Process quality detection method applied to aluminum-plated plate stamping die coating

Abstract

The invention relates to the technical field of coatings, in particular to an aluminum-plated plate stamping die coating process quality detection method, which comprises the following steps: arranging an ultrasonic transducer on a measuring side surface in a coupling manner, and collecting echo signals containing adjacent back wall echoes by adopting pulse echoes, multiple collection results at the same detection position are aggregated to obtain a target echo signal; performing band-pass band limiting and noise suppression on the target echo signal to obtain a filtering echo signal, and determining an echo event time window according to the envelope of the filtering echo signal; intercepting a first back wall echo and a second back wall echo which arrive in sequence in the time window, and adding a boundary smooth window; respectively performing frequency domain transformation to obtain a first echo frequency spectrum and a second echo frequency spectrum; performing regularization deconvolution based on adjacent back wall echoes, performing conjugate multiplication on the second echo spectrum and the first echo spectrum to obtain a cross-correlation spectrum, adding the amplitude square of the first echo spectrum and a regularization term to obtain a stabilizing denominator, and dividing the stabilizing denominator by the stabilizing denominator to obtain a medium transfer function spectrum; performing inverse transformation and normalization to obtain a reflectivity function, and inputting the reflectivity function into a convolutional neural network to output a smooth pulse sequence; positioning two interface reflection peaks in the smooth pulse sequence and performing interpolation fitting to obtain sub-sampling positioning, calculating a peak interval as coating back-and-forth propagation time and converting thickness, and comparing with a process quality criterion to output a conclusion. The method can be used for stably acquiring key quality information such as the thickness of the coating on the premise that the coating is not damaged.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of coating technology, in particular to a coating process quality detection method applied to an aluminum-plated plate stamping die. BACKGROUND

[0002] In the field of modern metal forming, stamping dies are widely used in the processing of various metal materials, especially in the stamping, drawing and forming operations of metal materials. The surface performance of the die plays a crucial role. Especially in the stamping process of metal materials such as galvanized plates, the die is often exposed to harsh environments of high temperature, high pressure and friction, which requires the die material to have excellent wear resistance, corrosion resistance and oxidation resistance.

[0003] In the surface protection system of key tooling such as stamping dies, the coating state and thickness are important indicators for evaluating the protection ability and service reliability. In the prior art, precise detection of the coating of the die is usually carried out by destructive detection, that is, the coating is destroyed to detect the coating quality. However, when the geometry is complex or the number of dies is large, destructive detection is not convenient to implement.

[0004] Therefore, a detection method sensitive to the coating state and suitable for non-destructive implementation on site is needed. The existing ultrasonic pulse echo detection is widely used because of its non-invasive and convenient implementation. However, in a multi-layer structure or when a coating exists, the echoes formed by multiple acoustic interfaces will be close in time and overlap, making it difficult to identify and separate the contributions of each surface of the die, and thus affecting the reliable estimation of the coating thickness and other parameters. SUMMARY

[0005] In view of the above technical problems, the present application provides a coating process quality detection method applied to an aluminum-plated plate stamping die, which is used to stably obtain key quality information such as coating thickness without damaging the coating. Especially for the problems caused by the close and overlapping ultrasonic echoes due to the existence of the coating, such as the difficulty in separating the interface echoes and the unreliability of the thickness estimation, the present application strives to maintain high detection accuracy and repeatability under the working conditions of fluctuating coupling states on site.

[0006] Other characteristics and advantages of the present application will become apparent from the following detailed description, or will be learned by practice of the present application.

[0007] According to one aspect of the present application, a coating process quality detection method applied to an aluminum-plated plate stamping die is provided, the stamping die comprising a die base and a coating formed on a working side surface of the die base, the working side surface being opposite to a measurement side surface of the die base; the method comprising the following steps S1-S7: S1, coupling an ultrasonic transducer to the measurement side surface, acquiring an ultrasonic echo signal containing adjacent back wall echoes in a pulse-echo mode, and aggregating multiple acquisition results at the same detection position to obtain a target echo signal; S2, performing band limitation and noise suppression processing on the target echo signal to obtain a filtered echo signal, and determining an echo event time window according to an envelope of the filtered echo signal; S3, intercepting a first back wall echo and a second back wall echo that arrive in sequence within the echo event time window to obtain the first back wall echo signal and the second back wall echo signal respectively, and applying a boundary smoothing window to the first back wall echo signal and the second back wall echo signal; S4, performing frequency domain transformation on the first back wall echo signal and the second back wall echo signal respectively to obtain a first echo spectrum and a second echo spectrum; S5, constructing deconvolution based on adjacent back wall echoes: multiplying the second echo spectrum and the conjugate spectrum of the first echo spectrum to obtain a cross-correlation spectrum, adding the amplitude square of the first echo spectrum to a regularization term determined by the power spectral density of the first echo spectrum to obtain a stabilization denominator, and dividing the cross-correlation spectrum by the stabilization denominator to obtain a medium transfer function spectrum; S6, performing inverse frequency domain transformation and normalization on the medium transfer function spectrum to obtain a reflectivity function, inputting the reflectivity function into a convolutional neural network trained by synthesized data, and outputting a smoothed pulse sequence with the same length as the reflectivity function; S7, determining the peak positions of the reflection pulse from the mold substrate to the coating and the reflection pulse from the coating to the environment medium in the smoothed pulse sequence, performing interpolation fitting on the peak positions to obtain sub-sampling time positioning accuracy, calculating the time interval of the two peak positions as the coating round-trip propagation time, converting the coating round-trip propagation time into coating thickness according to a preset sound velocity parameter or a calibration conversion relationship, and comparing the coating thickness with a preset process quality criterion to output a detection conclusion.

[0008] Further, in step S1, time domain alignment and abnormal echo rejection are performed on the ultrasonic echo signal obtained at the same detection position, and average aggregation is performed between multiple echoes after time domain alignment to reduce random noise and coupling fluctuations, the average aggregation including one of arithmetic average and weighted average; wherein the coupling state of each ultrasonic transducer and the measurement side surface is defined as an independent measurement channel, and the deconvolution of step S5 and the convolutional neural network inference of step S6 are respectively completed in each independent measurement channel to obtain the reflectivity function without relying on reference samples or external reference signals.

[0009] Further, the band limitation in step S2 is implemented by a finite impulse response band-pass filter, a passband of which is determined according to an effective working frequency band of the ultrasonic transducer and an acoustic response of the mold base and the coating, and filter coefficients are designed by a window function method to give consideration to passband flatness and stopband attenuation; the envelope in step S2 is obtained by rectifying the filtered echo signal and then low-pass filtering, and a peak value and a threshold value intersection point of the envelope are used to determine start and end positions of the echo event time window.

[0010] Further, the echo event time window is intercepted by: identifying a back wall echo family arranged in a propagation time sequence in the filtered echo signal, and selecting two back wall echoes adjacent to each other and satisfying a preset signal-to-noise quality condition as the first back wall echo and the second back wall echo; the boundary smoothing window is a windowing function with a smoothly changing edge, a change interval of which is controlled by a preset change coefficient to suppress boundary discontinuity introduced by echo truncation without significantly weakening echo main body energy.

[0011] Further, steps S3 to S6 are repeatedly performed on different adjacent back wall echo pairs, wherein the adjacent back wall echo pairs include: a back wall echo pair taking a first-arriving back wall echo as the first back wall echo and a back wall echo arriving immediately after the first-arriving back wall echo as the second back wall echo, and a back wall echo pair taking a subsequent pair of back wall echoes as the first back wall echo and the second back wall echo; a plurality of the smoothed pulse sequences obtained for each echo pair are fused, and one of the following is performed during fusion: average after amplitude normalization, confidence weighted summation, and voting selection according to peak consistency.

[0012] Further, the regularization term is determined by the following process: a power spectral density of the first echo spectrum is calculated, and a statistical feature of the power spectral density is extracted as a noise floor estimate, the statistical feature including at least one of a peak value, a quantile value, a mean value, or a mean square value; the noise floor estimate is multiplied by a preset proportion coefficient to obtain the regularization term, and the regularization term is added to a square of an amplitude of the first echo spectrum before the cross-correlation spectrum is divided by the stabilization denominator.

[0013] Further, the training of the convolutional neural network includes: a layered propagation medium model for characterizing the aluminum-plated plate stamping die is constructed, the layered propagation medium model including a coupling medium layer in contact with the ultrasonic transducer, the mold base, a coating on a working surface of the mold base, and an environmental medium adjacent to the coating; The numerical acoustic field simulation based on wave equation generates the training echo, which adopts a discrete solution of finite difference time domain, and sets an absorbing boundary on the boundary of the simulation domain to suppress non-physical reflection; In each simulation, the sound velocity, density and attenuation parameters of the coating and the mold base are sampled from a preset material attribute set, and a reference excitation waveform obtained by actually measuring the base without coating is applied to the ultrasonic transducer as an incident pulse; The simulation received echo signal is obtained by recording the time sequence of the pressure field at the sensor position corresponding to the ultrasonic transducer in the numerical acoustic field simulation to form a spatially sampled echo set, and spatially averaging the echo set within the sensor aperture to obtain a one-dimensional time sequence echo signal in the same format as the measured single-channel echo signal; The simulation received echo signal is processed in accordance with steps S2 to S6 to obtain the reflectivity function, a supervision label is constructed for the reflectivity function, the supervision label is a sparse pulse sequence, wherein the first non-zero pulse corresponds to the reflection event of the mold base to the coating, and the second non-zero pulse corresponds to the reflection event of the coating to the environment medium, and the kernel function smoothing is performed on the sparse pulse sequence to form a smoothed supervision label sequence, so as to complete the supervised training of the convolutional neural network.

[0014] Further, the convolutional neural network is a one-dimensional full convolution structure, the network main body of which is formed by cascading multiple convolution layers, each convolution layer adopts a linear rectifier activation function, and the convolution kernel is used in the front convolution layer to extract the global morphological features of the echo, and the convolution kernel is used in the rear convolution layer to extract the local overlapping mode features; The convolutional neural network introduces dilated convolution in part of the convolution layers to expand the receptive field while maintaining the time resolution, and sets a dropout layer to reduce overfitting; The output end of the convolutional neural network is provided with a point-by-point convolution layer, which transforms the feature map into the smoothed pulse sequence with the same length as the reflectivity function; In the training process, the convolutional neural network adopts an adaptive optimization algorithm to minimize the mean square error loss, takes the error between the smoothed pulse sequence and the smoothed supervision label sequence as the optimization objective, and combines the learning rate adaptive adjustment and early stopping strategy.

[0015] Further, in determining the peak position, it comprises: Local maxima satisfying a preset significance condition are searched in the smoothed pulse sequence, two main maxima corresponding to the reflection pulse of the mold base to the coating and the reflection pulse of the coating to the environment medium are taken as target peaks, and parabolic fitting is performed on each target peak in its neighborhood to obtain sub-sampling time positioning; The process quality criterion includes a target interval of the coating thickness, an uniformity index of the coating thickness between different detection positions, and a number of abnormal peaks in the smoothed pulse sequence except for the two main maxima, and an unqualified conclusion is output when the coating thickness deviates from the target interval or the uniformity index does not meet the requirements, or the abnormal peaks represent an interface reflection anomaly; and the detection results of different batches or different times are compared in trend to represent the stability of the coating process.

[0016] The technical solution of the present disclosure has the following beneficial effects: Compared with the prior art, the present application can more reliably extract information related to coating quality in a complex echo scene, reduce the uncertainty caused by interface echo overlap, noise interference and contact coupling changes, and improve the stability and consistency of coating thickness evaluation results. This method does not require sampling or destructive peeling of the mold coating, is suitable for batch detection of molds with complex geometry and large number, and can significantly improve process efficiency.

[0017] The present application combines signal processing mechanism with data-driven model, making the recognition of adjacent interface reflection events more robust, and better adapting to the differences in echo patterns under different mold base states, different coating formation conditions and different on-site detection conditions, thereby expanding the application range of the method. Thus, not only quantitative results such as coating thickness can be output, but also more reliable data support can be provided for process stability evaluation, abnormal batch early warning and quality traceability, promoting process control and continuous improvement of coating process quality. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 A flowchart of a coating process quality detection method for an aluminum-plated plate stamping die according to an embodiment of the present disclosure; Figure 2 A detection scene diagram of a stamping die and an ultrasonic transducer according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0019] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations can be implemented in any

[0020] The application provides a product applied to an aluminized plate stamping die coating process quality detection method. Referring to FIG. 1, a flowchart diagram of the aluminized plate stamping die coating process quality detection method provided by an embodiment of the application is shown. As shown in FIG. 1, the aluminized plate stamping die coating process quality detection method comprises the following steps: Figure 1 Figure 2 As shown in FIG. 2, the stamping die 3 comprises a die base body and a coating layer formed on a working side surface of the die base body. The working side surface is arranged opposite to a measuring side surface of the die base body. An ultrasonic transducer 2 is coupled and arranged on the measuring side surface. The ultrasonic transducer generates an incident wave 1. The incident wave generates an ultrasonic echo signal 4 in the die 3. The method can specifically comprise steps S1-S7: In step S1, the ultrasonic transducer is coupled and arranged on the measuring side surface. The ultrasonic echo signal containing the echo of the adjacent back wall is collected by using the pulse echo method. The multiple collection results of the same detection position are aggregated to obtain the target echo signal.

[0021] ​In the measurement side surface, the measurement side surface is cleaned to reduce the influence of surface contamination on coupling, and a coupling medium is applied between the measurement side surface and the ultrasonic transducer to form a stable acoustic channel, so that the longitudinal wave emitted by the transducer can enter the mold base in a nearly normal incidence manner and propagate inside. The measured ultrasonic response is easily affected by the sensor position and the contact quality, and the thickness of the coupling medium, the local surface roughness and the local differences caused by the coating process can all change the echo amplitude and phase. Therefore, each contact between the sensor and the measurement side surface is regarded as an independent acquisition channel, and the corresponding original echo data of each channel needs to be retained separately in the subsequent process to avoid errors caused by different contact states being used as references. The acquisition uses the pulse echo method: the test system applies an excitation pulse to the transducer, the transducer emits ultrasonic waves into the mold and receives the echoes returned from different interfaces. The recorded echo signal should contain at least two adjacent back wall main echoes in time to ensure that the continuous echoes of the same propagation path can be completely intercepted and used for subsequent processing. To ensure time resolution, the acquisition system can use high-speed analog-to-digital conversion to sample the echoes, for example, the sampling frequency is 100 MHz, and the corresponding discrete sampling interval is about 8 ns, so that the arrival interval of adjacent back wall echoes can be covered in the same acquisition window and the detailed features can be retained. To enhance the signal-to-noise ratio and reduce the influence of incidental disturbances, multiple echoes are usually continuously acquired at the same detection position before entering the subsequent process. The number of continuous acquisitions can be set according to the device capacity and stability. In practice, a group of continuous acquisitions are usually used to form multiple echo records, so that random errors and fluctuations can be suppressed by statistical methods. , the corresponding discrete sampling interval is about 8 ns, so that the arrival interval of adjacent back wall echoes can be covered in the same acquisition window and the detailed features can be retained. To enhance the signal-to-noise ratio and reduce the influence of incidental disturbances, multiple echoes are usually continuously acquired at the same detection position before entering the subsequent process. The number of continuous acquisitions can be set according to the device capacity and stability. In practice, a group of continuous acquisitions are usually used to form multiple echo records, so that random errors and fluctuations can be suppressed by statistical methods.

[0022] , the corresponding discrete sampling interval is about 8 ns, so that the arrival interval of adjacent back wall echoes can be covered in the same acquisition window and the detailed features can be retained. To enhance the signal-to-noise ratio and reduce the influence of incidental disturbances, multiple echoes are usually continuously acquired at the same detection position before entering the subsequent process. The number of continuous acquisitions can be set according to the device capacity and stability. In practice, a group of continuous acquisitions are usually used to form multiple echo records, so that random errors and fluctuations can be suppressed by statistical methods.

[0023] The multiple echo records obtained at the same detection position are time-domain aligned before aggregation, so that the key features of each echo are coincided on the time axis, avoiding the overall time delay difference caused by touch pressure changes, coupling layer changes or touch point microshift from being introduced into the spread after being averaged. The time-domain alignment can determine the alignment reference point based on the peak position of the back wall main echo, the envelope peak position or the cross-correlation maximum value position, and perform corresponding time shift compensation for each echo; when there is a sampling point level deviation, interpolation can be used to realize sub-sampling level alignment compensation to reduce residual time delay. After alignment, abnormal echo elimination is performed to remove acquisition records that deviate significantly from the group characteristics, such as coupling instantaneous failure, echo amplitude mutation, waveform distortion or echo loss caused by slip and the like; the abnormal criterion can be comprehensively judged by statistical quantities such as the peak amplitude, energy of the back wall main echo, the correlation coefficient of the average waveform, or the deviation of the echo arrival time. The remaining valid echoes are averaged and aggregated to obtain the target echo signal, and the arithmetic mean can be used to improve the steady-state performance under random noise, and the weighted mean can be used to reduce the influence of weak coupling or low confidence echoes on the result, and the weighting factor can be determined by the signal-to-noise ratio of a single echo, the back wall main echo amplitude stability or the correlation coefficient and the like. Through the above alignment, elimination and aggregation, the target echo signal can significantly reduce the uncertainty brought by random noise and coupling fluctuation while maintaining the structure information of adjacent back wall echoes, and provide stable input for subsequent deconvolution and neural network inference in each independent acquisition channel, so that the extraction of reflectivity related information does not depend on external reference samples or cross-sample reference signals.

[0024] In step S2, the target echo signal is subjected to band limitation and noise suppression processing to obtain a filtered echo signal, and the envelope of the filtered echo signal is used to determine an echo event time window.

[0025] In step S2, the band limitation is implemented by using a finite impulse response band-pass filter, the passband of the finite impulse response band-pass filter is determined according to the effective working frequency band of the ultrasonic transducer and the acoustic response of the mold base and the coating, and the filter coefficients are designed by using a window function method to consider the passband flatness and the stopband attenuation; the envelope in step S2 is obtained by rectifying the filtered echo signal and then low-pass filtering, and the start and end positions of the echo event time window are determined by the intersection of the peak value and the threshold value of the envelope.

[0026] When the target echo signal is band-limited, the band-pass range is first determined in combination with the effective working frequency band of the transducer and the frequency spectrum response of the mold base and coating system, so that the information related to the interface reflection in the echo is in the passband, and the noise components more likely to appear in the higher frequency band are suppressed as much as possible. The band limitation can be realized by using a finite impulse response band-pass filter. The filter coefficients are designed by using the window function method, which ensures the flatness of the passband and improves the stopband attenuation to reduce the leakage of out-of-band noise. The filter coefficients can be generated by using the common FIR design process based on the window function, and the coefficients are quantized according to the digital implementation capability of the actual acquisition system, so as to ensure the stable reproduction of the filtering effect in the self-developed or embedded measurement system. The filtered echo signal is obtained after band-pass filtering. The filtering process not only plays a frequency band limiting role, but also suppresses the out-of-band broadband noise, discrete peak interference and the like, so that the subsequent time window determination is more reliable.

[0027] To determine the echo event time window, the envelope can be extracted from the filtered echo signal and the gating can be implemented accordingly. The envelope can be obtained by first rectifying and then low-pass filtering: the filtered echo signal is full-wave rectified to change the bipolar oscillation into unipolar amplitude variation, and then the rectified result is smoothed by a FIR low-pass filter to obtain an envelope curve reflecting the energy change over time. When the envelope curve is used to window the echo, the peak value region corresponding to the back wall main echo can be located on the envelope, and then the intersection points of the envelope and the threshold value are searched forward and backward with the peak value as the center. The start and end time corresponding to the intersection points is determined as the boundary of the echo event time window. The threshold value can be determined adaptively according to the relative proportion of the envelope peak value or the noise floor, so as to balance the echo integrity and noise rejection effect. Through this envelope-based gating method, the effective time segment containing adjacent back wall echoes can be stably locked in the case of echo congestion or local signal-to-noise variation, and irrelevant pre-ringing, tail wave and random noise can be avoided from being brought into the subsequent processing. When the envelope peak value is not obvious, the threshold intersection points are not paired, or the time window width is abnormal, which represents poor echo quality, the signal can be marked as abnormal and not entered into the subsequent step, so as to improve the robustness of the whole detection process.

[0028] In step S3, the first back wall echo and the second back wall echo that arrive in sequence are intercepted in the echo event time window, and the first back wall echo signal and the second back wall echo signal are obtained, and the boundary smoothing window is applied to the first back wall echo signal and the second back wall echo signal.

[0029] The selection of the echo event time window includes: identifying a family of back wall echoes arranged in the propagation time order in the filtered echo signal, and selecting two adjacent back wall echoes whose amplitudes meet the preset signal-to-noise quality conditions as the first back wall echo and the second back wall echo; the boundary smoothing window is a windowing function with a smooth gradient edge, and its gradient range is controlled by a preset gradient coefficient, so as to suppress the boundary discontinuity introduced by the echo truncation without significantly weakening the energy of the echo body.

[0030] Based on the filtered echo signals and echo event time windows obtained in the previous stage, the backwall echo family is identified within this time window. The backwall echo family refers to a group of main echoes generated by multiple reflections from the backwall interface along the same propagation path, arranged chronologically. Adjacent main echoes have a fixed propagation relationship, and when the coating thickness is much smaller than the substrate thickness, the two main backwall echoes can be clearly separated on the time axis, facilitating stable interception within the same time window. During identification, a significant peak or a peak position based on the envelope can be searched within the time window. Two adjacent main backwall echoes are selected as the first and second backwall echoes according to their arrival order. Weak or distorted echoes are then filtered out using preset signal-to-noise quality conditions. These conditions are comprehensively judged by indicators such as the margin of the echo peak amplitude relative to the noise floor, the echo energy concentration, or the correlation with the mean waveform of echoes at the same location, to ensure the stability of subsequent frequency domain analysis and deconvolution. After determining the main echo pair, the first backwall echo and the second backwall echo are truncated within the time window to form the first backwall echo signal and the second backwall echo signal. This ensures that both signals cover the main lobe of their respective echoes and the necessary ringing tail without excessively introducing adjacent irrelevant components, thus balancing the integrity of the echo information and the purity of the window.

[0031] The first and second back-wall echo signals can be considered as time segments of two consecutive back-wall echoes in the received signal. Their formation mechanism can be described by a convolution model: the first back-wall echo is the convolution of the measurement system response and the impulse response of the first echo propagation path, plus noise; the second back-wall echo is the result of further propagation and attenuation based on the previous propagation, plus noise, respectively satisfying the following: ; ; in, This represents the equivalent impulse response of the measurement system. This represents the total propagation path impulse response corresponding to the first backwall echo. This represents the impulse response of the second backwall echo relative to the first backwall echo, representing the additional propagation path. , This represents additive noise. Converting to the frequency domain, it can be written as: ; ; In the case of high signal-to-noise ratio, the noise term is relatively small, but the window truncation will form discontinuity at the segment boundary, causing spectral leakage in the frequency domain and affecting the accuracy of the frequency domain representation, so before the first back wall echo signal and the second back wall echo signal are subjected to frequency domain transformation, the two echo segments need to be subjected to boundary smoothing window to smooth the segment edge. The boundary smoothing window can use a windowing function with a smooth transition edge, such as a Tukey window, by setting a small transition ratio, the window smoothly transitions to zero at both ends of the segment to reduce the spectral leakage caused by the boundary mutation, while trying to keep the echo main energy and detail information from being weakened; the transition interval is controlled by a preset transition coefficient, the smaller the transition coefficient, the more sufficient the main body is retained while the edge suppression is still effective. By windowing the two echoes respectively, the subsequent frequency domain analysis can be closer to the real spectral characteristics, and the pseudo-frequency components introduced by truncation can be reduced.

[0032] In step S4, the first back wall echo signal and the second back wall echo signal are subjected to frequency domain transformation respectively to obtain the first echo spectrum and the second echo spectrum.

[0033] After obtaining the first back wall echo signal and the second back wall echo signal, frequency domain transformation is performed on the two echoes respectively to obtain the corresponding echo spectrum. The frequency domain transformation can be realized by using fast Fourier transform, and the first back wall echo signal is mapped to the first echo spectrum , and the second back wall echo signal is mapped to the second echo spectrum . Since the phase information of the spectrum and the complex conjugate operation are needed for subsequent processing, the output of the frequency domain transformation should retain the complex form, including the real part and the imaginary part or the equivalent amplitude and phase information, rather than just extracting the amplitude spectrum. To ensure the comparability of the two echoes in the frequency domain, the two frequency domain transformations should use consistent transformation length and frequency resolution settings, and maintain consistent frequency axis mapping relationship with the sampling frequency during acquisition, so that the physical frequency corresponding to each frequency point is consistent; when it is necessary to improve the smoothness of the frequency domain interpolation or facilitate calculation, the time domain echo can be zero-padded when the length is insufficient, and then subjected to fast Fourier transform, the zero-padding only changes the frequency domain sampling density without introducing new physical information. In order to avoid the mutation of the time domain segment at the boundary causing non-real spectral leakage in the frequency domain, the frequency domain transformation is usually based on the echo segment that has been subjected to boundary smoothing processing in the previous step, so that the obtained first echo spectrum and second echo spectrum are closer to the spectral characteristics formed by the real propagation and reflection in the main frequency band, and provide stable input for the subsequent frequency domain operation based on and .

[0034] In step S5, deconvolution is constructed based on adjacent backwall echoes: the cross-correlation spectrum is obtained by multiplying the second echo spectrum with the conjugate spectrum of the first echo spectrum, the square of the amplitude of the first echo spectrum is added to the regularization term determined by the power spectral density of the first echo spectrum to obtain the stabilization denominator, and the medium transfer function spectrum is obtained by dividing the cross-correlation spectrum by the stabilization denominator.

[0035] The regularization term is determined through the following process: calculating the power spectral density of the first echo spectrum, extracting the statistical features of the power spectral density as a noise floor estimate, the statistical features including at least one of peak value, quantile value, mean value or mean square value; multiplying the noise floor estimate by a preset scaling factor to obtain the regularization term, and adding the regularization term to the square of the amplitude of the first echo spectrum before dividing the cross-correlation spectrum by the stabilization denominator.

[0036] When constructing the deconvolution, the propagation recursion relationship between two adjacent backwall main echoes is utilized. The second echo is considered as the result of the first echo propagating further and attenuating under the same measurement system, thus eliminating the unknown measurement system response term between the two echoes. Based on the linear time-invariant assumption of pulse echoes, the echo in the time domain can be represented by a convolution model as follows: ; in This represents the equivalent impulse response of the measurement system to electrical excitation. This represents the impulse response resulting from medium propagation and reflection. This represents additive noise. After representing two adjacent backwall echoes in the frequency domain, under conditions of high signal-to-noise ratio, the direct relationship between the two echoes in the frequency domain can be approximated: ; in The spectrum of the first echo. For the second echo spectrum, This is the frequency domain transfer function corresponding to the additional propagation path relative to the first echo, used to characterize the additional propagation effects and interface reflection information experienced by the second echo relative to the first echo. This allows for the estimation of... The problem is transformed into a deconvolution problem consisting of only two adjacent echoes, without the need to introduce external reference echoes or separately calibrate the system impulse response.

[0037] To improve stability at frequencies with smaller spectral amplitudes, a Wiener-type deconvolution with regularization is used to estimate the transfer function spectrum. Specifically, the second echo spectrum is first multiplied by the conjugate spectrum of the first echo spectrum to form a cross-correlation spectral term. ,in Represent the complex conjugate; then square the amplitude of the first echo spectrum. Adding this to the regularization term K yields the stabilized denominator, thus providing an estimate of the spectrum of the medium transfer function: ; In this formula, the numerator corresponds to the cross-correlation spectrum of the two echoes, and the denominator corresponds to the self-power spectrum of the first echo plus a regularization term, which can... When the value is close to zero or small, the unstable estimation caused by division amplification noise is suppressed, and a balance is achieved between noise suppression and signal fidelity. Through the above construction method, the output is obtained... This is the spectrum of the medium transfer function, which lays the foundation for the subsequent return from the frequency domain to the time domain and the extraction of reflection-related information.

[0038] The regularization term K can be determined by characterizing the noise floor using the power spectral density of the first echo spectrum, allowing the regularization intensity to adaptively vary with the ambient noise level. One feasible calculation method is to set the noise term scaling factor based on the maximum value of the square of the first echo spectrum amplitude, as given by: ; in This represents the maximum value of the squared amplitude of the first echo spectrum on the frequency axis. This form is equivalent to extracting statistical features from the power spectral density of the first echo as a noise floor estimate and multiplying it by a scaling factor to obtain a regularization term, making the regularization match the actual signal energy scale. Since adjacent backwall echoes are obtained in a short time, the measurement system remains approximately consistent between the two echoes. Using the first echo to estimate the noise floor can reflect the common noise environment of the two echoes, thereby improving the consistency and repeatability of the deconvolution results.

[0039] In step S6, the spectrum of the medium transfer function is subjected to inverse frequency domain transformation and normalized to obtain the reflectivity function. The reflectivity function is then input into a convolutional neural network trained with synthetic data, and the output is a smooth pulse sequence of the same length as the reflectivity function.

[0040] After obtaining the spectrum of the medium transfer function, it is transformed from the frequency domain back to the time domain to obtain the time-domain expression of the reflectivity function. This is specifically achieved through the inverse fast Fourier transform, and can be written as: ; Obtained by inverse transformation The impulse response characterizes the newly added propagation path between adjacent backwall echoes. In the case of a single matrix material, this impulse response corresponds to the reflectivity response of the medium. However, in the presence of a multilayer structure, reflection events at different interfaces in the reflectivity response appear in pulse form. To eliminate amplitude scale differences caused by different measurement points, different gain or attenuation conditions, and to ensure that different samples are within a consistent numerical range, the time-domain reflectivity response is normalized to obtain the reflectivity function. In the normalized reflectance function, the interface corresponding pulse usually presents a typical ultrasonic pulse echo shape instead of an ideal Dirac impulse, because the interface reflection is convolved with the impulse response of the band-limited transducer, and the non-ideal excitation, propagation attenuation, ringing caused by dispersion, multiple reflections inside the coating, and noise further change the pulse shape and side lobe structure. After normalization, the reflectance function is sent to the convolutional neural network as a single-channel one-dimensional input. The convolutional neural network learns the mapping relationship between the reflectance response and the reflection event through pure convolution operation in time sequence. The network output is a smooth pulse sequence as long as the reflectance function, which is used to represent the position and relative amplitude of each reflection event on the time axis. The smooth pulse sequence corresponds to the supervised target based on the sparse pulse sequence and the kernel smoothing. The typical theoretical sparse pulse sequence can be expressed as: ; wherein represents the Dirac function, represents the reference propagation time parameter related to the substrate layer, represents the propagation time parameter inside the coating layer, and represent the interlayer reflection coefficient and transmission coefficient, respectively, represents the amplitude factor related to the propagation attenuation; the network output adopts a smooth form to reduce the difficulty caused by the sparse label to learning, while keeping the time position of the reflection event unchanged.

[0041] In step S7, the peak positions of the reflection pulse from the mold substrate to the coating and the reflection pulse from the coating to the environment medium are determined in the smooth pulse sequence, and the peak positions are interpolated and fitted to obtain sub-sampling time positioning accuracy. The time interval between the two peak positions is calculated as the round-trip propagation time of the coating. The round-trip propagation time of the coating is converted into the coating thickness according to the preset sound velocity parameter or the calibration conversion relationship. The coating thickness is compared with the preset process quality criterion to output the detection conclusion.

[0042] wherein, in determining the peak position, including: searching for a local maximum value satisfying a preset significance condition in the smooth pulse sequence, taking two main maxima corresponding to the reflection pulse from the mold substrate to the coating and the reflection pulse from the coating to the environment medium as target peaks, and performing parabolic fitting on each target peak in its neighborhood to obtain sub-sampling time positioning; the process quality criterion includes a target interval of the coating thickness, a uniformity index of the coating thickness between different detection positions, and the number of abnormal peaks in the smooth pulse sequence except the two main maxima. When the coating thickness deviates from the target interval or the uniformity index does not meet the requirements, or the interface reflection is abnormal as represented by the abnormal peak, an unqualified conclusion is output. The detection results of different batches or different times are compared in trend to represent the stability of the coating process.

[0043] When extracting coating-related information from a smooth pulse sequence, the first step is to search for local maxima that satisfy the significance criteria across the entire sequence. These criteria can be constrained by factors such as peak prominence, relative background amplitude ratio, and minimum inter-peak spacing to avoid mistaking minor ringing or noise fluctuations as valid reflection events. Additional small spikes may appear in the prediction results; these spikes typically originate from secondary reflections or localized weak reflection events. However, the two principal maxima are identifiable under all operating conditions. Therefore, the two principal maxima with the largest amplitudes are selected as target peaks, corresponding to the reflection events at the mold substrate-to-coating interface and the coating-to-environmental medium interface, respectively. To overcome the resolution limitations caused by the sampling point interval, parabolic interpolation fitting is performed on each target peak within its neighborhood: the sampling points of the target peak and its adjacent sampling points constitute the fitting data, and the vertex position is obtained on the local quadratic curve. The x-coordinate of the vertex corresponds to the fine position where the peak occurs, thus obtaining the sub-sampling level peak position estimate; the difference between the fine positions of two target peaks is the interval between the two reflection events, which is used to calculate the round-trip propagation time of the coating, and this interval can be expressed as the number of sampling points or converted into a time quantity.

[0044] The interval between two target peaks is defined as the coating's round-trip time of flight (ToF3), which can be directly obtained from the peak-to-peak distance under sampling point measurement; under time measurement, it can be obtained by dividing the sampling point measurement by the sampling frequency. The flight time and coating thickness satisfy a conversion relationship, and the expression for sampling point measurement is: ; in Sampling frequency, This refers to the one-way propagation time of the coating. For coating thickness, The sound velocity is the velocity of the coating. Therefore, the interpeak distance can be converted into coating thickness: when preset sound velocity parameters are available, it can be... First convert to Press again The coating thickness is obtained. When the sound velocity of the coating is difficult to obtain accurately, a pre-established calibration conversion relationship can be used to directly map the inter-peak distance or flight time to the estimated thickness. After obtaining the coating thickness, it is compared with the process quality criteria to output the test conclusion. The process quality criteria include at least the target thickness range, the thickness uniformity index between different test locations, and the number of abnormal peaks other than the two main maxima. When the thickness deviates from the target range, the uniformity index does not meet the requirements, or abnormal peaks indicate abnormal interface reflection, it is judged as unqualified. The test results of different batches or different times are summarized and the trends are compared, which can be used to characterize the stability of the coating process and the quality fluctuation.

[0045] In an embodiment, steps S3 to S6 are repeated for different adjacent backwall echo pairs, wherein the adjacent backwall echo pairs include: an echo pair with the first backwall echo being the earliest-arriving backwall echo and the second backwall echo being the next-arriving backwall echo, and an echo pair with the first backwall echo being the next-arriving backwall echo and the second backwall echo being the later-arriving backwall echo; and the multiple smoothed pulse sequences obtained for each echo pair are fused by one of the following: amplitude-normalized averaging, confidence-weighted summation, and peak-consistency voting.

[0046] In order to improve the stability under the conditions of echo shape change, attenuation difference, and local noise disturbance, not only one set of adjacent backwall echoes is selected in the same echo event time window, but also the subsequent processing procedure is repeated for different adjacent backwall echo pairs. The selection of adjacent backwall echo pairs follows the order of propagation time: the first backwall echo and the next-arriving backwall echo form a first echo pair, and the next backwall echo and the next-arriving backwall echo form another echo pair. Since different backwall echo pairs may have differences in energy, spectral shape, and noise influence, processing multiple echo pairs can provide more reliable results from another echo pair when a certain echo pair has low energy or distortion, thereby reducing accidental errors caused by poor selection of a single echo pair. Each echo pair independently completes the steps of intercepting, windowing, frequency domain transformation, deconvolution, and network inference to obtain a smoothed pulse sequence with the same length as the input. The main pulses in the smoothed pulse sequence still correspond to the interface reflection events, but the pulse amplitude scale, sidelobe shape, and local noise spikes obtained by different echo pairs may not be completely consistent.

[0047] To synthesize the results of multiple echo pairs into a more robust output, the multiple smoothed pulse sequences obtained from each echo pair are fused. Amplitude normalization can be performed before fusion to put different smoothed pulse sequences on a uniform scale, avoiding the situation where a sequence with higher energy dominates the result when added directly. Averaging after amplitude normalization can be used to suppress random noise and non-uniform spikes, making the uniform main pulse more prominent after averaging. Confidence-weighted summation can be used to perform a preferential fusion between echo pairs with significant quality difference, where the confidence can be determined by the prominence of the main pulse in the smoothed pulse sequence, the contrast between the main pulse and the background, the smoothing degree in the neighborhood of the main pulse, or a stability indicator after deconvolution in the frequency domain, making the sequence with higher quality contribute more. Voting according to peak consistency is used to extract the consistent interface event position from multiple smoothed pulse sequences: first determine the candidate main peak position in each smoothed pulse sequence, then cluster or determine consistency according to the consistency degree of the candidate main peak position on the time axis, take the main peak with consistent position in most sequences as the final main peak, and optionally take the statistical representative of the corresponding sequence or the corresponding peak position as the fusion result, so that the reliability of interface event positioning can be maintained even when abnormal peaks or main peak shifts occur in individual sequences. The above fusion methods can be implemented optionally, or a more suitable fusion strategy can be selected according to the noise level, echo pair quality difference, and computing resource conditions in actual applications.

[0048] In an embodiment, the training of the convolutional neural network comprises: The layered propagation medium model for characterizing the aluminum-plated plate stamping die is constructed, the layered propagation medium model comprises a coupling medium layer in contact with an ultrasonic transducer, a die base, a coating layer located at a working surface of the die base, and an environmental medium adjacent to the coating layer; a numerical acoustic field simulation based on a wave equation is used to generate a training echo, the numerical acoustic field simulation adopts a discrete solution of a finite difference time domain, and an absorbing boundary is set at a simulation domain boundary to suppress non-physical reflection; in each simulation, the sound velocity, density and attenuation parameters of the coating layer and the die base are sampled from a preset material property set, and a reference excitation waveform obtained from a die base without a coating layer is applied to the ultrasonic transducer as an incident pulse; a simulation received echo signal is obtained, the simulation received echo signal is: recording a time sequence of a pressure field at a sensor position corresponding to the ultrasonic transducer in the numerical acoustic field simulation to form a spatially sampled echo set, and spatially averaging the echo set within the sensor aperture to obtain a one-dimensional time sequence echo signal in the same format as the measured single-channel echo signal; the simulation received echo signal is processed in accordance with steps S2 to S6 to obtain a reflectivity function; a supervision label is constructed for the reflectivity function, the supervision label is a sparse pulse sequence, wherein a first non-zero pulse corresponds to a reflection event of the die base to the coating layer, and a second non-zero pulse corresponds to a reflection event of the coating layer to the environmental medium, and the sparse pulse sequence is smoothed by a kernel function to form a smoothed supervision label sequence, to complete the supervision training of the convolutional neural network.

[0049] Wherein, when constructing the layered propagation medium model, the detection object is abstracted as a structure of multiple layers of media stacked along the axial direction in the ultrasonic propagation direction, and the coupling medium between the transducer and the sample, the die base, the coating layer located at the working surface of the die base, and the environmental medium outside the coating layer are included in the same propagation path, so as to simultaneously reflect the effects of incidence, transmission, reflection and attenuation in the simulation. In order to obtain large-scale data for supervision training, a numerical acoustic field simulation based on a wave equation is used to generate a training echo, the numerical solution adopts a finite difference time domain discrete update, and an absorbing boundary layer is set at the calculation domain boundary to suppress the pollution of non-physical reflection to the echo time sequence and amplitude; the time step is determined according to the maximum wave velocity in the system and satisfies the stability constraint, which can be: ; Calculation, wherein The maximum wave speed in the medium is taken. Each simulation samples the speed of sound, density and attenuation parameters of the coating and the mold base from a preset material property set, so that the generated data covers the acoustic differences caused by material and process fluctuations; instead of using an ideal pulse, the incident excitation directly introduces the measured reference excitation waveform of the pulse echo of the uncoated base, and after time resolution matching, it is applied to the transducer model as the incident pulse, thereby preserving the spectral form and ringing characteristics of the transducer excitation in the simulation domain. The transducer can be represented by a line source in the simulation, and a Gaussian modulation is applied along the source line in the spatial domain to simulate the focused sound beam; the same source line also serves as a sensor array, and the pressure field time series recorded at the sensor positions form a set of spatially sampled echoes, which are then spatially averaged within the sensor aperture to obtain a one-dimensional time series echo signal in the same format as the measured single-channel echo signal, which is used for subsequent data processing link consistent with the measured data.

[0050] The one-dimensional time series echo signal is used to generate training samples, and a processing procedure consistent with the measured data is performed to obtain a reflectivity function, which is then input into a convolutional neural network. The supervision label is in the form of a sparse pulse sequence, and two non-zero pulses correspond to the reflection events of the mold base to the coating and the reflection events of the coating to the environment medium, respectively. The label amplitude is determined by the reflection, transmission and attenuation coefficients; in the idealized pulse propagation approximation of layered media, the propagation impulse response of each layer can be approximated as an ideal impulse: ; and the theoretical pulse expression composed of two reflection events is obtained accordingly. In order to facilitate network learning and highlight the relative amplitude and time interval, the supervision label is in the form of normalized and time-shifted theoretical impulse response: ; This formula is consistent with the explanation of step S6 above. Since sparse pulses are not friendly to network training, the sparse pulse sequence is convolved with a Gaussian kernel to obtain a smoothed supervision label sequence, which reduces the discrete sharpness of the supervision target while keeping the time position of the reflection events unchanged; at the same time, the supervision label and the reflectivity function are aligned in the sampling length, so that the sample interval of the two pulses in the label satisfies .

[0051] And the convolutional neural network is a one-dimensional full convolution structure, and a main body of the network is formed by cascading multiple convolutional layers, each convolutional layer adopts a linear rectifier activation function, and a convolution kernel is used in a front-stage convolutional layer to extract a global morphology feature of an echo, and a convolution kernel is used in a rear-stage convolutional layer to extract a local overlapping mode feature; the convolutional neural network introduces dilated convolution in part of the convolutional layers to expand a receptive field while maintaining a time resolution, and a dropout layer is arranged to reduce overfitting; a point-by-point convolutional layer is arranged at an output end of the convolutional neural network, and feature mapping is transformed into a smooth pulse sequence with the same length as a reflectivity function; in a training process, an adaptive optimization algorithm is used to minimize a mean square error loss, to smooth an error between the smooth pulse sequence and a smooth supervision label sequence, and a learning rate adaptive adjustment and an early stopping strategy are combined.

[0052] The convolutional neural network adopts a one-dimensional full convolution structure, and a main body of the network is formed by cascading multiple convolutional layers, and the convolutional layers are matched with linear rectifier activation functions to enhance nonlinear representation capabilities; a wider convolution receptive field is used in a front stage to extract global morphology and background structure of a reflectivity function, and a narrower convolution receptive field is used in a rear stage to extract details and edge features of a local overlapping pulse, and dilated convolution is introduced in part of the convolutional layers to expand a receptive field without reducing a time resolution, to improve a capture capability for long-term dependence and multi-echo structure. To reduce overfitting and improve generalization, a dropout layer is introduced in the network for regularization; a point-by-point convolutional layer is arranged at an output end to transform feature mapping into a smooth pulse sequence with the same length as the reflectivity function, so that the output directly corresponds to positions and relative amplitudes of reflection events. In a training process, an adaptive optimization algorithm is used to minimize an error between the smooth pulse sequence and a smooth supervision label sequence, taking a mean square error as a loss function and an optimization target, and a learning rate adaptive adjustment and an early stopping strategy are combined, and the training is terminated and the best parameters are reserved when a validation error no longer improves, to obtain a model more stable for positioning of the reflection events.

[0053] The technical scheme of the present application has the following beneficial effects: Compared with the prior art, the present application can more reliably extract information related to coating quality in a complex echo scene, reduce uncertainty caused by interface echo overlap, noise interference and contact coupling changes, and thus improve the stability and consistency of coating thickness evaluation results. The method does not need to sample or destructively peel the mold coating, is suitable for batch detection requirements of molds with complex geometry and large quantity, and can significantly improve process efficiency.

[0054] The present application combines signal processing mechanism with data-driven model, so that the identification of the reflection event close to the interface is more robust, and the echo shape difference under different mold base states, different coating forming conditions and different field detection conditions can be better adapted, thereby expanding the application range of the method. Therefore, not only quantitative results such as coating thickness can be output, but also more reliable data support can be provided for process stability evaluation, abnormal batch early warning and quality traceability, and process control and continuous improvement of coating process quality are promoted.

[0055] Other embodiments of the present disclosure will be apparent to those skilled in the art from consideration of the specification and practice of the present disclosure disclosed herein. This application is intended to cover any variations, uses or adaptations of the present disclosure that follow, in general, the principles of the present disclosure and include such departures from the present disclosure that come within known or customary practice within the art to which the present disclosure pertains. The specification and examples are to be regarded as illustrative only, and the true scope and spirit of the present disclosure are indicated by the claims.

[0056] It should be understood that the present disclosure is not limited to the precise structures as herein described and illustrated in the drawings, and that various modifications and changes can be made without departing from its scope. The scope of the present disclosure is limited only by the claims that follow.

Claims

1. A method for quality inspection of coating process in stamping dies for aluminized sheets, characterized in that, The stamping die includes a die base and a coating formed on the working side surface of the die base, the working side surface being opposite to the measuring side surface of the die base; the method includes the following steps S1-S7: S1, An ultrasonic transducer is coupled and arranged on the measuring side surface, and ultrasonic echo signals containing echoes from adjacent back walls are acquired using pulse echo mode, and the multiple acquisition results at the same detection position are aggregated to obtain the target echo signal. S2, perform frequency band limiting and noise suppression processing on the target echo signal to obtain a filtered echo signal, and determine the echo event time window based on the envelope of the filtered echo signal; S3, within the echo event time window, the first back wall echo and the second back wall echo that arrive sequentially are intercepted, and the first back wall echo signal and the second back wall echo signal are obtained respectively, and a boundary smoothing window is applied to the first back wall echo signal and the second back wall echo signal. S4, perform frequency domain transformation on the first back wall echo signal and the second back wall echo signal respectively to obtain the first echo spectrum and the second echo spectrum; S5, construct deconvolution based on adjacent backwall echoes: multiply the second echo spectrum with the conjugate spectrum of the first echo spectrum to obtain the cross-correlation spectrum, add the square of the amplitude of the first echo spectrum to the regularization term determined by the power spectral density of the first echo spectrum to obtain the stabilization denominator, and divide the cross-correlation spectrum by the stabilization denominator to obtain the medium transfer function spectrum. S6, perform inverse frequency domain transformation and normalize the spectrum of the medium transfer function to obtain the reflectivity function, input the reflectivity function into a convolutional neural network trained with synthetic data, and output a smooth pulse sequence of the same length as the reflectivity function; S7. In the smooth pulse sequence, determine the peak positions of the reflected pulses from the mold substrate to the coating and the reflected pulses from the coating to the environmental medium, and perform interpolation fitting on the peak positions to obtain subsampling time positioning accuracy. Calculate the time interval between the two peak positions as the coating round-trip propagation time. Convert the coating round-trip propagation time into coating thickness according to the preset sound velocity parameter or calibration conversion relationship, and compare the coating thickness with the preset process quality criteria to output the detection conclusion.

2. The method for quality inspection of coating process in stamping dies for aluminized sheets according to claim 1, characterized in that, In step S1, temporal alignment and abnormal echo removal are performed on the ultrasonic echo signals acquired at the same detection location, and average aggregation is performed among the multiple echoes after temporal alignment to reduce random noise and coupling fluctuations. The average aggregation includes one of arithmetic average and weighted average. Each coupling state between the ultrasonic transducer and the measurement side surface is defined as an independent measurement channel. The deconvolution in step S5 and the convolutional neural network inference in step S6 are performed separately in each of the independent measurement channels to obtain the reflectivity function without relying on reference samples or external reference signals.

3. The method for quality inspection of coating process in stamping dies for aluminized sheets according to claim 1, characterized in that, The frequency band limitation in step S2 is achieved using a finite impulse response bandpass filter. The passband of the finite impulse response bandpass filter is determined based on the effective operating frequency band of the ultrasonic transducer and the acoustic response of the mold substrate and coating. The filter coefficients are designed using the window function method to balance passband flatness and stopband attenuation. The envelope in step S2 is obtained by rectifying the filtered echo signal and then low-pass filtering it. The start and end positions of the echo event time window are determined by the intersection of the peak value and the threshold value of the envelope.

4. The method for quality inspection of coating process in aluminized sheet stamping dies according to claim 1, characterized in that, The extraction of the echo event time window includes: In the filtered echo signal, a family of back wall echoes arranged in the order of propagation time is identified, and two adjacent back wall echoes whose amplitudes meet the preset signal-to-noise quality conditions are selected as the first back wall echo and the second back wall echo. The boundary smoothing window is a windowing function with a smooth gradient edge, and its gradient range is controlled by a preset gradient coefficient to suppress the boundary discontinuity introduced by echo truncation without significantly weakening the energy of the echo body.

5. The method for quality inspection of coating process in stamping dies for aluminized sheets according to claim 1, characterized in that, Steps S3 to S6 are repeated for different adjacent backwall echo pairs, wherein adjacent backwall echo pairs include: The first back wall echo is taken as the first back wall echo and the next back wall echo is taken as the second back wall echo, and the subsequent pair of back wall echoes are taken as the echo pair of the first back wall echo and the second back wall echo. The multiple smooth pulse sequences obtained from each echo pair are fused. During fusion, one of the following methods is performed: averaging after amplitude normalization, weighted summation based on confidence level, or voting based on peak consistency.

6. The method for quality inspection of coating process in stamping dies for aluminized sheets according to claim 1, characterized in that, The regularization term is determined through the following process: Calculate the power spectral density of the first echo spectrum, and extract the statistical features of the power spectral density as a noise floor estimate. The statistical features include at least one of peak value, quantile value, mean value, or mean square value. The regularization term is obtained by multiplying the noise floor estimate by a preset scaling factor, and before dividing the cross-correlation spectrum by the stabilization denominator, the regularization term is added to the square of the amplitude of the first echo spectrum.

7. The method for quality inspection of coating process in stamping dies for aluminized sheets according to claim 1, characterized in that, The training of the convolutional neural network includes: A layered propagation medium model is constructed to characterize the stamping die for the aluminized sheet. The layered propagation medium model includes a coupling medium layer in contact with the ultrasonic transducer, the die substrate, a coating on the working surface of the die substrate, and an environmental medium adjacent to the coating. Training echoes are generated by numerical sound field simulation based on the wave equation. The numerical sound field simulation adopts the discrete solution method of finite difference time domain and sets an absorbing boundary at the boundary of the simulation domain to suppress non-physical reflections. In each simulation, the sound velocity, density and attenuation parameters of the coating and the mold substrate are sampled from a preset set of material properties, and the reference excitation waveform obtained from the actual measurement of the uncoated substrate is applied to the ultrasonic transducer as the incident pulse. The simulated received echo signal is obtained by recording the time series of the pressure field at the sensor position corresponding to the ultrasonic transducer in the numerical sound field simulation to form a spatially sampled echo set, and spatially averaging the echo set within the sensor aperture range to obtain a one-dimensional time series echo signal with the same format as the measured single-channel echo signal. The simulated received echo signal is processed in the same way as steps S2 to S6 to obtain the reflectivity function; a supervision label is constructed for the reflectivity function, the supervision label is a sparse pulse sequence, wherein the first non-zero pulse corresponds to the reflection event from the mold substrate to the coating, the second non-zero pulse corresponds to the reflection event from the coating to the environmental medium, and the sparse pulse sequence is smoothed by a kernel function to form a smooth supervision label sequence to complete the supervised training of the convolutional neural network.

8. The method for quality inspection of coating process in stamping dies for aluminized sheets according to claim 7, characterized in that, The convolutional neural network is a one-dimensional fully convolutional structure. Its main body is formed by cascading multiple levels of convolutional layers. Each level of convolutional layer uses a linear rectified activation function. Convolutional kernels are used in the previous level convolutional layer to extract global echo morphological features, and convolutional kernels are used in the subsequent level convolutional layer to extract local overlapping pattern features. The convolutional neural network introduces dilated convolutions in some convolutional layers to expand the receptive field while maintaining temporal resolution, and sets dropout layers to reduce overfitting. The output of the convolutional neural network is provided with a pointwise convolutional layer to transform the feature map into the smooth pulse sequence of the same length as the reflectivity function; The convolutional neural network employs an adaptive optimization algorithm to minimize the mean squared error loss during training, with the error between the smoothed pulse sequence and the smoothed supervision label sequence as the optimization objective, combined with adaptive learning rate adjustment and early stopping strategies.

9. The method for quality inspection of coating process in stamping dies for aluminized sheets according to claim 1, characterized in that, Determining the peak position includes: Search for local maxima that satisfy a preset significance condition in the smooth pulse sequence, and take the two principal maxima corresponding to the reflected pulse from the mold substrate to the coating and the reflected pulse from the coating to the environmental medium as target peaks, and perform parabolic fitting on each target peak in its neighborhood to obtain subsampling time localization. The process quality criteria include the target range of the coating thickness, the uniformity index of the coating thickness between different detection positions, and the number of abnormal peaks in the smooth pulse sequence excluding the two main maxima. When the coating thickness deviates from the target range or the uniformity index does not meet the requirements, or the interface reflection characterized by the abnormal peak is abnormal, an unqualified conclusion is output. The detection results of different batches or different times are compared to characterize the stability of the coating process.

Images