An acousto-optic Q-switching method for a carbon dioxide laser

Abstract

This invention discloses an acousto-optic Q-switching method for carbon dioxide lasers, specifically relating to the field of acousto-optic Q-switching of lasers. It addresses the problem that existing Q-switching control methods rely on fixed delays or single feedback, making it difficult to adapt to changes in discharge state, resulting in large pulse energy fluctuations and poor stability. By acquiring and segmenting the voltage and current waveforms within each discharge cycle, multi-dimensional features of the pre-ionization, main discharge, and relaxation stages are extracted to construct a full-cycle state vector. Based on this state vector, a timing prediction network is used to output the Q-switching turn-on delay and acousto-optic diffraction efficiency setpoints for the next cycle, achieving dynamic prediction and adaptive adjustment of Q-switching parameters. This improves the stability and consistency of laser pulse output, enhances peak power control accuracy and adaptability under different operating conditions, while reducing the reliance on manual experience for parameter tuning and improving the long-term reliability and robustness of the system.

Description

Technical Field

[0001] This invention relates to the field of acousto-optic Q-switching technology for lasers, and more specifically, to an acousto-optic Q-switching method for carbon dioxide lasers. Background Technology

[0002] In applications such as industrial material processing, medical device manufacturing, and precision micromachining, CO2 laser gas discharge is widely used in processes such as cutting, welding, marking, and surface treatment due to its moderate wavelength, high output power, and stable beam quality. To obtain high peak power and narrow pulse width output, acousto-optic Q-switching technology is typically used to rapidly modulate the laser resonator, allowing the laser energy to be released in a concentrated manner after sufficient accumulation of particles in the upper energy level.

[0003] However, under high repetition frequency and complex operating conditions, the gas discharge process has obvious periodic fluctuation characteristics. There are differences in ignition conditions, discharge stability and relaxation recovery between different discharge cycles. This makes it difficult for a single fixed Q switch turn-on delay or constant diffraction efficiency setting to adapt to the dynamically changing discharge state, thus causing problems such as output pulse energy fluctuation, peak power reduction and pulse width inconsistency.

[0004] Existing technologies are mostly based on empirical settings or simple feedback adjustment methods. The characterization of the discharge process is usually limited to a single stage or a single parameter, lacking the ability to jointly model the entire process of pre-ionization, main discharge, and relaxation, making it difficult to accurately reflect the comprehensive impact of plasma evolution on laser output. Under continuous multi-cycle operation, the uncertainty of the discharge process is further aggravated by insufficient relaxation and the accumulation of thermal effects in the previous cycle, making it difficult to achieve stable optimization of Q-switching control. There is an urgent need for a technical solution that can combine the characteristics of multi-stage discharge and dynamically predict and adjust the Q-switching parameters for the next cycle. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide an acousto-optic Q-switching method for carbon dioxide lasers to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] An acousto-optic Q-switching method for carbon dioxide lasers includes the following steps:

[0008] S1. At the rising edge of the pump power supply trigger pulse, capture the pump voltage waveform and current waveform at a sampling rate higher than the discharge oscillation frequency. With the voltage collapse point as the time zero point, align and segment the waveforms to store them as two time waveform pairs: the pre-ionization segment and the main discharge segment.

[0009] S2. In the pre-ionization stage, obtain the electric field establishment rate characterization value from the voltage waveform and the current ramp-up rate from the current waveform. Combine the electric field establishment rate characterization value and the current ramp-up rate into an ignition feature vector.

[0010] S3. In the main discharge stage, when the current waveform shows the first peak point, the current waveform is decomposed in the time-frequency domain to extract the characteristic harmonic energy ratio and the time-domain fluctuation rate. At the same time, the voltage waveform is integrated and the corresponding second derivative is calculated as a characterization of the plasma conductivity state change. The characteristic harmonic energy ratio, the time-domain fluctuation rate and the characterization of conductivity state change are combined into a stability feature vector.

[0011] S4. During the relaxation phase after the main discharge phase ends, the relaxation feature vector is extracted from the slope of the current waveform envelope and the recovery rate of the voltage waveform.

[0012] S5. Concatenate the ignition feature vector, stability feature vector and relaxation feature vector into a state vector, input it into the pre-trained temporal prediction network, and output the Q-switch turn-on delay and acousto-optic diffraction efficiency setting value for the next control cycle.

[0013] S6. Taking the end of the current discharge cycle as the starting point of the timing, press the Q switch of the next control cycle to start the delay start timer. After the timing is reached, adjust the amplitude of the radio frequency drive signal according to the acousto-optic diffraction efficiency setting value and apply it to the acousto-optic transducer.

[0014] As a further aspect of the present invention, in S1, storing the waveform in segments as two time-series waveform pairs—a pre-ionization segment and a main discharge segment—specifically includes:

[0015] When the rising edge of the pump power trigger pulse arrives, the pump voltage waveform and pump current waveform are continuously acquired, and the first derivative of the voltage waveform is monitored in real time. When the first derivative of the voltage shows a negative jump and the jump amplitude exceeds the preset background noise threshold, this moment is taken as the time zero point corresponding to the voltage collapse point.

[0016] The waveform segment of a set duration is intercepted forward from the zero point of time as the pre-ionization segment waveform pair, and the waveform segment intercepted backward from the zero point of time to the end of the pump discharge is used as the main discharge segment waveform pair.

[0017] Align the voltage and current waveforms within the pre-ionization phase waveform pair on the time axis, and align the voltage and current waveforms within the main discharge phase waveform pair on the time axis.

[0018] As a further aspect of the present invention, in step S2, combining the electric field establishment rate characterization value and the current ramp-up rate into an ignition feature vector specifically includes:

[0019] The voltage waveform is retrieved from the pre-ionization section waveform pair. The voltage waveform is then calculated on the time axis according to the sampling order. The ratio of the voltage change between adjacent sampling points to the corresponding time interval is obtained, and the maximum voltage change rate is used as the characterization value of the electric field establishment rate.

[0020] The current waveform is retrieved from the pre-ionization waveform pair. The first zero-crossing point of the current waveform in the pre-ionization section is detected. The rising edge segment of the current waveform from zero value to the first peak value is intercepted along the time axis starting from the zero-crossing point. The ratio of the difference between the current value at the end of the rising edge segment and the current value at the beginning of the rising edge segment to the time span from the beginning to the end is calculated as the current ramp-up rate.

[0021] The electric field establishment rate and current ramp-up rate are arranged in a preset dimension order to form an ignition feature vector.

[0022] As a further aspect of the present invention, in step S3, combining the characteristic harmonic energy ratio, time-domain volatility, and conductivity state change characterization quantity into a stability feature vector specifically includes:

[0023] Within the main discharge waveform pair, when the current waveform first reaches a peak, a waveform segment of a set duration is extracted from the peak of the current waveform as an analysis window. A sliding windowed short-time Fourier transform is performed on the current waveform within the analysis window. The spectral amplitudes within each time window are accumulated according to frequency to calculate the proportion of energy in the discharge characteristic frequency band to the total energy of the analysis window, thus obtaining the characteristic harmonic energy proportion. The standard deviation of this characteristic harmonic energy proportion between each time window within the analysis window is then calculated to obtain the time-domain volatility.

[0024] The voltage waveform is retrieved from the main discharge segment waveform. Starting from the voltage collapse point, the voltage value is accumulated point by point and integrated to obtain the injected energy accumulation sequence. Then, a second-order difference operation is performed. The maximum absolute value of the second-order difference sequence is used as the characteristic quantity of the plasma conductivity state change.

[0025] The characteristic harmonic energy ratio, time-domain volatility, and conductivity state change parameters are arranged in a predetermined dimension order to form a stability feature vector.

[0026] As a further aspect of the present invention, in step S4, extracting the relaxation feature vector from the slope of the current waveform envelope and the recovery rate of the voltage waveform specifically includes:

[0027] After the main discharge phase ends, the moment when the detected current waveform drops from its peak point to the first time below the preset background noise threshold is taken as the discharge extinction point, and a waveform segment of a set duration is extracted as the relaxation analysis window.

[0028] Extract the peak amplitudes from the current waveform within the relaxation analysis window, take the natural logarithm of each peak amplitude, and then perform linear fitting in time order. Take the absolute value of the slope obtained from the fitting as the composite decay rate.

[0029] Extract the rising segment of the voltage waveform from the relaxation analysis window, perform linear fitting on the voltage amplitude in the rising segment in time sequence, and use the slope obtained by fitting as the electric field recovery rate.

[0030] The composite decay rate and the electric field recovery rate are arranged in a predetermined order according to a preset dimension to form a relaxation feature vector.

[0031] As a further aspect of the present invention, in step S5, the training method of the pre-trained temporal prediction network is as follows:

[0032] Within multiple consecutive test discharge cycles, diffraction efficiency constraints are set and the Q-switch turn-on delay parameters are traversed in a stepwise manner. The actual output pulse energy value is collected and optimized. The marginal gain saturation point corresponding to each discharge cycle is determined as the optimal reference point. The optimal reference point of each discharge cycle is associated with the corresponding state vector and stored in the sample database.

[0033] The discharge cycle includes a pre-ionization phase, a main discharge phase, and a relaxation phase. The state vector is a combination vector of the fusion ignition feature vector, the stability feature vector, and the relaxation feature vector.

[0034] State vectors from several consecutive test discharge cycles are extracted from the sample database in chronological order to form a state sequence. The state sequence is used as input, and the Q-switch turn-on delay and acousto-optic diffraction efficiency setpoints corresponding to the optimal reference point are used as output labels to train a time-series prediction network constructed from convolutional layers and gated recurrent units.

[0035] As a further aspect of the present invention, in step S5, the Q-switch turn-on delay and the acousto-optic diffraction efficiency setting value for the next control cycle specifically include:

[0036] The state vectors of continuous discharge cycles are integrated into a state sequence that conforms to the input standard of the pre-trained time-series prediction network. The corresponding control period is set based on the discharge cycle. The output of the time-series prediction network is used as the Q-switch turn-on delay and acousto-optic diffraction efficiency setting value for the next control period.

[0037] As a further aspect of the present invention, in S6, activating the delay start timer by pressing the Q switch of the next control cycle, and adjusting the amplitude of the radio frequency drive signal according to the acousto-optic diffraction efficiency setting value and applying it to the acousto-optic transducer after the timer expires specifically includes:

[0038] The system obtains the Q-switch turn-on delay and acousto-optic diffraction efficiency setting value for the next control cycle output by the timing prediction network. It loads the turn-on delay count value and starts the count-down process with the end time of the current discharge cycle as the timing start point. When the count-down reaches zero, it adjusts the programmable gain amplifier coefficient of the RF drive circuit according to the acousto-optic diffraction efficiency setting value and outputs the adjusted RF drive signal to the acousto-optic transducer.

[0039] The technical effects and advantages of the acousto-optic Q-switching method for carbon dioxide lasers of the present invention are as follows:

[0040] This invention employs structured modeling of the entire discharge process, unifying and integrating the ignition characteristics of the pre-ionization stage, the stability characteristics of the main discharge stage, and the recovery characteristics of the relaxation stage to construct a full-cycle state vector. Based on this state vector, it predictively controls the Q-switching turn-on delay and acousto-optic diffraction efficiency for the next discharge cycle. Compared to traditional fixed-parameter or single-feedback adjustment methods, this approach accurately reflects the impact of plasma evolution on laser output, achieving dynamic adaptive optimization of Q-switching parameters. By introducing the marginal gain saturation point as the optimal reference point, it avoids the instability caused by simply using the maximum pulse energy as the optimization target, improving the robustness and repeatability of parameter selection. Furthermore, by using a time-series prediction network to model the correlation between consecutive discharge cycles, the control strategy possesses cross-cycle memory capability, effectively suppressing output fluctuations caused by insufficient relaxation and thermal effect accumulation under high repetition frequency conditions. This results in improved pulse energy stability, enhanced peak power consistency, and improved pulse width controllability. Simultaneously, this invention achieves control optimization through deep utilization of existing voltage and current signals without adding complex hardware structures, offering advantages such as low implementation cost, strong adaptability, and high engineering promotion value. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of an acousto-optic Q-switching method for a carbon dioxide laser according to the present invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0043] Example 1

[0044] Figure 1 The present invention provides an acousto-optic Q-switching method for a carbon dioxide laser, comprising the following steps:

[0045] S1. At the rising edge of the pump power supply trigger pulse, capture the pump voltage waveform and current waveform at a sampling rate higher than the discharge oscillation frequency. With the voltage collapse point as the time zero point, align and segment the waveforms to store them as two time waveform pairs: the pre-ionization segment and the main discharge segment.

[0046] S2. In the pre-ionization stage, obtain the electric field establishment rate characterization value from the voltage waveform and the current ramp-up rate from the current waveform. Combine the electric field establishment rate characterization value and the current ramp-up rate into an ignition feature vector.

[0047] S3. In the main discharge stage, when the current waveform shows the first peak point, the current waveform is decomposed in the time-frequency domain to extract the characteristic harmonic energy ratio and the time-domain fluctuation rate. At the same time, the voltage waveform is integrated and the corresponding second derivative is calculated as a characterization of the plasma conductivity state change. The characteristic harmonic energy ratio, the time-domain fluctuation rate and the characterization of conductivity state change are combined into a stability feature vector.

[0048] S4. During the relaxation phase after the main discharge phase ends, the relaxation feature vector is extracted from the slope of the current waveform envelope and the recovery rate of the voltage waveform.

[0049] S5. Concatenate the ignition feature vector, stability feature vector and relaxation feature vector into a state vector, input it into the pre-trained temporal prediction network, and output the Q-switch turn-on delay and acousto-optic diffraction efficiency setting value for the next control cycle.

[0050] S6. Taking the end of the current discharge cycle as the starting point of the timing, press the Q switch of the next control cycle to start the delay start timer. After the timing is reached, adjust the amplitude of the radio frequency drive signal according to the acousto-optic diffraction efficiency setting value and apply it to the acousto-optic transducer.

[0051] In step S1, the waveform is aligned and segmented and stored as two time-series waveform pairs: a pre-ionization segment and a main discharge segment.

[0052] At the instant the rising edge of the pump power supply trigger pulse arrives, continuous waveform data of the pump voltage and current are synchronously acquired through a high-voltage probe and a current sampling element. The sampling frequency is set to at least five times the discharge oscillation frequency. For example, when the typical oscillation frequency is 1MHz, the sampling frequency is selected in the range of 5MHz to 20MHz to ensure complete capture of voltage change details. Before the acquisition begins, a stable, discharge-free background time window is selected before the arrival of the trigger pulse. The voltage waveform within this window is statistically analyzed to calculate its amplitude fluctuation range. Specifically, the absolute value sequence of the difference between consecutive sampling points is used, and the maximum value is taken as the background noise amplitude reference value. This value is then amplified by 2 to 3 times as the noise judgment threshold. For example, when the maximum background fluctuation is 0.5V, the threshold is set to 1.5V. Subsequently, during real-time acquisition, the change between adjacent sampling points is calculated point by point according to the sampling sequence of the voltage waveform. The change trend is obtained by combining the sampling time interval. When the voltage change changes from stable to a rapid decrease, and the change of several consecutive sampling points exceeds the threshold and remains in the same direction, this position is determined to be the voltage collapse initiation point. To avoid false triggering by single-point spike noise, the criterion is that at least three consecutive sampling points simultaneously meet the condition, thus ensuring stable and reliable detection results. The sampling time corresponding to the starting point of the voltage collapse is uniformly defined as the zero point, and the time coordinates of all subsequent waveform data are recalibrated so that the time value at this point is zero, with the time before and after it extended in both positive and negative directions with this point as a reference.

[0053] After the zero-point calibration is completed, the acquired voltage and current waveforms are segmented around this zero-point. First, the pre-ionization segment's range is determined. A waveform of a fixed duration before the zero-point is selected as the pre-ionization segment. This duration is determined based on the equipment's operating characteristics, for example, a fixed value within the range of 5 to 20 microseconds. In practice, 10 microseconds can be used as an example value. The number of sampling points corresponding to this duration is traced back from the zero-point, and the voltage and current waveforms within the corresponding interval are extracted to form the pre-ionization segment waveform pair. Next, the main discharge segment's range is determined, extending backward from the zero-point until the current waveform falls back to near the background level and enters a stable range. This stable range is determined by comparing it with the aforementioned background noise window. When the current amplitude is within the background fluctuation range for several consecutive sampling points and no longer shows a significant upward trend, the main discharge is considered to have ended, thus determining the termination position of the main discharge segment and avoiding the mistaken inclusion of the relaxation segment within the main discharge segment. After segmentation, the voltage and current waveforms within each segment undergo sample-level time alignment. Due to slight delay differences in the sampling of the two signals, the data is first resampled according to a unified sampling clock to ensure consistent sampling point positions for both signals at the same time scale. Then, using the zero point of time as the alignment reference, the two waveforms are synchronously shifted to ensure strict correspondence between voltage and current values ​​corresponding to the same timestamp. For any missing points during sampling, linear interpolation between adjacent sampling points is used to fill in the gaps, ensuring sequence continuity.

[0054] In S2, the electric field establishment rate characterization value and the current ramp-up rate are combined to form the ignition feature vector.

[0055] Voltage waveform data is retrieved from the pre-ionization phase waveform pair, and the voltage sequence is processed point-by-point according to a unified sampling time sequence. The voltage waveform is lightly smoothed to suppress the amplified effect of sampling noise on the differential results of adjacent points. A moving average method is used, where the average of the current sampling point and the sampling points before and after it is used as the smoothed value for that point, thus ensuring the stability of subsequent rate of change calculations. On the smoothed voltage sequence, the voltage change between adjacent sampling points is calculated point-by-point according to the sampling order, and the voltage rate of change sequence is obtained by combining the corresponding sampling time intervals. The sampling time interval is derived from a unified clock trigger, and its value is fixed during the system initialization phase. For example, at a sampling frequency of 10MHz, the sampling interval is 0.1 microseconds. This time interval remains consistent throughout the pre-ionization phase, thus ensuring a uniform scale for the rate of change calculation. The obtained voltage rate of change sequence is then iterated, and the maximum value of the rate of change is found within the interval from the start position of the pre-ionization phase to the zero time point, serving as the characterization value for the electric field establishment rate. To avoid occasional spikes in noise causing abnormally high maximum values, a continuous judgment condition is set during the search process. That is, a value is only considered a valid candidate value if it maintains a high level within at least two adjacent sampling points; if a value is an isolated spike, it is discarded. Furthermore, the rate of change sequence is sorted, and the average of the first few larger values ​​is selected as a reference for the final representation value. For example, the average of the first three maximum values ​​is used to enhance anti-interference capability.

[0056] The corresponding current waveform data is retrieved from the same pre-ionization waveform pair, and baseline correction is performed on the current waveform. Specifically, a section with no significant discharge current is selected at the beginning of the pre-ionization section, and the average current value within this section is calculated as the baseline bias. This baseline bias is then subtracted from the entire current waveform to eliminate the influence of sensor zero drift. Subsequently, zero-crossing detection is performed on the corrected current waveform using a sign change determination method. When the current value at a sampling point is negative and the next sampling point is positive, and both absolute values ​​exceed the baseline noise range, that position is determined to be a valid zero-crossing. The baseline noise range is obtained by statistically analyzing current fluctuations in the non-discharge section; for example, twice the maximum fluctuation value is used as the determination threshold to avoid false zero-crossings caused by noise. After determining the zero-crossing point, the current rise segment is truncated along the time axis starting from that point until the first significant peak value appears. Peak value determination employs a local maximum detection method, where a current value at a sampling point is considered a candidate peak if it exceeds the values ​​of the two sampling points before and after it. Simultaneously, this peak value must be several times the amplitude of the zero-crossing current, for example, more than three times, to eliminate false positives caused by minor fluctuations. After determining the range of the rising segment, the difference between the current value at the end of the segment and the current value at the beginning is calculated and divided by the corresponding time span. The time span is obtained by multiplying the number of sampling points by the sampling interval, thus yielding the current ramp-up rate. To improve stability, the ramp-up rates of multiple adjacent sub-intervals within the rising segment are statistically analyzed in actual calculations, and their average value is taken as the final characterization value, avoiding the influence of single-endpoint errors. Finally, the aforementioned electric field establishment rate characterization value and the current ramp-up rate are combined in a fixed order, for example, using the electric field establishment rate as the first dimension and the current ramp-up rate as the second dimension, to form a two-dimensional ignition feature vector.

[0057] In S3, the characteristic harmonic energy ratio, time-domain volatility, and conductivity state change characterization are combined into a stability feature vector.

[0058] After retrieving the current waveform from the main discharge waveform, the current data is preprocessed to eliminate the impact of sampling noise on subsequent frequency domain analysis. The processing method involves smoothing the current waveform using a fixed-length sliding window. The window length is selected as an odd number of sampling points (5 to 9 points), for example, 7 sampling points are selected as example values. The average value within the window is taken as the smoothed value for the current point, avoiding instantaneous spike interference. Subsequently, peak identification is performed on the smoothed current waveform using a local maximum detection method. When the current value at a certain sampling point is simultaneously greater than the values ​​of at least two sampling points before and after it, and the current value at that point exceeds a set proportion (e.g., 1.2 times) of the average current amplitude of the main discharge stage, that point is identified as a valid peak point. To avoid mistakenly identifying small local peaks in the oscillations as main peaks, the time interval between the peak point and the previous candidate peak is further required to be greater than a set minimum interval (e.g., 1 microsecond), thus ensuring that the first significant peak of the main discharge stage is identified. After determining the peak point, a waveform segment of fixed duration is extracted from that point as an analysis window. This duration is set according to the discharge characteristics, for example, a fixed value within the range of 10 to 30 microseconds, with 20 microseconds selected as an example value. The corresponding number of sampling points is calculated from the sampling frequency. After extracting the analysis window, a sliding windowed short-time Fourier transform is performed on the current waveform within that window. The sliding window length is selected as a fixed number of sampling points, such as 128 or 256 points. A Hamming window is used as the window function to reduce spectral leakage, and the sliding step size is set to half the window length to ensure a balance between time resolution and frequency resolution. Within each time window, the obtained spectral amplitude is accumulated according to frequency, and the energy proportion within that frequency band is calculated in conjunction with a pre-determined discharge characteristic frequency band. The discharge characteristic frequency band is obtained by statistically analyzing multiple test discharge experimental data. For example, if there is a significant energy concentration in the range of 10 kHz to 200 kHz, then that range is set as the characteristic frequency band. Subsequently, the proportion of characteristic harmonic energy corresponding to each time window is statistically analyzed within the entire analysis window range, and its standard deviation is calculated as the time-domain volatility.

[0059] After retrieving the corresponding voltage waveform from the main discharge waveform, the voltage data is integrated point-by-point, starting from the zero point of time, i.e., the voltage collapse point. Since the sampling process uses a unified clock trigger and the sampling interval is a fixed value, for example, 0.1 microseconds at a 10MHz sampling frequency, the voltage value at each sampling point is multiplied by this fixed time interval and accumulated point by point during integration to obtain a time-increasing sequence of injected energy accumulation. To reduce the cumulative amplification of the integration result by noise, the voltage waveform is lightly filtered before integration using a moving average with a length of 5 sampling points to keep the integration input data stable. After constructing the energy accumulation sequence, a second-order difference operation is performed on the sequence. This involves first calculating the difference sequence between adjacent sampling points, and then recalculating the adjacent differences to obtain a sequence reflecting the curvature of energy change. Because the second-order difference is sensitive to noise, a continuous three-point averaging method is used to smooth the difference result during the calculation to suppress high-frequency noise interference. Subsequently, the second-order difference sequence is traversed throughout the entire main discharge segment, and the maximum absolute value is taken as the characterization quantity of the plasma conductivity state change. This maximum value corresponds to the moment of most drastic energy change, reflecting the degree of abrupt change in conductivity characteristics during the discharge process. To prevent single-point anomalies from causing result deviations, the average of the first few larger values ​​(e.g., the first three maximum values) is selected as the final characterization value. Finally, the obtained characteristic harmonic energy ratio and time-domain fluctuation rate are combined with the obtained conductivity state change characterization quantity in a fixed order, for example, arranged sequentially as the first dimension, the second dimension, and the third dimension, to form a three-dimensional stability feature vector. This arrangement order is determined and maintained during system initialization, thereby ensuring the uniformity of the feature vector structure across different discharge cycles.

[0060] In step S4, relaxation feature vectors are extracted from the slope of the current waveform envelope and the recovery rate of the voltage waveform.

[0061] To identify the discharge extinction point from the current waveform, a background time interval with no obvious current flow is selected before the discharge begins. The current waveform within this interval is statistically analyzed to obtain its maximum fluctuation amplitude. Twice this maximum fluctuation value is used as the background noise threshold. For example, if the maximum background fluctuation is 0.02A, the threshold is set to 0.04A to ensure that the judgment process is not affected by random noise. In actual detection, the current waveform is traversed along the time axis from the peak point of the main discharge segment. When the current amplitude at at least five consecutive sampling points is lower than the noise threshold and its trend no longer shows an upward trend, this moment is determined as the discharge extinction point, thus avoiding misjudgments caused by single points or brief drops. After determining the discharge extinction point, a waveform segment of a fixed duration is extracted from this point as a relaxation analysis window. The duration is determined according to the actual operating frequency of the equipment. For example, under the condition of a repetition frequency of 10kHz, 20 microseconds is selected as the analysis window length so that the window covers the complete decay process. Subsequently, peak extraction was performed on the current waveform within the analysis window using a local maximum detection method. This method required that the current value at a given sampling point be greater than that of at least two sampling points before and after it, and that this value exceed 1.5 times the noise threshold, thus filtering out low-amplitude oscillations. After extracting each peak, its amplitude was recorded in chronological order, and the natural logarithm of each amplitude was taken to transform the exponential decay process into a linear change process. Linear fitting was then performed on the processed sequence using a least-squares method. A straight line was obtained by fitting all peak points, and the slope of this line reflects the rate of current amplitude decay. To improve fitting stability, cases with insufficient or unevenly distributed peaks were screened, requiring at least four valid peak points for fitting.

[0062] The voltage waveform is synchronously retrieved within the relaxation analysis window to analyze the voltage recovery process. First, baseline correction is performed on the voltage waveform. The average voltage value within the stable interval before the start of discharge is calculated as the reference baseline, and the entire voltage sequence is biased to eliminate the influence of measurement errors. Then, the lowest point of the voltage waveform is identified within the relaxation analysis window. This point typically appears shortly after the end of discharge. By traversing all sampling points within this window, the position corresponding to the minimum voltage value is selected as the starting point. Starting from this starting point, the voltage change trend is detected forward along the time axis. When the voltage shows a stable increase without significant decline at several consecutive sampling points, this interval is defined as the voltage rise segment. The termination position is set at the moment the voltage enters the stable plateau region, i.e., the moment when the change amplitude of multiple consecutive sampling points is below the background noise range. To ensure the effectiveness of the rise segment, its length must be no less than a preset minimum time length, for example, no less than 5 microseconds. After determining the rise segment, the voltage amplitude within this segment is linearly fitted in time sequence. The fitting method also uses the least squares method. The slope of the fitted line is calculated using all sampling points, and this slope represents the electric field recovery rate. To avoid the influence of local fluctuations on the results, the voltage sequence was smoothed using a three-point moving average before fitting. Finally, the obtained composite decay rate and electric field recovery rate were combined in a fixed order, with the composite decay rate as the first dimension and the electric field recovery rate as the second dimension, to form a two-dimensional relaxation feature vector.

[0063] In S5, the training method for the pre-trained temporal prediction network.

[0064] In multiple consecutive test discharge cycles, the Q-switch turn-on delay parameters are systematically scanned to establish the mapping relationship between the state vector and the optimal control parameters. First, the acousto-optic diffraction efficiency setpoint is constrained during the testing phase, specifically using a graded fixing method. The diffraction efficiency is divided into several discrete levels, for example, 30%, 50%, and 70% are selected as example values. In each scan, one diffraction efficiency level is fixed, ensuring the delay parameter traversal is performed under a single efficiency condition, thus avoiding the complexity caused by two-dimensional parameter coupling. Subsequently, under this fixed efficiency condition, the Q-switch turn-on delay parameters are traversed in a step-by-step manner. The step interval is set according to the device response accuracy, for example, 1 microsecond or 2 microseconds are selected as the step unit, gradually increasing from the minimum delay value to the set maximum delay range. At each delay level, the laser output pulse energy value is continuously collected for multiple discharge cycles, for example, 5 to 10 cycles, and the average value is taken as the energy characterization for that delay level to eliminate the influence of single-cycle fluctuations. A discrete response sequence is constructed by assigning delay levels and their corresponding average pulse energies. The energy increment between adjacent levels is calculated point-by-point along the direction of delay increase, and the continuity of this increment sequence is assessed. When the energy increment of several consecutive delay levels (e.g., three consecutive levels) is lower than a set gain threshold, the delay level where this condition first occurs is identified as the marginal gain saturation point. This gain threshold is obtained through statistical analysis of historical scan data, for example, by statistically analyzing the average level of energy changes across multiple scans and selecting 20% ​​of this average as an example value to ensure the stability of the judgment criterion. After determining the marginal gain saturation point, the corresponding turn-on delay and the current diffraction efficiency level are combined to form the optimal reference point. Simultaneously, feature vectors for the pre-ionization, main discharge, and relaxation stages are extracted from the corresponding discharge cycle and concatenated into a state vector in a predetermined order. This state vector is associated with the optimal reference point to form a complete sample entry, which is then stored in the sample database in chronological order.

[0065] After constructing the sample database, data is extracted from it for training the time-series prediction network. The samples are sorted according to the chronological order of the discharge cycles to ensure the data reflects the true temporal evolution. A sliding window approach is then used to construct the state sequence, where the state vectors of several consecutive discharge cycles are used as one input sample. The window length is set according to the system's dynamic characteristics; for example, selecting five consecutive cycles as a sequence unit results in each input sample containing five time-ordered state vectors. The corresponding output label is selected from the optimal reference point parameters corresponding to the last cycle of the sequence, namely the Q-switch turn-on delay and acousto-optic diffraction efficiency settings. By gradually moving the sliding window across the entire dataset, a large number of paired input-output samples are constructed, thus forming the training dataset. Subsequently, a temporal prediction network model was constructed. The network structure consists of convolutional layers and gated recurrent units (ROUs). The convolutional layers are used to extract local correlation features between different dimensions within the state vector. The convolutional layers adopt a one-dimensional convolutional form with 32 kernels, and the kernel length covers multiple combinations of dimensions of the state vector. The output of the convolutional layers is non-linearly activated and then input into the gated recurrent units. The gated recurrent units adopt a single-layer structure with 64 hidden units to capture the temporal dependencies between different discharge cycles. The output of the gated recurrent units is mapped to a two-dimensional output space through a fully connected layer, corresponding to the two parameters of activation delay and diffraction efficiency. During training, mean squared error is used as the loss function to measure the difference between the predicted and labeled values. The training batch size is set to 32, and the number of training epochs is set to 100. After each training epoch, the error of the validation set is evaluated. Training stops when the error decreases below a set threshold for several consecutive epochs. Through the above training process, the network can learn the mapping relationship between the state sequence and the optimal control parameters. In actual operation, by inputting the state sequence constructed in real time, the network can output the start-up delay and diffraction efficiency setpoint corresponding to the next discharge cycle, thereby realizing dynamic Q-switching control based on the characteristics of the entire discharge process.

[0066] In step S5, the Q-switch turn-on delay and acousto-optic diffraction efficiency setting value for the next control cycle are output.

[0067] During continuous discharge operation, the constructed full-cycle state vector is acquired cycle by cycle in chronological order, and the state vector is formatted in a unified way to ensure that it meets the input requirements of the pre-trained temporal prediction network. The length of the state sequence is set, i.e., a fixed number of consecutive discharge cycles are selected as an input unit. This number is determined based on the dynamic range of the actual discharge process; for example, five consecutive discharge cycles are selected as example values. This ensures that the input data contains sufficient historical information while avoiding the computational burden of excessively long sequences. During data organization, multiple state vectors acquired in chronological order are arranged sequentially to form an ordered sequence, where the earliest cycle corresponds to the first position of the sequence, and the most recent cycle corresponds to the last position. To ensure that different batches of input data have a consistent scale, a unified range mapping process is performed on each feature value. Specifically, by statistically analyzing the maximum and minimum values ​​of each feature in the historical sample database, the current data is mapped to a fixed interval, such as the range of 0 to 1, thereby eliminating differences between different units. Subsequently, the state sequence is organized according to the input structure required by the pre-trained network. For example, each state vector is used as an input time step, and the entire sequence constitutes a two-dimensional data structure, where the time dimension is the sequence length and the feature dimension is the state vector dimension. In the actual implementation, the continuously acquired state vectors are updated through a fixed data buffer. After each new discharge cycle, the latest state vector is added to the end of the sequence, and the oldest item is removed, forming a sliding update mechanism to ensure that the input sequence always reflects the real state changes of the most recent few cycles.

[0068] After constructing the state sequence, a correspondence between the control period and the discharge period is established using the discharge cycle as a time reference, meaning that each discharge cycle corresponds to one Q-switching control decision process. If the state vector of the current discharge cycle is acquired and updated to the end of the state sequence, an inference operation of the time-series prediction network is performed before the end of that cycle. The constructed state sequence is then input into the trained network model for calculation. The network output is a set of control parameters corresponding to the next discharge cycle, including the Q-switch turn-on delay and the acousto-optic diffraction efficiency setpoint.

[0069] In step S6, the delay start timer is activated by pressing the Q switch of the next control cycle. After the timer expires, the amplitude of the radio frequency drive signal is adjusted according to the acousto-optic diffraction efficiency setting value and applied to the acousto-optic transducer.

[0070] After obtaining the Q-switch turn-on delay and acousto-optic diffraction efficiency setpoints for the next discharge cycle from the timing prediction network output, the turn-on delay parameters are loaded and controlled for timing. A unified definition is established for the end time of the current discharge cycle, using a fixed reference point after the current waveform enters the stable relaxation phase as the cycle end marker. This reference point is determined by detecting that the current remains stable for several consecutive sampling points below the background noise range, thus avoiding time base offset caused by transient fluctuations. Subsequently, using this moment as the timing start point, the predicted turn-on delay parameters are converted into the corresponding counter count value. The count value conversion is based on the system clock frequency. For example, with a clock frequency of 10MHz, each counting cycle is 0.1 microseconds. When the predicted delay is 20 microseconds, the corresponding count value is set to 200 count units. During loading, this count value is written to the preset register of the hardware counting module, and a decrementing count start signal is triggered immediately after writing. The counting module uses a synchronous decrementing method, automatically decreasing by one in each clock cycle until the count value reaches zero. To avoid the accumulation of counting errors during the counting process, a fixed clock source is used to drive the counting unit, ensuring a stable and consistent counting cycle. Simultaneously, the counting register is verified before counting begins to confirm that the loaded value matches the target delay. Furthermore, during continuous operation, when a new predicted delay value is updated, a complete load is always performed at the end of the current cycle to prevent control malfunctions caused by parameter modifications during the counting process.

[0071] When the countdown reaches zero, the RF drive signal amplitude adjustment and output process is triggered. Based on the acousto-optic diffraction efficiency setpoint output by the timing prediction network, the corresponding RF drive amplitude control parameters are determined. This mapping relationship is established experimentally during the equipment calibration phase. Specifically, this is achieved by progressively adjusting the programmable gain amplifier coefficient of the RF drive circuit and measuring the corresponding actual diffraction efficiency, forming a discrete mapping table. For example, a coefficient of 1.2 corresponds to a diffraction efficiency of approximately 30%, 1.8 corresponds to approximately 50%, and 2.5 corresponds to approximately 70%. During operation, the amplifier coefficient value closest to the target diffraction efficiency is directly obtained from the table. This coefficient value is then written to the control register of the programmable gain amplifier to complete the amplitude setting. To ensure smooth signal switching, a synchronous update method is used when writing new coefficients, i.e., coefficient switching is completed near the zero-crossing point of the RF signal to avoid abrupt interference. After amplitude adjustment, the RF drive signal is amplified and output to the acousto-optic transducer, causing the transducer to generate sound waves of corresponding intensity, thereby forming a stable refractive index modulation structure inside the acousto-optic crystal. In actual operation, to ensure output consistency, the amplifier's current coefficients are read and verified before each output to confirm their consistency with the target setting. This mapping relationship remains stable for multiple consecutive cycles unless the network output is updated. Through the above steps, accurate conversion from predicted diffraction efficiency to RF drive signal amplitude is achieved, and the corresponding control signal is stably output at the timing point, thus completing the Q-switching process.

[0072] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0073] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0074] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0075] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0076] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0077] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0078] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0079] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0080] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An acousto-optic Q-switching method for a carbon dioxide laser, characterized by, Includes the following steps: S1. At the rising edge of the pump power supply trigger pulse, capture the pump voltage waveform and current waveform at a sampling rate higher than the discharge oscillation frequency. With the voltage collapse point as the time zero point, align and segment the waveforms to store them as two time waveform pairs: the pre-ionization segment and the main discharge segment. S2. In the pre-ionization stage, obtain the electric field establishment rate characterization value from the voltage waveform and the current ramp-up rate from the current waveform. Combine the electric field establishment rate characterization value and the current ramp-up rate into an ignition feature vector. S3. In the main discharge stage, when the current waveform shows the first peak point, the current waveform is decomposed in the time-frequency domain to extract the characteristic harmonic energy ratio and the time-domain fluctuation rate. At the same time, the voltage waveform is integrated and the corresponding second derivative is calculated as a characterization of the plasma conductivity state change. The characteristic harmonic energy ratio, the time-domain fluctuation rate and the characterization of conductivity state change are combined into a stability feature vector. S4. During the relaxation phase after the main discharge phase ends, the relaxation feature vector is extracted from the slope of the current waveform envelope and the recovery rate of the voltage waveform. S5. Concatenate the ignition feature vector, stability feature vector and relaxation feature vector into a state vector, input it into the pre-trained temporal prediction network, and output the Q-switch turn-on delay and acousto-optic diffraction efficiency setting value for the next control cycle. S6. Taking the end of the current discharge cycle as the starting point of the timing, press the Q switch of the next control cycle to start the delay start timer. After the timing is reached, adjust the amplitude of the radio frequency drive signal according to the acousto-optic diffraction efficiency setting value and apply it to the acousto-optic transducer.

2. The acousto-optic Q-switching method for a carbon dioxide laser according to claim 1, characterized in that, In step S1, storing the waveform in segments as two time-series waveform pairs—a pre-ionization segment and a main discharge segment—specifically includes: When the rising edge of the pump power trigger pulse arrives, the pump voltage waveform and pump current waveform are continuously acquired, and the first derivative of the voltage waveform is monitored in real time. When the first derivative of the voltage shows a negative jump and the jump amplitude exceeds the preset background noise threshold, this moment is taken as the time zero point corresponding to the voltage collapse point. The waveform segment of a set duration is intercepted forward from the zero point of time as the pre-ionization segment waveform pair, and the waveform segment intercepted backward from the zero point of time to the end of the pump discharge is used as the main discharge segment waveform pair. Align the voltage and current waveforms within the pre-ionization phase waveform pair on the time axis, and align the voltage and current waveforms within the main discharge phase waveform pair on the time axis.

3. The acousto-optic Q-switching method for a carbon dioxide laser according to claim 1, characterized in that, In step S2, combining the electric field establishment rate characterization value and the current ramp-up rate into an ignition feature vector specifically includes: The voltage waveform is retrieved from the pre-ionization section waveform pair. The voltage waveform is then calculated on the time axis according to the sampling order. The ratio of the voltage change between adjacent sampling points to the corresponding time interval is obtained, and the maximum voltage change rate is used as the characterization value of the electric field establishment rate. The current waveform is retrieved from the pre-ionization waveform pair. The first zero-crossing point of the current waveform in the pre-ionization section is detected. The rising edge segment of the current waveform from zero value to the first peak value is intercepted along the time axis starting from the zero-crossing point. The ratio of the difference between the current value at the end of the rising edge segment and the current value at the beginning of the rising edge segment to the time span from the beginning to the end is calculated as the current ramp-up rate. The electric field establishment rate and current ramp-up rate are arranged in a preset dimension order to form an ignition feature vector.

4. The acousto-optic Q-switching method for a carbon dioxide laser according to claim 1, characterized in that, In step S3, combining the characteristic harmonic energy ratio, time-domain volatility, and conductivity state change characterization quantities into a stability feature vector specifically includes: Within the main discharge waveform pair, when the current waveform first reaches a peak, a waveform segment of a set duration is extracted from the peak of the current waveform as an analysis window. A sliding windowed short-time Fourier transform is performed on the current waveform within the analysis window. The spectral amplitudes within each time window are accumulated according to frequency to calculate the proportion of energy in the discharge characteristic frequency band to the total energy of the analysis window, thus obtaining the characteristic harmonic energy proportion. The standard deviation of this characteristic harmonic energy proportion between each time window within the analysis window is then calculated to obtain the time-domain volatility. The voltage waveform is retrieved from the main discharge segment waveform. Starting from the voltage collapse point, the voltage value is accumulated point by point and integrated to obtain the injected energy accumulation sequence. Then, a second-order difference operation is performed. The maximum absolute value of the second-order difference sequence is used as the characteristic quantity of the plasma conductivity state change. The characteristic harmonic energy ratio, time-domain volatility, and conductivity state change parameters are arranged in a predetermined dimension order to form a stability feature vector.

5. The acousto-optic Q-switching method for a carbon dioxide laser according to claim 1, characterized in that, In step S4, extracting the relaxation feature vector from the slope of the current waveform envelope and the recovery rate of the voltage waveform specifically includes: After the main discharge phase ends, the moment when the detected current waveform drops from its peak point to the first time below the preset background noise threshold is taken as the discharge extinction point, and a waveform segment of a set duration is extracted as the relaxation analysis window. Extract the peak amplitudes from the current waveform within the relaxation analysis window, take the natural logarithm of each peak amplitude, and then perform linear fitting in time order. Take the absolute value of the slope obtained from the fitting as the composite decay rate. Extract the rising segment of the voltage waveform from the relaxation analysis window, perform linear fitting on the voltage amplitude in the rising segment in time sequence, and use the slope obtained by fitting as the electric field recovery rate. The composite decay rate and the electric field recovery rate are arranged in a predetermined order according to a preset dimension to form a relaxation feature vector.

6. The acousto-optic Q-switching method for a carbon dioxide laser according to claim 1, characterized in that, In step S5, the training method for the pre-trained temporal prediction network is as follows: Within multiple consecutive test discharge cycles, diffraction efficiency constraints are set and the Q-switch turn-on delay parameters are traversed in a stepwise manner. The actual output pulse energy value is collected and optimized. The marginal gain saturation point corresponding to each discharge cycle is determined as the optimal reference point. The optimal reference point of each discharge cycle is associated with the corresponding state vector and stored in the sample database. The discharge cycle includes a pre-ionization phase, a main discharge phase, and a relaxation phase. The state vector is a combination vector of the fusion ignition feature vector, the stability feature vector, and the relaxation feature vector. State vectors from several consecutive test discharge cycles are extracted from the sample database in chronological order to form a state sequence. The state sequence is used as input, and the Q-switch turn-on delay and acousto-optic diffraction efficiency setpoints corresponding to the optimal reference point are used as output labels to train a time-series prediction network constructed from convolutional layers and gated recurrent units.

7. The acousto-optic Q-switching method for a carbon dioxide laser according to claim 1, characterized in that, In step S5, the Q-switch turn-on delay and acousto-optic diffraction efficiency setting values ​​for the next control cycle specifically include: The state vectors of continuous discharge cycles are integrated into a state sequence that conforms to the input standard of the pre-trained time-series prediction network. The corresponding control period is set based on the discharge cycle. The output of the time-series prediction network is used as the Q-switch turn-on delay and acousto-optic diffraction efficiency setting value for the next control period.

8. The acousto-optic Q-switching method for a carbon dioxide laser according to claim 1, characterized in that, In step S6, pressing the Q switch of the next control cycle to start the delay start timer, and after the timer expires, adjusting the amplitude of the radio frequency drive signal according to the acousto-optic diffraction efficiency setting value and applying it to the acousto-optic transducer specifically includes: The system obtains the Q-switch turn-on delay and acousto-optic diffraction efficiency setting value for the next control cycle output by the timing prediction network. It loads the turn-on delay count value and starts the count-down process with the end time of the current discharge cycle as the timing start point. When the count-down reaches zero, it adjusts the programmable gain amplifier coefficient of the RF drive circuit according to the acousto-optic diffraction efficiency setting value and outputs the adjusted RF drive signal to the acousto-optic transducer.

Images