematic diagram of the intelligent control system of an intelligent stamping machine in one embodiment. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0058] refer to Figure 1 The document presents a flowchart illustrating an intelligent control method for an intelligent stamping machine provided in this application. The method includes the following steps:

[0059] S1. Obtain the environmental status data and target boundary data of the current stamping task.

[0060] For example, acquiring environmental status data relies on the multi-dimensional sensing module integrated into the stamping machine to achieve comprehensive perception of the stamping scene. Ambient light intensity is collected using a high-precision photoresistor sensor deployed around the stamping machine's document scanner. This sensor captures dynamic changes in ambient light in real time at a set sampling frequency, accurately reflecting the impact of different lighting conditions, such as natural light and artificial light, on infrared signal propagation. Paper characteristic parameters are acquired through a built-in thickness detection sensor and transmittance detection module. Thickness detection uses a piezoresistive sensor attached to the paper placement area at the bottom of the document scanner to detect the thickness of different types of paper. Transmittance detection uses the infrared transmission principle, emitting infrared signals of a specific wavelength that penetrate the paper. The transmittance is calculated based on the received signal strength, thus accurately identifying the characteristics of different types of paper, such as ordinary printing paper, cardstock, and transparent film. The infrared sensing system's own status data is collected by the sensor's built-in performance monitoring unit, including the operating voltage of each infrared emitter, the temperature of the emitting tube, and the signal noise intensity of the receiver, monitoring in real time for any abnormalities such as performance drift or aging.

[0061] The acquisition of target boundary data is achieved through a combination of a user interaction module and a preset database. Users can directly select the standard stamping range option via the stamping machine's touchscreen, a companion computer client, or a mobile app, or set the boundary coordinates of non-standard stamping areas through a custom input interface. A two-dimensional Cartesian coordinate system is established with a corner of the stamping machine as the origin. Users input the coordinate axis values ​​of the four vertices of the stamping area, and the system automatically calculates the length, width, and range threshold of the target boundary. Simultaneously, the system's built-in standard size database stores boundary parameters corresponding to each standard stamping range, historical data of optimal infrared sensor parameters, etc., which can quickly match basic boundary data for the user's selected standard size, improving data acquisition efficiency. All collected environmental state data and target boundary data are converted into digital signals through an analog-to-digital converter module and transmitted to the Central Processing Unit (CPU) via an internal high-speed bus for caching and preprocessing. The preprocessing process includes data filtering and outlier removal. Data filtering uses a Kalman filter algorithm to remove random noise, and outlier removal is based on criteria to judge and remove data exceeding a reasonable range, ensuring data accuracy and reliability.

[0062] S2. Based on environmental state data and target boundary data, obtain the state feature vector.

[0063] Optionally, the state feature vector is used to characterize the current operating condition of the stamping machine.

[0064] Optionally, firstly, feature extraction and standardization are performed on the preprocessed environmental state data and target boundary data. For continuous variables in the environmental state data, the min-max standardization method is used to map them to intervals, and the standardization formula is:

[0065]

[0066] in, This is the original data. This is the historical minimum value of the variable. This represents the historical maximum value of the variable. Standardization eliminates the dimensional differences between different variables, ensuring a balanced weighting of each feature in model training and decision-making. For the target boundary data, key features such as the length, width, area, and extreme values ​​of the boundary coordinates of the printing area are extracted and also standardized.

[0067] Secondly, a dimensional system for the state feature vector is constructed. The dimensions of the feature vector are set according to the actual working conditions, specifically including standardized ambient light intensity, paper thickness, paper transmittance, operating voltage of each infrared emitter, temperature of each infrared emitter, infrared receiver signal noise intensity, stamping area length, stamping area width, and stamping area area. Each dimension corresponds to a feature value, comprehensively covering the core factors affecting the working performance of the infrared sensing system. During the feature vector construction process, the effectiveness of each feature needs to be verified. The correlation between each feature and the detection accuracy of the infrared sensor is analyzed using the Pearson correlation coefficient. Redundant features with extremely low correlation are eliminated, while highly correlated features are retained to ensure that the state feature vector can accurately represent the current working condition of the stamping machine and avoid increasing the computational complexity of the model.

[0068] Finally, the verified feature values ​​are arranged in a preset order to form a fixed-dimensional state feature vector. The vector is stored in a one-dimensional array format for easy input into the reinforcement learning decision model. Each element corresponds to a specific working condition feature parameter, achieving a comprehensive and accurate quantitative representation of the current stamping scenario's working condition.

[0069] S3. Input the state feature vector into the pre-trained reinforcement learning decision model. The reinforcement learning decision model performs joint optimization decision-making on multiple adjustable parameters of the infrared sensing system on the stamping machine and outputs the current optimal parameter combination.

[0070] Specifically, the core of the reinforcement learning decision model employs a deep deterministic policy gradient algorithm. This model consists of an Actor network (policy network) and a Critic network (value function network), both using a three-layer fully connected neural network structure. The Actor network is responsible for generating parameter combination decisions for the infrared sensing system, while the Critic network evaluates the quality of these decisions. The model's pre-training process is based on a large amount of operational sample data. This sample data comes from measured data under different environmental conditions, paper types, and printing ranges. Each sample includes environmental state data, target boundary data, adjustable parameter combinations of the infrared sensing system, and corresponding detection accuracy evaluation indicators. A large number of valid samples are collected to construct the training dataset.

[0071] During training, the reward function is to maximize the boundary detection accuracy of the infrared sensing system and minimize the false positive rate. The reward function formula is designed as follows:

[0072]

[0073] in, The weighting coefficient for detection accuracy. The weighting coefficients represent the false positive rate. The network parameters are continuously adjusted using the gradient descent algorithm, enabling the model to learn to output the optimal parameter combination decision based on the feature vectors of different operating conditions. The adjustable parameters of the infrared sensing system include the transmission power of each infrared transmitter, the signal modulation frequency, and the gain of the infrared receiver, totaling multiple adjustable parameters. The model needs to jointly optimize these parameters.

[0074] After the state feature vector is input into the model, the Actor network generates a set of candidate solutions for parameter combinations based on the current operating conditions through forward propagation. The Critic network then evaluates the value of these candidate solutions using historical evaluation data, determining the expected detection performance of the parameter combination under the current operating conditions. If the evaluation value does not reach a preset threshold (based on historical best detection performance), the model adjusts its parameters through backpropagation and regenerates candidate solutions until the optimal parameter combination is found. This optimal parameter combination is the best parameter combination under the current operating conditions, achieving the best balance between infrared signal propagation efficiency, anti-interference capability, and boundary detection accuracy. The optimal parameter combination is output to the CPU in the form of digital instructions, providing a basis for subsequent parameter configuration of the infrared sensing system.

[0075] S4. Configure the working parameters of the infrared sensing system according to the current optimal parameter combination, drive the infrared sensing system to scan and detect the paper boundary, and obtain the actual boundary information of the paper.

[0076] Specifically, after receiving the optimal parameter combination, the CPU transmits each parameter instruction to the corresponding execution module of the infrared sensing system through a dedicated parameter configuration interface. For the transmission power configuration of the infrared transmitters, each transmitter is equipped with an independent high-precision power drive circuit. The drive circuit uses a DAC (Digital-to-Analog Converter) chip to convert the digital power instruction into a corresponding analog voltage signal. Precise control of the transmission power is achieved by adjusting the operating current of the transmitting tube, ensuring that the power of each transmitter strictly matches the requirements of the optimal parameter combination. The signal modulation frequency configuration is achieved through a frequency synthesizer. The frequency control instruction output by the CPU drives the frequency synthesizer to generate a corresponding square wave signal. This square wave signal is used to modulate the output signal of the infrared transmitter, avoiding the influence of external infrared interference signals. The modulated infrared signal is a pulse signal, and the pulse width and duty cycle are adaptively adjusted according to the modulation frequency to ensure signal stability and recognizability. The gain configuration of the infrared receiver is achieved through a programmable gain amplifier. According to the optimal gain parameter instruction, the amplifier automatically adjusts its internal amplification factor, enabling the receiver to maintain optimal receiving sensitivity under different signal strengths while suppressing the amplification of noise signals.

[0077] After parameter configuration, the CPU sends a scan start command to drive the infrared sensing system to start working. Each infrared transmitter starts synchronously according to its preset position, continuously emitting infrared signals according to the optimal parameter combination and set transmission power and modulation frequency, forming an infrared scanning network covering the target printing area. The scanning process uses a combination of line-by-line scanning and edge positioning. After the infrared signal is emitted from the transmitter, it is reflected or transmitted when it encounters paper. The infrared receiver works with the transmitter according to a preset pairing relationship, receiving the reflected or transmitted infrared signals, converting the optical signals into electrical signals, and extracting the signal intensity characteristic values ​​after amplification, filtering, demodulation, and other processing.

[0078] The system determines the paper boundary position by monitoring the signal strength changes of each receiver in real time. When the signal strength abruptly changes from below a threshold to above a threshold, it is determined as the starting boundary of the paper; when the signal strength abruptly changes from above a threshold to below a threshold, it is determined as the ending boundary of the paper. Combined with the scanned position coordinates, which are obtained based on the coordinate axis coordinates of the scanned point obtained by the displacement sensor built into the stamping machine, the actual boundary coordinates of each edge of the paper can be determined. After scanning, the system performs fitting processing on the boundary coordinates of each edge, using the least squares method to eliminate coordinate fluctuations caused by scanning noise, and obtains the coordinates of the four vertices and contour parameters of the actual boundary of the paper, forming complete actual boundary information, which is stored in the CPU cache for subsequent stamping command judgment.

[0079] S5. Based on the actual boundary information and target boundary data, obtain the stamping instruction.

[0080] Optionally, the stamping instruction is used to instruct the stamping machine to perform a stamping operation to obtain a stamping result.

[0081] Specifically, the CPU invokes a boundary matching algorithm to perform a precise comparison and analysis between the actual boundary information and the target boundary data. First, the overlap between the actual boundary and the target boundary is calculated. The overlap calculation uses the intersection-union ratio algorithm, which is the area of ​​the intersection of the actual printing area and the target printing area divided by the area of ​​the union. At the same time, the coordinate deviation between each vertex of the actual boundary and the corresponding vertex of the target boundary is calculated, and this deviation needs to be compared with a set deviation threshold.

[0082] If the intersection-union ratio reaches a preset standard and all vertex coordinate deviations are within the deviation threshold range, the paper is determined to be completely within the target stamping area, and the CPU generates a stamping permission command. This command includes parameters such as stamping position coordinates, stamping pressure, and stamping speed. The stamping position coordinates are calculated based on the center position of the target stamping area or the user-preset stamping point. The stamping pressure is dynamically adjusted according to the paper thickness parameter to ensure a clear and uniform imprint. The stamping permission command is transmitted to the stamping execution mechanism through the drive interface. This mechanism consists of a stepper motor, a transmission gear set, and a stamp fixing device. After receiving the command, the stepper motor drives the transmission gear set to move according to the preset speed and rotation angle, driving the stamp fixing device to move downwards to complete the stamping action. After stamping is completed, the stepper motor rotates in the reverse direction, driving the stamp back to the initial position and sending a stamping completion signal back to the CPU.

[0083] If the intersection-to-union ratio does not meet the preset standard or if the vertex coordinate deviation exceeds the deviation threshold, the paper is determined to be outside the target stamping range or its position is offset. The CPU immediately generates a prohibition stamping command and sends it to the locking mechanism. This mechanism uses an electromagnetic lock design. After receiving the command, the electromagnetic lock is quickly energized, generating electromagnetic attraction to lock the stamp fixing device in the initial position, preventing it from moving downwards. At the same time, the stamping machine's operating terminal displays locking prompt information, including the current paper offset direction and offset distance, reminding the user to adjust the paper position. After the user repositions the paper, the system automatically triggers a new round of environmental status data acquisition, boundary detection, and comparison processes until the paper meets the stamping conditions, generating a stamping permission command and executing the stamping operation. This ensures that stamping is only completed within the target stamping range, completely eliminating the security risks of stamping in unauthorized areas and guaranteeing the compliance and accuracy of the stamping results.

[0084] To further illustrate the solutions of the embodiments of this application, the following description uses a specific embodiment applied to the intelligent control method of the above-mentioned intelligent stamping machine.

[0085] refer to Figure 2 The document presents a structural schematic diagram of an intelligent stamping machine provided in this application. Infrared sensors are arranged at the top of the machine, and other sensors are positioned at the four edges of the bottom to minimize misjudgments between the four sensor transmitters and receivers and to ensure accuracy in the stamping range. The stamping machine internally incorporates an adjustable infrared sensor (i.e., an infrared transmitter), allowing users to set the sensor parameters via options to flexibly meet various stamping range requirements.

[0086] refer to Figure 3The document presents a structural diagram of an intelligent stamping machine provided in this application. The user selects the stamping area A3 or other based on their needs. After the CPU receives the option, it automatically adjusts the power parameters of the four infrared transmitters. The parameters of each infrared transmitter are pre-configured according to the stamping area. When the stamping machine is located in corner A, the infrared receiver will only receive signals from three sensors: D corresponds to sensor 2, B to sensor 4, and C to sensor 1. The edges of AD correspond to sensor 1 or 4, the edges of AB to sensor 1 or 2, the edges of BC to sensor 2 or 3, and the edges of CD to sensor 3 or 4. Outside this area, no sensors can be received, and the stamping machine automatically locks the stamp and stops stamping.

[0087] In the aforementioned intelligent control method for an intelligent stamping machine, the following steps are taken: First, the environmental state data and target boundary data of the current stamping task are acquired. Then, based on these data, a state feature vector representing the current working condition of the stamping machine is constructed. Subsequently, the state feature vector is input into a pre-trained reinforcement learning decision model to jointly optimize multiple adjustable parameters of the infrared sensing system to output the current optimal parameter combination. Next, the infrared sensing system is configured according to the optimal parameter combination and driven to scan and detect the paper boundary to obtain the actual boundary information. Finally, a stamping instruction is generated based on the actual boundary information and the target boundary data, and the stamping operation is executed in a coherent manner. This achieves dynamic adaptation of the infrared sensing system parameters, improves the accuracy of paper boundary detection, and ensures that the stamping operation is performed compliantly only within the target range. This effectively solves the problems of poor environmental adaptability, insufficient boundary detection accuracy, and security risks of stamping in unauthorized areas caused by the fixed parameters of traditional stamping machines, thereby improving the accuracy, security, and intelligence level of stamping.

[0088] In an optional embodiment, obtaining the environmental state data and target boundary data of the current stamping task includes the following steps:

[0089] S11. Control all infrared transmitters in the infrared sensing system to transmit signals at a reference power, and synchronously acquire the raw signal values ​​of all infrared receivers to obtain the reference signal vector.

[0090] Optionally, the CPU first sends a reference transmission command to the infrared sensing system. This command specifies that all infrared transmitters will transmit infrared signals at a preset reference power. This reference power is a fixed power value calibrated beforehand, ensuring stable propagation of the infrared signal in an unobstructed, standard environment, providing a unified reference for subsequent calculations of environmental impact parameters. Simultaneously with the infrared transmitters initiating signal transmission, the CPU triggers all infrared receivers to synchronously enter signal acquisition mode. The receivers continuously acquire the received infrared signals at a preset sampling frequency for a preset duration to obtain sufficient raw signal samples and reduce the impact of random factors on signal quality. The acquired raw signal values ​​are analog electrical signals, which are converted into digital signals by the receiver's built-in analog-to-digital converter and transmitted to the CPU. The CPU organizes the digitized raw signal values ​​transmitted by each receiver, arranging and combining the corresponding raw signal values ​​according to a preset receiver number order to form a one-dimensional reference signal vector. Each element of this vector corresponds to the digitized value of the reference power signal received by an infrared receiver in the current environment, fully reflecting the reference signal status of each receiving channel in the current environment.

[0091] S12. Based on the reference signal vector and the pre-calibrated no-load environment reference signal vector, calculate the signal attenuation coefficient vector and background noise intensity vector under the current environment.

[0092] Optionally, the pre-calibrated airborne environmental reference signal vector is a reference signal vector obtained under standard environmental conditions with no obstructions, using the same reference power transmission signal and the same sampling parameters as the current reference signal. This vector is pre-stored in the CPU's storage module as a benchmark reference for environmental impact assessment. The CPU calls a preset calculation algorithm to first calculate the difference between the corresponding elements in the reference signal vector and the airborne environmental reference signal vector. This difference reflects the comprehensive impact of the current environment on the infrared signals of each channel relative to the standard environment. Based on this difference and the corresponding elements in the airborne environmental reference signal vector, the signal attenuation coefficient of each channel is calculated. The signal attenuation coefficient is the ratio of the current reference signal value to the airborne environmental reference signal value, which can quantify the attenuation degree of the current environment on the propagation of the infrared signals of each channel. The signal attenuation coefficients of all channels are arranged in order of receiver number to form a signal attenuation coefficient vector. At the same time, the CPU performs noise analysis on the currently acquired reference signal samples, and characterizes the background noise intensity by calculating the variance of the original signal samples. The larger the variance value, the more serious the interference of the current environment's background noise on the signal. The variance of the original signal sample corresponding to each receiver is calculated to obtain the background noise intensity value of each channel. The background noise intensity vector is formed by arranging the values ​​in order of receiver number. This vector, together with the signal attenuation coefficient vector, constitutes the core parameter of the current environment's influence on infrared signal propagation, providing a key basis for the subsequent optimization of infrared sensing system parameters.

[0093] S13. Read the historical task logs of the stamping machine, obtain the final decision results and boundary error statistics of historical tasks similar to the current target boundary data, and use them as historical performance reference data.

[0094] Optionally, the CPU first retrieves the historical task log of the stamping machine from the storage module. This log records key information for each historical stamping task, including the target boundary data corresponding to the task, the parameter configuration of the infrared sensing system, the final decision result, and the boundary detection error statistics. Then, the CPU initiates a similarity matching algorithm, using the current target boundary data as the matching benchmark, to filter historical tasks with similar target boundary data from the historical task log. The similarity matching is determined by the similarity of the feature parameters between the current target boundary data and the target boundary data of the historical task. The feature parameters of the target boundary data include key parameters such as the shape and size range of the stamping area. When the similarity of the feature parameters reaches a preset threshold, the historical task is determined to be similar. For the filtered similar historical tasks, the CPU extracts their final decision result and boundary error statistics. The final decision result includes the optimal parameter combination of the infrared sensing system in the historical task, the instruction on whether stamping is allowed, etc., while the boundary error statistics include the deviation statistics between the actual boundary and the target boundary in the historical task, the detection error rate, etc. The extracted information is integrated to form historical performance reference data. This data can provide historical experience support for parameter optimization decisions for the current task, and help improve the accuracy and efficiency of current decisions.

[0095] S14. Combine the signal attenuation coefficient vector, background noise intensity vector, and historical performance reference data to form environmental status data.

[0096] Optionally, the CPU standardizes the signal attenuation coefficient vector, background noise intensity vector, and historical performance reference data to ensure uniformity and compatibility of various data formats. The signal attenuation coefficient vector and background noise intensity vector are already in a standardized one-dimensional array format, while the historical performance reference data is converted into a standardized structured data format, including a vector representation of the historical optimal parameter combination and a numerical representation of boundary error statistics. Subsequently, the CPU integrates the standardized signal attenuation coefficient vector, background noise intensity vector, and historical performance reference data to form complete environmental state data. This environmental state data comprehensively covers the parameters affecting the infrared signal from the current environment and performance reference information from similar historical tasks, providing comprehensive and accurate environment-related input data for the subsequent construction of state feature vectors, ensuring that the state feature vectors accurately represent the operating condition of the stamping machine in the current environment.

[0097] The acquisition of target boundary data is achieved through a combination of a user interaction module and a preset database. Users can directly select the standard stamping range option via the stamping machine's touchscreen, a companion computer client, or a mobile app, or set the boundary coordinates of non-standard stamping areas through a custom input interface. A two-dimensional Cartesian coordinate system is established with a corner of the stamping machine as the origin. Users input the coordinate axis values ​​of the four vertices of the stamping area, and the system automatically calculates the length, width, and range threshold of the target boundary. Simultaneously, the system's built-in standard size database stores boundary parameters corresponding to each standard stamping range, historical data of optimal infrared sensor parameters, etc., which can quickly match basic boundary data for the user's selected standard size, improving data acquisition efficiency. All collected environmental state data and target boundary data are converted into digital signals through an analog-to-digital conversion module (if already digitized, they are used directly), and transmitted to the CPU via an internal high-speed bus for caching and preprocessing. The preprocessing process includes data filtering and outlier removal. Data filtering uses a Kalman filter algorithm to remove random noise, and outlier removal is based on criteria to judge and remove data exceeding a reasonable range, ensuring data accuracy and reliability.

[0098] In an optional embodiment, a state feature vector is obtained based on environmental state data and target boundary data, including the following steps:

[0099] S21. Normalize the environmental state data to obtain a normalized environmental vector.

[0100] Optionally, the environmental state data includes a signal attenuation coefficient vector, a background noise intensity vector, and historical performance reference data. The signal attenuation coefficient vector and background noise intensity vector are continuous numerical vectors, while the historical performance reference data is structured data. First, key numerical information from the historical performance reference data needs to be extracted and converted into vector form to form unified numerical environmental data. Then, the Z-score normalization method is used to normalize all elements in the numerical environmental data. This method eliminates dimensional and numerical range differences between different dimensions by converting each data point into its deviation from the dataset mean and standard deviation. After normalizing all numerical environmental data, the vector elements are concatenated and integrated according to a preset dimensional order to form a one-dimensional normalized environmental vector. This vector fully preserves the core information of the environmental state data and has a unified numerical scale.

[0101] S22. Encode and transform the target boundary data to obtain the boundary encoding vector.

[0102] Optionally, the target boundary data includes key parameters such as the boundary coordinates, length, width, and area of ​​the stamping area. These parameters have different physical meanings and data formats, and need to be converted into a vector form suitable for feature fusion through encoding. First, the coordinate parameters of the target boundary are normalized. Using the origin preset by the stamping machine as a reference, all boundary coordinates are converted into relative coordinate values ​​to ensure the consistency of the coordinate data. Then, the target boundary data is transformed using an embedded encoding method. The length, width, area, and normalized boundary coordinates of the stamping area are taken as input and converted into a fixed-dimensional low-dimensional vector through a preset encoding network or mapping function. During the encoding process, the geometric and dimensional features of the target boundary data must be preserved to ensure that the encoded vector can accurately represent the core attributes of the target stamping area. For example, for a rectangular stamping area, the encoding process should focus on preserving key features such as the ratio of its length and width and the center coordinate position, so that the boundary encoding vector can form an effective feature complementarity with the normalized environment vector. After encoding, a fixed-dimensional boundary encoding vector is obtained. Each element of this vector is a numerical value that has been encoded and can be directly used for subsequent feature fusion operations.

[0103] S23. Based on the normalized environment vector and the boundary encoding vector, feature fusion is performed to obtain the fused feature vector.

[0104] Optionally, a combination of feature concatenation and attention weighting is used to fuse the normalized environment vector and the boundary encoding vector. First, the normalized environment vector and the boundary encoding vector are directly concatenated in a preset order to form an initial fusion vector. Then, an attention mechanism is introduced to assign weights to the features of each dimension in the initial fusion vector. By calculating the correlation between each dimension feature and the detection accuracy of the infrared sensing system, features with high correlation are assigned higher weights, enhancing the influence of key features on subsequent decisions while reducing the weights of redundant features. The calculation of attention weights is based on a preset attention model, which, through learning from historical data, can adaptively identify key feature dimensions affecting detection accuracy under different operating conditions. The initial fusion vector is multiplied element-wise with the corresponding attention weights to obtain weighted fusion features. These weighted features are then dimensionally integrated to finally form a fusion feature vector. This vector simultaneously covers the core information of the environmental state and the key features of the target boundary, and the effectiveness of the features is optimized through attention weighting, providing high-quality feature data for subsequent dimensional transformations and model input.

[0105] S24. The dimensionality of the fused feature vector is transformed by a fully connected layer to obtain the state feature vector.

[0106] Optionally, the dimension of the fused feature vector may not directly match the input dimension requirements of the reinforcement learning decision model, thus requiring dimensionality transformation through a fully connected layer. The fully connected layer consists of a predetermined number of neurons, each connected to all dimensions of the fused feature vector. Linear transformations and activation functions are used to adjust the dimensions and extract features from the fused feature vector. First, the fused feature vector is input into the fully connected layer. A preliminary dimensionality transformation is achieved through matrix multiplication of the weight matrix and the feature vector, mapping the fused feature vector to a predetermined intermediate dimensional space. Subsequently, a nonlinear activation function is used to perform a nonlinear mapping on the transformed features, enhancing their expressive power and capturing complex nonlinear relationships between features. The ReLU activation function is used, effectively mitigating the gradient vanishing problem and improving the stability of the feature transformation. After the linear transformation and nonlinear activation processing of the fully connected layer, the dimension of the output feature vector perfectly matches the input dimension of the reinforcement learning decision model; this vector is the state feature vector. The state feature vector retains the core information of the environmental state and target boundary, and through dimensionality transformation and feature optimization, it adapts to the model's input requirements, accurately representing the current operating condition of the stamping machine and providing a reliable input basis for the model's optimization decision.

[0107] In an optional embodiment, the pre-training process of the reinforcement learning decision model includes:

[0108] S31. Construct the state space of the reinforcement learning decision model. The state vector of the state space includes the state feature vector, the boundary detection error of the previous cycle in real time feedback, and the signal-to-noise ratio of the sensor signal.

[0109] Optionally, the state space is the foundation for the reinforcement learning decision model to perceive the external environment and its own operating state. Its construction needs to comprehensively cover the key state information that affects parameter adjustment decisions. The core component of the state vector is the state feature vector obtained above. This vector integrates the core information of environmental state data and target boundary data, and can characterize the basic operating conditions of the stamping machine. On this basis, the boundary detection error of the previous cycle is introduced in real time feedback. This error is the statistical value of the deviation between the actual boundary information and the target boundary data in the previous cycle. It can reflect the current detection accuracy of the infrared sensing system under the current parameter configuration, and provide a direct performance feedback basis for model parameter adjustment. At the same time, the sensor signal signal-to-noise ratio is included. This index is calculated by the ratio of the effective signal strength received by the infrared receiver to the background noise intensity. It can quantify the quality of the infrared signal in the current environment and directly affect the accuracy of boundary detection. It is a key factor to consider in parameter adjustment. The above three parts of information are integrated in a preset dimensional order to form the state vector of the state space. Each state vector corresponds to a specific operating state of the model. All possible state vectors together constitute the complete state space, providing a comprehensive state input foundation for the model's decision learning.

[0110] S32. Construct the action space of the reinforcement learning decision-making model;

[0111] Optionally, the motion vector in the motion space is defined as the adjustment amount of multiple adjustable parameters in the infrared sensing system, including the transmission power of each infrared transmitter, the gain threshold of each infrared receiver, and the signal sampling frequency.

[0112] Optionally, the action space is constructed around the adjustable parameters of the infrared sensing system. The core meaning of the action vector is the adjustment amount of the adjustable parameters rather than their absolute values. This definition allows the model to focus on the optimization direction and magnitude of the parameters, improving the accuracy and flexibility of decision-making. The adjustable parameters are selected from three categories that have a key impact on infrared signal propagation and detection: the transmission power of each infrared transmitter, whose adjustment directly affects the coverage and penetration capability of the infrared signal; the gain threshold of each infrared receiver, whose adjustment can change the receiver's sensitivity to the signal and adapt to signal environments of different intensities; and the signal sampling frequency, whose adjustment can balance detection accuracy and system resource consumption. High-frequency sampling can improve detection accuracy but increase resource consumption, while low-frequency sampling has the opposite effect. Each element in the action vector corresponds to the adjustment amount of an adjustable parameter. The range of the adjustment amount is calibrated through prior experiments to ensure that adjusting the parameters within this range will not cause abnormal operation of the infrared sensing system. All possible action vectors together constitute the action space. The model's decision-making process involves selecting the optimal action vector from the action space to achieve precise adjustment of the infrared sensing system parameters.

[0113] S33. Define the reward function for the reinforcement learning decision-making model;

[0114] Optionally, the reward function calculates the instantaneous reward based on the degree of agreement between the actual boundary information generated after the action is executed and the target boundary data, as well as the sensor resources consumed to achieve the agreement.

[0115] Optionally, the reward function is the core of guiding the reinforcement learning decision model to learn the optimal strategy. Its design must balance detection accuracy and resource economy, ensuring that the parameter adjustment actions output by the model can guarantee the accuracy of boundary detection while avoiding unnecessary waste of sensor resources. The calculation of the reward function is based on two core indicators after the action is executed: first, the degree of agreement between the actual boundary information and the target boundary data, which is calculated using the intersection-over-union (IoU) algorithm mentioned earlier. The larger the IoU value, the higher the degree of agreement, and the higher the corresponding reward value; second, the sensor resources consumed to achieve this degree of agreement. Resource consumption is mainly quantified by indicators such as the power consumption of the infrared emitter and the computational resource usage corresponding to the signal sampling frequency. The greater the resource consumption, the greater the corresponding reduction in reward value. The specific expression of the reward function is constructed through a linear weighted method, combining the positive reward corresponding to the degree of agreement with the negative penalty corresponding to the resource consumption to form the final immediate reward value. Through this reward function, the model can gradually master the parameter adjustment strategy that balances detection accuracy and resource consumption during the learning process, ensuring the optimality and economy of the decision.

[0116] S34. Use the proximal policy optimization algorithm to train the reinforcement learning decision model offline until the model can stably output parameter adjustment actions that maximize the reward function in response to simulated state inputs, thus completing the pre-training.

[0117] Optionally, preparing training data is fundamental for offline training. By building a simulated training environment, a large number of simulated working condition samples are generated, covering different environmental conditions, target boundary data, and paper types. Each sample contains a state vector in the state space, an action vector in the action space, and a corresponding reward value, thus constructing a large-scale offline training dataset. The training process employs a proximal policy optimization algorithm. This algorithm, by limiting the magnitude of policy updates, avoids training instability caused by policy mutations, thereby improving the convergence and stability of model training. During training, the state vectors from the simulated training dataset are input into the model. The Actor network generates action vectors based on the current state, and the Critic network calculates the reward value corresponding to the action using the reward function and feeds it back to the Actor network. The model continuously adjusts the network parameters using the gradient descent algorithm, ensuring that the action vectors generated by the Actor network maximize the reward function value. During training, the stability of the model's output and the mean change of the reward function are continuously monitored. When the model can stably output parameter adjustment actions that maximize the reward function for different simulated input states, and the mean of the reward function tends to stabilize and no longer increases significantly in multiple consecutive training cycles, the model training is considered complete. At this time, the model has the ability to adaptively output the optimal parameter combination according to the actual working conditions. The model parameters after pre-training are fixed and stored in the CPU for use in actual stamping tasks.

[0118] In an optional embodiment, the operating parameters of the infrared sensing system are configured according to the current optimal parameter combination, and the infrared sensing system is driven to scan and detect the paper boundary to obtain the actual boundary information of the paper, including the following steps:

[0119] S41. Based on the current optimal parameter combination, set the modulation frequency and phase of each infrared transmitter in the infrared sensing system to transmit pulse signals in a time-division manner according to the preset timing sequence.

[0120] Optionally, the CPU first analyzes the modulation frequency, phase, and timing control parameters of the infrared transmitters in the current optimal parameter combination. These parameters are optimized by the reinforcement learning decision model for the current operating conditions, maximizing signal anti-interference capability and detection accuracy. Based on the analysis results, the CPU sends independent configuration instructions to the driver module of each infrared transmitter, specifying the modulation frequency and phase parameters of each transmitter. This ensures that the transmitted signals from different transmitters have distinguishable time-frequency characteristics, avoiding superimposed interference between multi-channel signals. Simultaneously, the CPU controls each infrared transmitter to start transmitting pulse signals in a time-division manner according to a preset timing plan. The timing plan is based on the spatial distribution design of the infrared detection path, ensuring that transmitters corresponding to adjacent detection paths do not transmit signals simultaneously, further reducing crosstalk between channels. The pulse signal transmission of each transmitter follows the pulse width and duty cycle set by the optimal parameter combination. The amplitude of the pulse signal is determined by the optimal transmission power parameter, precisely controlled by the power drive circuit to ensure that the energy of the transmitted signal matches the signal attenuation characteristics of the current environment, balancing detection range and energy efficiency.

[0121] S42. During each transmission pulse cycle, the response signal strength and transit time of the corresponding infrared receiver are synchronously acquired to obtain multi-channel signal strength and time series data.

[0122] Optionally, the CPU constructs a transmission-acquisition synchronization control mechanism. Simultaneously with each infrared transmitter emitting a pulse signal, the corresponding infrared receiver is triggered to enter synchronous acquisition mode, ensuring that the acquired response signal strictly corresponds to the transmitted pulse and avoiding data distortion caused by time offset. During acquisition, the receiver captures the infrared signal reflected or transmitted through the paper in real time. The built-in signal conditioning module performs preliminary amplification and shaping of the signal, followed by a high-precision analog-to-digital converter that converts the analog signal into a digital signal. Two core data points are extracted simultaneously: response signal strength and transit time. Response signal strength reflects the degree of energy attenuation of the infrared signal after passing through the paper, and is directly related to the paper's thickness, transmittance, and occlusion status. Transit time refers to the time interval between the infrared signal being emitted from the transmitter, propagating, interacting, and being captured by the receiver, and is related to the length of the signal propagation path and the medium characteristics. For each transmission pulse cycle, a corresponding signal strength and transit time acquisition is completed. Through continuous acquisition over multiple pulse cycles, signal strength sequences and time series data for each channel are formed, providing sufficient data support for subsequent occlusion location calculation. The acquired data from all channels are organized and stored in chronological order according to transmitter number, forming a structured multi-channel data set.

[0123] S43. Filter and perform time-frequency analysis on the multi-channel signal strength and time series data to calculate the occlusion position of the paper edge on each independent infrared detection path and obtain the occlusion calculation result.

[0124] Optionally, the multi-channel signal strength and time-series data are first preprocessed. An adaptive median filtering algorithm is used to remove impulse noise and random interference from the data. This algorithm can dynamically adjust the size of the filtering window according to the local characteristics of the data, effectively suppressing noise while preserving signal edge features. After filtering, time-frequency analysis is performed on the data. A short-time Fourier transform is used to convert the time-domain signal into a time-frequency domain signal. The time nodes of signal strength abrupt changes are identified through the time-frequency spectrum, which correspond to the moment when the infrared signal encounters the edge of the paper. Combining the known infrared signal propagation speed, the propagation distance of the infrared signal from the transmitter to the edge of the paper is calculated using transit time data. The propagation distance is calculated as the product of the propagation speed and the transit time. Based on the preset installation coordinates of the infrared transmitter and receiver, the propagation distance is converted into the specific occlusion position coordinates of the paper edge on the current infrared detection path through geometric derivation. For each independent infrared detection path, the above filtering, time-frequency analysis, and calculation process is repeated to obtain the occlusion position information on that path. The occlusion position information of all paths constitutes the occlusion calculation result, which accurately reflects the projected position of the paper edge on each detection path.

[0125] S44. By fusing the occlusion calculation results of all infrared detection paths, the complete actual boundary contour coordinate sequence of the paper is reconstructed through geometric calculation, and the actual boundary information of the paper is obtained.

[0126] Optionally, the CPU first verifies the validity of the occlusion calculation results for all infrared detection paths, eliminating abnormal calculation results caused by signal interference or data errors. The verification criteria are the reasonable deviation range between the calculation results and the target boundary data, and the consistency of the calculation results of adjacent detection paths. Subsequently, a multi-source data fusion algorithm is used to integrate the valid calculation results. The influence of single detection errors is weakened by weighted averaging, improving the overall reliability of the occlusion position. The weight values ​​are determined according to the signal-to-noise ratio of each detection path, with paths having higher signal-to-noise ratios being assigned higher weights. After fusion, a geometric reconstruction model in a global coordinate system is established based on the spatial arrangement parameters of the infrared transmitter and receiver, mapping the occlusion position coordinates on each detection path to the global coordinate system. A polygon fitting algorithm is used to fit all mapped occlusion position coordinates to construct the preliminary boundary contour of the paper. Then, the least squares method is used to smooth and optimize the preliminary contour, eliminating contour fluctuations caused by discrete points. Finally, the optimized coordinate sequence of the complete actual boundary contour of the paper is output. This coordinate sequence contains the coordinates of the key vertices of the paper boundary and the coordinates of continuous points on the contour, which fully represents the actual boundary shape and position of the paper, that is, the required actual boundary information, and is stored in the CPU for subsequent stamping instructions to judge.

[0127] In an optional embodiment, filtering and time-frequency analysis are performed on the multi-channel signal strength and time-series data to calculate the occlusion position of the paper edge on each independent infrared detection path, obtaining the occlusion calculation result, including the following steps:

[0128] S51. Filter the received signal of each channel to obtain the filtered signal.

[0129] Optionally, the multi-channel received signal contains various interference components such as environmental noise, circuit noise, and channel crosstalk. These interferences can cause signal feature distortion and affect the accuracy of subsequent occlusion location calculation. Therefore, each channel's received signal needs to be filtered separately. The filtering scheme adopts a layered strategy of "coarse filtering followed by fine filtering." First, a finite impulse response (FIR) low-pass filter is used to coarsely filter the original received signal, removing high-frequency noise components. The cutoff frequency of the low-pass filter is determined based on the modulation frequency of the infrared signal to ensure effective preservation of useful frequency components. Subsequently, an adaptive median filtering algorithm is used for fine filtering. This algorithm can dynamically adjust the size of the filtering window according to the local grayscale characteristics of the signal. When impulse noise is detected in the signal, the window is automatically increased to enhance noise suppression capability; when the signal is in a stable region, the window is decreased to preserve the detailed features of the signal and avoid over-filtering that leads to the loss of signal edge information. Through two layers of filtering, various interference components are effectively removed, resulting in a smooth filtered signal that retains complete edge features, providing a high-quality data foundation for subsequent signal feature extraction.

[0130] S52. Extract the peak intensity and peak time of the filtered signal envelope, and calculate the signal's flight time by combining the known signal transmission start time.

[0131] Optionally, the filtered signal is a pulse-modulated signal, and its envelope characteristics directly reflect the energy change law after the infrared signal interacts with the paper. Therefore, the envelope of the filtered signal needs to be extracted first. A Hilbert transform is used to process the filtered signal, converting the real signal into an analytic signal. The envelope curve is obtained by taking the modulus of the analytic signal, which accurately characterizes the signal intensity change trend over time. Based on the extracted envelope curve, a peak detection algorithm is used to identify the peak intensity of the envelope and the corresponding peak arrival time. The peak intensity reflects the maximum energy of the infrared signal after reflection or transmission through the paper, while the peak arrival time corresponds to the key time node when the infrared signal is captured by the receiver after interacting with the edge of the paper. The signal transmission start time is recorded by the CPU's synchronous timing module and is a known fixed value. By calculating the difference between the peak arrival time and the transmission start time, the flight time of the signal can be obtained. This flight time directly corresponds to the propagation time of the infrared signal from the transmitter to its return to the receiver after interacting with the edge of the paper, and is the core parameter for subsequent calculation of the obstruction distance.

[0132] S53. Construct an obstruction location calculation model based on the filtered signal and the signal's time of flight. The obstruction location calculation model is used to map the received signal characteristics to the physical obstruction distance to obtain the obstruction location.

[0133] Optionally, the obstruction location calculation model is constructed based on the propagation laws of infrared signals and the principles of geometric optics. The core is to establish a mapping relationship between the characteristics of the received signal and the physical obstruction distance. The model input parameters include the peak intensity of the filtered signal, the time of flight, and preset system parameters. These system parameters include known fixed parameters such as the installation coordinates of the infrared transmitter and receiver, and the propagation speed of the infrared signal in the air. First, based on the principle of uniform propagation, the propagation path of the infrared signal is calculated using the time of flight and the propagation speed. The propagation path, time of flight, and propagation speed satisfy a linear relationship of "propagation path = propagation speed × time of flight". Then, combining the spatial relative positions of the transmitter and receiver, the propagation path is converted into the vertical distance of the paper edge relative to the transmitter-receiver detection path, i.e., the physical obstruction distance, through geometric derivation. Simultaneously, peak intensity is introduced to correct the obstruction distance, as peak intensity is related to the reflection / transmission characteristics of the paper. Different paper characteristics will lead to differences in peak intensity, thus affecting the accuracy of the propagation path calculation. A preset correction coefficient is used to correlate the peak intensity with the obstruction distance, calibrating the initially calculated obstruction distance. Finally, the solution model outputs the specific occlusion position coordinates of the paper edge on the current infrared detection path. These coordinates are determined by taking the detection path as the reference axis and combining the mapping relationship of the system's global coordinate system.

[0134] S54. A Kalman filter is used to filter and track the occlusion positions of each channel calculated within a continuous sampling period, and the real-time occlusion position sequence is output to obtain the occlusion calculation result.

[0135] Optionally, the occlusion position calculated within a single sampling period may contain random errors, and the occlusion position data from continuous sampling may fluctuate. Therefore, a Kalman filter algorithm is needed for time-domain filtering and tracking to improve the stability and accuracy of the occlusion position. The core of the Kalman filter is to achieve optimal estimation of the state through an iterative "prediction-update" process. First, based on the occlusion position and motion state of the previous sampling period, a state prediction equation is established to predict the occlusion position value for the current sampling period. Then, the measured value of the occlusion position calculated in the current sampling period is fused with the predicted value, and the optimal occlusion position estimate for the current period is calculated through the state update equation. The update equation introduces the observation noise covariance matrix and the process noise covariance matrix to characterize the noise level of the measured data and the uncertainty of the system model, respectively. The initial values ​​of the two covariance matrices are calibrated through previous experiments and adaptively adjusted during the iteration process. The above "prediction-update" process is performed on the occlusion positions for multiple consecutive sampling periods to form a smooth real-time occlusion position sequence. This sequence can accurately track the positional changes of the paper edge on the detection path and effectively suppress the random errors of single-period calculation. By integrating the real-time occlusion location sequences of all channels, a complete occlusion solution is obtained, providing reliable single-path location data for subsequent multi-path fusion reconstruction of boundary contours.

[0136] In an optional embodiment, a stamping instruction is obtained based on the actual boundary information and the target boundary data, including:

[0137] S61. Calculate the overlap between the contour reconstructed from the actual boundary information and the region defined by the target boundary data, and the set of shortest distances from each point of the actual contour to the boundary of the target region.

[0138] Optionally, the CPU first extracts the complete contour coordinate sequence reconstructed from the actual boundary information and the region contour coordinates defined by the target boundary data. Based on these two sets of coordinate data, it performs dual feature quantization calculations. The first calculation is the overlap calculation, implemented using the intersection-union ratio algorithm. By solving the intersection area and union area of ​​the actual contour and the target region, the ratio of the intersection area to the union area is taken as the overlap. This ratio can intuitively reflect the degree of overlap between the actual position of the paper and the target printing area. The overlap value ranges from 0 to 100%, with a higher value indicating a higher degree of overlap. The second calculation is the calculation of the set of shortest distances from each point on the actual contour to the boundary of the target region. It traverses all coordinate points on the actual contour, and for each point, calculates its distance to all points on the boundary of the target region using a geometric distance algorithm. The minimum distance value is selected as the shortest distance corresponding to that point. The shortest distances of all actual contour points are arranged in order of coordinate points to form the set of shortest distances. When the shortest distance is positive, it means the corresponding contour point is inside the target area; when the shortest distance is negative, it means the corresponding contour point is outside the target area; and when the shortest distance is zero, it means the corresponding contour point falls exactly on the boundary of the target area. Through the dual quantification of overlap and the set of shortest distances, the deviation between the actual position of the paper and the target printing area can be comprehensively and accurately characterized, providing core data support for subsequent instruction determination.

[0139] S62. When the overlap is 100% and all values ​​in the shortest distance set are greater than the first security threshold, generate a stamping permission instruction.

[0140] Optionally, a 100% overlap indicates that the actual outline completely coincides with the target area, and the paper as a whole does not exceed the target stamping area. However, it is necessary to further verify whether the paper is in a safe stamping position through the shortest distance set to avoid the risk of stamping deviation due to the paper being too close to the target area boundary. The first safety threshold is a pre-calibrated safety distance threshold, representing the minimum safe gap that must be maintained between the paper outline and the target area boundary. This threshold is determined based on factors such as stamp size and stamping accuracy requirements to ensure that the stamp edge does not exceed the target stamping area when stamping within this gap range. When all values ​​in the shortest distance set are greater than the first safety threshold, it indicates that all points on the actual outline are inside the target area and far from the boundary, and the paper position meets the high-precision stamping requirements. The CPU then generates a stamping permission instruction. This instruction contains a complete stamping parameter configuration, including key parameters such as the stamping position coordinates based on the preset points in the target stamping area, the stamping force adapted to the current paper characteristics, and the stamping speed to ensure clear imprints. This provides a clear instruction basis for the precise actions of the subsequent stamping execution mechanism.

[0141] S63. When the overlap is less than 100% and greater than the minimum allowable threshold, and the maximum value in the shortest distance set is less than zero but the absolute value is less than the second fault tolerance threshold, generate a fine-tuning instruction.

[0142] Optionally, the fine-tuning command is used to control the micro-motion of the stamping platform to push the paper to the target area.

[0143] Optionally, if the overlap is less than 100% but greater than the minimum allowable threshold, it indicates that the paper has a slight offset but not a serious deviation, and it is still feasible to correct it through fine-tuning. The minimum allowable threshold is a pre-set critical value for correctable offset. When the overlap is below this threshold, the paper offset is too large to be corrected through fine-tuning. If the maximum value in the shortest distance set is less than zero, it means that all points on the actual contour are located inside the target area, and there is no situation where the paper is outside the target area. That is, the overall offset direction of the paper is offset within the target area. If the absolute value of the maximum value is less than the second fault tolerance threshold, it means that the offset is within the range of fine-tunable correction. The second fault tolerance threshold is a pre-calibrated maximum fine-tunable offset to ensure that the fine-tuning action will not damage the paper or cause secondary offset. At this time, the CPU generates a fine-tuning instruction, which includes parameters such as the micro-motion direction, micro-motion distance, and micro-motion speed of the stamping platform. The micro-motion direction and distance are calculated based on the distribution characteristics of the shortest distance set, that is, fine-tuning in the direction corresponding to the minimum value of the shortest distance until the actual contour meets the conditions for allowing stamping. The fine-tuning command is transmitted to the micro-motion drive mechanism of the stamping platform. This mechanism is driven by a high-precision stepper motor and achieves the micro-displacement of the platform through the lead screw transmission, ensuring that the fine-tuning accuracy meets the stamping requirements.

[0144] S64. When the overlap is less than the minimum allowed threshold, or the minimum value in the shortest distance set is less than a negative locking threshold, generate an immediate lock command.

[0145] Optionally, the two judgment conditions correspond to different severe offset scenarios. Meeting either condition requires generating an immediate locking command to eliminate potential safety hazards. The first condition, where the overlap is less than the minimum allowable threshold, indicates that the paper offset is too large, exceeding the fine-tuning correction range. If a stamp is forcibly applied under these conditions, the stamp will inevitably exceed the target stamping area, resulting in an invalid stamp or even document damage. The second condition, where the minimum value in the shortest distance set is less than a negative locking threshold (a pre-defined critical value for dangerous offset), indicates that a point on the actual outline severely exceeds the target area, and the excess distance has reached a dangerous level. Even fine-tuning cannot correct it to the target position while ensuring document integrity. When either condition is met, the CPU immediately generates an immediate locking command. This command has the highest priority and will be directly transmitted to the locking mechanism of the stamping machine, simultaneously interrupting all currently unfinished operations. Upon receiving the command, the locking mechanism quickly executes a locking action, firmly locking the stamp fixing device in its initial position, prohibiting any stamping action, and completely blocking the possibility of unauthorized stamping from a mechanical structure perspective.

[0146] S65. Generate a stamping instruction based on the stamping permission instruction, fine-tuning instruction, and immediate stamping lock instruction.

[0147] Optionally, the CPU constructs an instruction priority determination mechanism to prioritize the stamping permission instruction, fine-tuning instruction, and immediate locking instruction. The immediate locking instruction has the highest priority, followed by the fine-tuning instruction, while the stamping permission instruction is a regular execution instruction. First, the corresponding instruction type is matched based on the calculation result of S61. If an immediate locking instruction is matched, it is directly output as the final stamping instruction to ensure a rapid locking response in hazardous scenarios. If no immediate locking instruction is matched but a fine-tuning instruction is matched, the fine-tuning instruction is output as the final stamping instruction, controlling the stamping platform to perform fine-tuning correction. After correction, the system will re-trigger the actual boundary detection process, re-executing the determination logic of S61 to S65 until a stamping permission instruction is generated. If a stamping permission instruction is directly matched, it is output as the final stamping instruction, controlling the stamping execution mechanism to perform the stamping operation. The finally generated stamping instruction is transmitted to the corresponding execution module via an internal high-speed bus, ensuring the real-time performance and accuracy of instruction transmission. Simultaneously, the instruction type, generation time, and corresponding determination data are recorded in the historical task log, providing data traceability for parameter optimization and fault diagnosis in subsequent tasks.

[0148] In the aforementioned intelligent control method for an intelligent stamping machine, the following steps are taken: First, the environmental state data and target boundary data of the current stamping task are acquired. Then, based on these data, a state feature vector representing the current working condition of the stamping machine is constructed. Subsequently, the state feature vector is input into a pre-trained reinforcement learning decision model to jointly optimize multiple adjustable parameters of the infrared sensing system to output the current optimal parameter combination. Next, the infrared sensing system is configured according to the optimal parameter combination and driven to scan and detect the paper boundary to obtain the actual boundary information. Finally, a stamping instruction is generated and the stamping operation is executed based on the actual boundary information and the target boundary data. This coherent method realizes the dynamic adaptation of the infrared sensing system parameters, improves the accuracy of paper boundary detection, and ensures that the stamping operation is performed compliantly only within the target range. This solves the problems of poor environmental adaptability, insufficient boundary detection accuracy, and security risks of stamping in unauthorized areas caused by the fixed parameters of traditional stamping machines, thereby improving the accuracy, security, and intelligence level of stamping.

[0149] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0150] Based on the same inventive concept, this application also provides an intelligent control system for an intelligent stamping machine to implement the intelligent control method of the intelligent stamping machine described above. The solution provided by this system is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more intelligent control system embodiments of intelligent stamping machines provided below can be found in the limitations of the intelligent control method of an intelligent stamping machine described above, and will not be repeated here.

[0151] In one exemplary embodiment, such as Figure 4 As shown, a schematic diagram of the intelligent control system 10 for an intelligent stamping machine is provided, including:

[0152] Data acquisition module 11 is used to acquire environmental status data and target boundary data of the current stamping task;

[0153] The stamping machine status module 12 is used to obtain a status feature vector based on environmental status data and target boundary data; the status feature vector is used to characterize the current working condition of the stamping machine.

[0154] The sensor parameter optimization module 13 is used to input the state feature vector into the pre-trained reinforcement learning decision model, and to jointly optimize multiple adjustable parameters of the infrared sensing system on the stamping machine through the reinforcement learning decision model, and output the current optimal parameter combination.

[0155] The paper actual boundary recognition module 14 is used to configure the working parameters of the infrared sensing system according to the current optimal parameter combination, drive the infrared sensing system to scan and detect the paper boundary, and obtain the actual boundary information of the paper.

[0156] The stamping execution module 15 is used to obtain stamping instructions based on actual boundary information and target boundary data; the stamping instructions are used to instruct the stamping machine to perform the stamping operation and obtain the stamping result.

[0157] Furthermore, the data acquisition module 11 is also used for:

[0158] S11. Control all infrared transmitters in the infrared sensing system to transmit signals at a reference power, and synchronously acquire the raw signal values ​​of all infrared receivers to obtain the reference signal vector.

[0159] S12. Based on the reference signal vector and the pre-calibrated no-load environment reference signal vector, calculate the signal attenuation coefficient vector and background noise intensity vector under the current environment;

[0160] S13. Read the historical task log of the stamping machine, obtain the final decision results and boundary error statistics of historical tasks similar to the current target boundary data, and use them as historical performance reference data.

[0161] S14. Combine the signal attenuation coefficient vector, background noise intensity vector, and historical performance reference data to form environmental status data.

[0162] Furthermore, the stamping machine status module 12 is also used for:

[0163] S21. Normalize the environmental state data to obtain a normalized environmental vector;

[0164] S22. Encode and transform the target boundary data to obtain the boundary encoding vector;

[0165] S23. Based on the normalized environment vector and the boundary coding vector, feature fusion is performed to obtain the fused feature vector;

[0166] S24. The dimensionality of the fused feature vector is transformed by a fully connected layer to obtain the state feature vector.

[0167] Furthermore, the sensor parameter optimization module 13 is also used for:

[0168] S31. Construct the state space of the reinforcement learning decision model. The state vector of the state space includes the state feature vector, the boundary detection error of the previous cycle in real time feedback, and the signal-to-noise ratio of the sensor signal.

[0169] S32. Construct the action space of the reinforcement learning decision model. The action vector of the action space is defined as the adjustment amount of multiple adjustable parameters in the infrared sensing system. The adjustable parameters include the transmission power of each infrared transmitter, the gain threshold of each infrared receiver, and the signal sampling frequency.

[0170] S33. Define the reward function of the reinforcement learning decision model. The reward function calculates the instant reward based on the degree of agreement between the actual boundary information generated after the action is executed and the target boundary data, as well as the sensor resources consumed to achieve the agreement.

[0171] S34. Use the proximal policy optimization algorithm to train the reinforcement learning decision model offline until the model can stably output parameter adjustment actions that maximize the reward function in response to simulated state inputs, thus completing the pre-training.

[0172] Furthermore, the paper actual boundary recognition module 14 is also used for:

[0173] S41. Based on the current optimal parameter combination, set the modulation frequency and phase of each infrared transmitter in the infrared sensing system to transmit pulse signals in a time-division manner according to a preset time sequence;

[0174] S42. During each transmission pulse cycle, the response signal strength and transit time of the corresponding infrared receiver are synchronously acquired to obtain multi-channel signal strength and time series data.

[0175] S43. Filter and perform time-frequency analysis on the multi-channel signal strength and time series data to calculate the occlusion position of the paper edge on each independent infrared detection path and obtain the occlusion calculation result.

[0176] S44. By fusing the occlusion calculation results of all infrared detection paths, the complete actual boundary contour coordinate sequence of the paper is reconstructed through geometric calculation, and the actual boundary information of the paper is obtained.

[0177] Furthermore, the paper actual boundary recognition module 14 is also used for:

[0178] S51. Filter the received signal of each channel to obtain the filtered signal;

[0179] S52. Extract the peak intensity and the time to reach the peak value of the filtered signal envelope, and calculate the flight time of the signal by combining the known signal transmission start time.

[0180] S53. Construct an obstruction location calculation model based on the filtered signal and the signal's time of flight. The obstruction location calculation model is used to map the received signal features to the physical obstruction distance to obtain the obstruction location.

[0181] S54. A Kalman filter is used to filter and track the occlusion positions of each channel calculated within a continuous sampling period, and the real-time occlusion position sequence is output to obtain the occlusion calculation result.

[0182] Furthermore, the stamping permission module 15 is also used for:

[0183] S61. Calculate the overlap between the contour reconstructed from the actual boundary information and the region defined by the target boundary data, and the set of shortest distances from each point of the actual contour to the boundary of the target region.

[0184] S62. When the overlap is 100% and all values ​​in the shortest distance set are greater than the first security threshold, generate a stamping permission instruction.

[0185] S63. When the overlap is less than 100% and greater than the minimum allowable threshold, and the maximum value in the shortest distance set is less than zero but the absolute value is less than the second fault tolerance threshold, a fine-tuning instruction is generated; the fine-tuning instruction is used to control the micro-motion of the stamping platform to push the paper to the target area.

[0186] S64. When the overlap is less than the minimum allowable threshold, or the minimum value in the shortest distance set is less than a negative locking threshold, generate an immediate lock command.

[0187] S65. Generate a stamping instruction based on the stamping permission instruction, fine-tuning instruction, and immediate stamping lock instruction.

[0188] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the intelligent control method for an intelligent stamping machine as described above.

[0189] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0190] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0191] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. An intelligent control method for an intelligent stamping machine, characterized in that, The method includes: S1. Obtain the environmental status data and target boundary data of the current stamping task; S2. Based on the environmental state data and the target boundary data, a state feature vector is obtained; the state feature vector is used to characterize the current operating condition of the stamping machine. S3. Input the state feature vector into the pre-trained reinforcement learning decision model, and use the reinforcement learning decision model to jointly optimize multiple adjustable parameters of the infrared sensing system on the stamping machine, and output the current optimal parameter combination. S4. Configure the working parameters of the infrared sensing system according to the current optimal parameter combination, and drive the infrared sensing system to scan and detect the paper boundary to obtain the actual boundary information of the paper. S5. Based on the actual boundary information and the target boundary data, a stamping instruction is obtained; the stamping instruction is used to instruct the stamping machine to perform a stamping operation to obtain a stamping result.

2. The method according to claim 1, characterized in that, The acquisition of environmental state data and target boundary data for the current stamping task includes: S11. Control all infrared transmitters in the infrared sensing system to transmit signals at a reference power, and synchronously acquire the raw signal values ​​of all infrared receivers to obtain a reference signal vector. S12. Calculate the signal attenuation coefficient vector and background noise intensity vector under the current environment based on the reference signal vector and the pre-calibrated no-load environment reference signal vector. S13. Read the historical task log of the stamping machine, obtain the final decision results and boundary error statistics of historical tasks similar to the current target boundary data, and use them as historical performance reference data. S14. The signal attenuation coefficient vector, the background noise intensity vector, and the historical performance reference data are combined to form the environmental state data.

3. The method according to claim 2, characterized in that, The process of obtaining a state feature vector based on the environmental state data and the target boundary data includes: S21. Normalize the environmental state data to obtain a normalized environmental vector; S22. Encode and convert the target boundary data to obtain a boundary encoding vector; S23. Based on the normalized environment vector and the boundary encoding vector, feature fusion is performed to obtain a fused feature vector; S24. The fusion feature vector is transformed in dimension through a fully connected layer to obtain the state feature vector.

4. The method according to claim 1, characterized in that, The pre-training process of the reinforcement learning decision model includes: S31. Construct the state space of the reinforcement learning decision model, wherein the state vector of the state space includes the state feature vector, the boundary detection error of the previous cycle in real time feedback, and the signal-to-noise ratio of the sensor signal; S32. Construct the action space of the reinforcement learning decision model. The action vector of the action space is defined as the adjustment amount of multiple adjustable parameters in the infrared sensing system. The adjustable parameters include the transmission power of each infrared transmitter, the gain threshold of each infrared receiver, and the signal sampling frequency. S33. Define the reward function of the reinforcement learning decision model. The reward function calculates the instant reward based on the degree of match between the actual boundary information generated after the action is executed and the target boundary data, as well as the sensor resources consumed to achieve the degree of match. S34. The reinforcement learning decision model is trained offline using the proximal policy optimization algorithm until the model can stably output parameter adjustment actions that maximize the reward function in response to simulated state inputs, thus completing the pre-training.

5. The method according to claim 4, characterized in that, The step of configuring the operating parameters of the infrared sensing system according to the current optimal parameter combination, and driving the infrared sensing system to scan and detect the paper boundary to obtain the actual boundary information of the paper, includes: S41. Based on the current optimal parameter combination, the modulation frequency and phase of each infrared transmitter in the infrared sensing system are set to transmit pulse signals in a time-division manner according to a preset time sequence; S42. During each transmission pulse cycle, the response signal strength and transit time of the corresponding infrared receiver are synchronously acquired to obtain multi-channel signal strength and time series data. S43. Filter and perform time-frequency analysis on the multi-channel signal strength and the time series data to calculate the occlusion position of the paper edge on each independent infrared detection path and obtain the occlusion calculation result. S44. By fusing the occlusion calculation results of all infrared detection paths, the complete actual boundary contour coordinate sequence of the paper is reconstructed through geometric calculation to obtain the actual boundary information of the paper.

6. The method according to claim 5, characterized in that, The filtering and time-frequency analysis of the multi-channel signal intensity and the time series data are performed to calculate the occlusion position of the paper edge on each independent infrared detection path, obtaining the occlusion calculation result, including: S51. Filter the received signal of each channel to obtain the filtered signal; S52. Extract the peak intensity and the time to reach the peak value of the filtered signal envelope, and calculate the flight time of the signal by combining it with the known signal transmission start time. S53. Construct an obstruction location calculation model based on the filtered signal and the time of flight of the signal. The obstruction location calculation model is used to map the characteristics of the received signal into physical obstruction distance to obtain the obstruction location. S54. A Kalman filter is used to filter and track the occlusion positions of each channel calculated within a continuous sampling period, and a real-time occlusion position sequence is output to obtain the occlusion calculation result.

7. The method according to claim 6, characterized in that, The step of obtaining the stamping instruction based on the actual boundary information and the target boundary data includes: S61. Calculate the overlap between the contour reconstructed from the actual boundary information and the region defined by the target boundary data, and the set of shortest distances from each point of the actual contour to the boundary of the target region. S62. When the overlap is 100% and all values ​​in the shortest distance set are greater than the first security threshold, a stamping permission instruction is generated. S63. When the overlap is less than 100% and greater than the minimum allowable threshold, and the maximum value in the shortest distance set is less than zero but the absolute value is less than the second fault tolerance threshold, a fine-tuning instruction is generated; the fine-tuning instruction is used to control the stamping platform to move the paper to the target area. S64. When the overlap is less than the minimum allowable threshold, or the minimum value in the shortest distance set is less than a negative locking threshold, generate an immediate locking command. S65. Generate the stamping instruction based on the permission stamping instruction, the fine-tuning instruction, and the immediate stamping instruction.

8. An intelligent control system for an intelligent stamping machine, characterized in that, The system includes: The data acquisition module is used to acquire environmental status data and target boundary data for the current stamping task; The stamping machine status module is used to obtain a status feature vector based on the environmental status data and the target boundary data; the status feature vector is used to characterize the current working condition of the stamping machine. The sensor parameter optimization module is used to input the state feature vector into a pre-trained reinforcement learning decision model, and to perform joint optimization decision on multiple adjustable parameters of the infrared sensing system on the stamping machine through the reinforcement learning decision model, and output the current optimal parameter combination. The paper actual boundary recognition module is used to configure the working parameters of the infrared sensing system according to the current optimal parameter combination, drive the infrared sensing system to scan and detect the paper boundary, and obtain the actual boundary information of the paper. The stamping permission execution module is used to obtain a stamping instruction based on the actual boundary information and the target boundary data; the stamping instruction is used to instruct the stamping machine to perform a stamping operation to obtain a stamping result.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images