Sensor-based metal powder blend uniformity monitoring method and system

Abstract

The application relates to the technical field of sensor-based metal powder mixing uniformity monitoring, and discloses a sensor-based metal powder mixing uniformity monitoring method and system, the method comprising the following steps: acquiring an acoustic emission signal generated by metal powder tumbling and performing frequency spectrum feature transformation processing to obtain a fluctuation frequency spectrum matrix; inputting the matrix into a preset residual convolutional neural network to perform feature mapping and obtaining a powder space distribution matrix; dividing the powder space distribution matrix into regions and calculating regional segmentation thresholds, and then performing spatial aggregation analysis according to the segmentation thresholds to obtain a local aggregation matrix; calculating the variance of the local aggregation matrix as a current uniformity value, taking the average of the current uniformity values obtained through continuous monitoring to obtain an average uniformity value, and if the average uniformity value is less than or equal to a preset uniformity value, it is determined that the metal powder reaches a uniform mixing state, and an end point instruction is generated. The method can realize the judgment of whether the metal powder is uniformly mixed according to a sensor signal.

Description

Technical Field

[0001] This invention relates to the field of metal powder mixing uniformity monitoring technology, and in particular to a sensor-based method and system for monitoring metal powder mixing uniformity. Background Technology

[0002] Currently, powder metallurgy technology occupies a pivotal position in modern high-end manufacturing, and the mixing of metal powders is the core process that determines the mechanical properties and quality stability of the final product.

[0003] In existing technologies, the uniformity of metal powder mixing is typically monitored online using smart sensors. However, under the complex operating conditions of a mixer running at high speed, metal powders of different compositions and densities undergo violent three-dimensional tumbling and interweaving within a closed space. This causes the signals captured by the smart sensors to exhibit violent and irregular fluctuations, containing a large amount of interference. It is difficult to extract the spatial distribution of the powder from these signals, thus making it impossible to accurately determine the uniformity of mixing and easily leading to over-mixing or uneven mixing.

[0004] Therefore, existing technologies cannot accurately determine whether the metal powder in the closed mixing cylinder is mixed evenly based on sensor signals. Summary of the Invention

[0005] This invention provides a sensor-based method and system for monitoring the uniformity of metal powder mixing, in order to solve the problem of not being able to accurately determine whether the metal powder in a closed mixing cylinder is mixed uniformly based on sensor signals.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a sensor-based method for monitoring the uniformity of metal powder mixing, comprising: Acoustic emission sensors are deployed at equal intervals on the outer wall of the mixer to acquire acoustic emission signals generated by the tumbling of metal powder, and noise is filtered out to obtain pure emission signals; The pure transmitted signal is subjected to spectral feature transformation processing to obtain the wave spectrum matrix; The values ​​in the fluctuation spectrum matrix are normalized to obtain a normalized two-dimensional data matrix, which is then input into a preset residual convolutional neural network for feature mapping to obtain a powder spatial distribution matrix; wherein, the preset residual convolutional neural network is pre-trained under supervision using a training sample set generated by physical simulation. The powder spatial distribution matrix is ​​uniformly divided into spatial sub-regions, and the segmentation threshold of each spatial sub-region is calculated. For each spatial sub-region, data above the corresponding segmentation threshold is extracted as representative data of heavy powder for spatial clustering analysis to obtain a local clustering matrix. The variance of the local clustering matrix is ​​calculated as the current uniform value. The average uniform value is obtained by averaging the current uniform values ​​obtained by continuous monitoring. If the average uniform value is less than or equal to the preset uniform value, it is determined that the metal powder has reached a uniform mixing state, and an endpoint command is generated to control the metal powder mixer to stop operating.

[0007] In one optional implementation, the noise filtering to obtain a clean transmitted signal includes: The acoustic emission signal is converted to obtain a digital time-domain signal; The digital time-domain signal is filtered and cleaned to obtain the filtered signal; Calculate the signal-to-noise ratio (SNR) and peak factor (CPF) of the filtered signal. If the SNR is greater than a preset SNR and the CPF is greater than a preset CPF, then the filtered signal is determined to be a clean transmission signal.

[0008] In one optional implementation, the step of performing spectral feature transformation processing on the pure transmitted signal to obtain the fluctuation spectrum matrix includes: The pure transmitted signal is smoothed to obtain a smoothed signal; The smoothed signal is then transformed in the frequency domain to obtain a complex sequence in the frequency domain; The modulus of the frequency domain complex number sequence is extracted and squared to obtain the power spectrum sequence; Interpolation fitting is performed on the power spectrum sequence to obtain a continuous fluctuation spectrum curve; The continuous fluctuation spectrum curve is encapsulated as a fluctuation spectrum matrix.

[0009] In one optional implementation, the preset residual convolutional neural network includes an input layer, a convolutional layer, a pooling layer, four cascaded residual blocks, a flattening layer, a fully connected layer, and an output layer connected in sequence.

[0010] In one optional implementation, the input preset residual convolutional neural network performs feature mapping to obtain a powder spatial distribution matrix, including: After the normalized two-dimensional data matrix is ​​input through the input layer, the first feature map is obtained by sliding scan in the convolutional layer. The first feature map is input into the pooling layer for spatial dimensionality reduction to obtain the second feature map; The second feature map is input into four cascaded residual blocks for feature extraction to obtain a high-dimensional feature map. The high-dimensional feature map is input into the flattening layer for flattening to obtain the first feature vector; The first feature vector is input into a fully connected layer for feature recombination to obtain a second feature vector; The second feature vector output layer is subjected to regression calculation to obtain the powder spatial distribution matrix.

[0011] In one optional implementation, calculating the segmentation threshold for each spatial sub-region includes: Take a spatial sub-region as an example: The values ​​in this spatial sub-region are used as the candidate threshold set; Choose any candidate threshold from the candidate threshold set. If the value in the spatial sub-region is less than or equal to the candidate threshold, the value is assigned to the low-density group. If the value in the spatial sub-region is greater than the candidate threshold, the value is assigned to the high-density group. Calculate the inter-class variance of the candidate threshold after classifying the values ​​into the high-density group and the low-density group; Iterate through all candidate thresholds, repeat the above steps of grouping and calculating inter-class variance, and obtain the inter-class variance for each candidate threshold. The candidate threshold with the largest inter-class variance is selected as the segmentation threshold for the spatial sub-region.

[0012] In one optional implementation, the step of extracting data above the corresponding segmentation threshold for each spatial sub-region as representative data of heavy powder for spatial clustering analysis to obtain a local clustering matrix includes: Within each of the spatial sub-regions, data exceeding the segmentation threshold are extracted and weighted summed to obtain the weighted sum value of that sub-region. The weighted sum of all spatial sub-regions is combined and output as a local clustering matrix.

[0013] In one alternative implementation, calculating the variance of the local clustering matrix as the current uniformity includes: Calculate the arithmetic mean of all elements in the local clustering matrix to obtain the global mean; Calculate the deviation of each element in the local clustering matrix from the global mean, and square each deviation to obtain the squared deviation of each element; Sum the squared deviations of all elements to obtain the total sum of squared deviations; The current uniform value is obtained by dividing the sum of squared total deviations by the total number of elements in the local clustering matrix.

[0014] In an optional implementation, the method further includes: If the average uniform value is greater than the preset uniform value, it is determined that the current metal powder has not reached a uniform mixing state, and the system continues to collect acoustic emission signals for continuous monitoring.

[0015] Secondly, the present invention provides a sensor-based metal powder mixing uniformity monitoring device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the sensor-based metal powder mixing uniformity monitoring method described above.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention obtains the acoustic emission signal generated by the tumbling of metal powder, and obtains the wave spectrum matrix after noise filtering and spectral feature transformation. Then, the normalized matrix is ​​input into a preset residual convolutional neural network for feature mapping, and the output powder spatial distribution matrix is ​​obtained. This solves the signal interference problem and utilizes the powerful image feature extraction capability of the convolutional neural network to deduce the spatial distribution of powder, comprehensively reflecting the powder distribution in various regions inside the closed container, and providing a global spatial distribution information basis for the accurate judgment of subsequent mixing uniformity.

[0017] (2) This invention divides the powder spatial distribution matrix evenly into multiple spatial sub-regions, dynamically calculates the segmentation threshold for each sub-region using adaptive threshold segmentation, extracts the numerical values ​​representing heavy powder, and calculates and outputs a local aggregation matrix. The above operations automatically separate light powder from heavy powder, eliminate interference from light powder, accurately reflect the degree of aggregation of heavy powder in each spatial region, provide more targeted characteristic data for accurate evaluation of mixing uniformity, and thus improve the accuracy of monitoring.

[0018] (3) This invention calculates the variance of the local aggregation matrix as the current uniformity value, calculates the average uniformity value by averaging multiple current uniformity values ​​obtained through continuous monitoring, and then compares the average uniformity value with a preset uniformity threshold to determine whether the metal powder is uniform. This realizes the transformation of the local aggregation matrix into an intuitive uniformity index, and filters out the instantaneous fluctuations caused by local powder agglomeration through sliding window smoothing, making the judgment result of mixing uniformity more stable and reliable. Attached Figure Description

[0019] Figure 1 This is a schematic flowchart of the sensor-based metal powder mixing uniformity monitoring method provided in the first embodiment of the present invention. Detailed Implementation

[0020] 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Reference Figure 1 The first embodiment of the present invention provides a sensor-based method for monitoring the mixing uniformity of metal powder, comprising the following steps: S11, acoustic emission sensors are deployed at equal intervals on the outer wall of the mixer to acquire acoustic emission signals generated by the rolling of metal powder, and noise is filtered out to obtain pure emission signals; S12, Perform spectral feature transformation processing on the pure transmitted signal to obtain the wave spectrum matrix; S13, normalize the values ​​in the fluctuation spectrum matrix to obtain a normalized two-dimensional data matrix, and input it into a preset residual convolutional neural network for feature mapping to obtain a powder spatial distribution matrix; wherein, the preset residual convolutional neural network is pre-trained under supervision using a training sample set generated by physical simulation. S14, the powder spatial distribution matrix is ​​uniformly divided into spatial sub-regions, and the segmentation threshold of each spatial sub-region is calculated. For each spatial sub-region, data higher than the corresponding segmentation threshold is extracted as representative data of heavy powder for spatial aggregation analysis to obtain a local aggregation matrix. S15, calculate the variance of the local aggregation matrix as the current uniform value, take the average value of the current uniform value obtained by continuous monitoring to obtain the average uniform value, if the average uniform value is less than or equal to the preset uniform value, it is determined that the metal powder has reached a uniform mixing state, and an endpoint command is generated to control the metal powder mixer to stop operating.

[0022] In step S11, it should be noted that the acoustic emission signal is acquired by deploying four piezoelectric acoustic sensors at equal intervals on the outer wall of the enclosed mixer, and collecting the acoustic emission signal generated by the tumbling of metal powder by setting the sampling frequency through a programmable logic controller. To ensure signal quality, the system calculates the signal-to-noise ratio (SNR) of the signal collected by each sensor channel and selects the signal from the channel with the highest SNR as the acoustic emission signal for subsequent processing; or, it performs an equal-weighted average of the signals from all channels to obtain a comprehensive acoustic emission signal representing the overall state of the mixer. This embodiment adopts the former, that is, selecting the signal from the channel with the highest SNR for processing.

[0023] In one implementation, noise filtering is performed to obtain a clean transmitted signal, including: The acoustic emission signal is converted to obtain a digital time-domain signal; The digital time-domain signal is filtered and cleaned to obtain the filtered signal; Calculate the signal-to-noise ratio (SNR) and peak factor (CPF) of the filtered signal. If the SNR is greater than a preset SNR and the CPF is greater than a preset CPF, then the filtered signal is determined to be a clean transmission signal.

[0024] It should be noted that after the sensor captures the acoustic emission signal, the analog-to-digital converter in the data acquisition terminal automatically samples and quantizes the acoustic emission signal into a digital time-domain signal. In the digital time-domain signal, in addition to the effective acoustic emission signal generated by the collision of metal powder, various background noises are also present, such as mechanical vibration noise generated by the rotation of the mixer motor, reducer, and drive shaft, and random noise generated by the friction between the metal powder and the cylinder wall. If these noises are not filtered out, the accuracy of subsequent spectrum analysis and feature extraction will be severely reduced, making it impossible to establish a reliable mapping relationship between acoustic features and the spatial distribution of powder.

[0025] In this example, a sixth-order Butterworth infinite impulse response bandpass filter is used for cleaning. This filter has a maximally flat amplitude response characteristic within its passband, which can preserve the signal waveform to the greatest extent possible without distortion. The core parameters of the filter are configured based on the spectral characteristics of the digital time-domain signal from the metal powder collision. For example, for common metal powders such as iron powder and copper powder, the effective acoustic emission signal energy generated by the collision is mainly concentrated in the mid-frequency band between 10kHz and 150kHz. Therefore, the lower cutoff frequency of the filter can be set to 10kHz and the upper cutoff frequency to 150kHz, which significantly suppresses frequency components below the lower cutoff frequency and above the upper cutoff frequency, thereby filtering out background noise and obtaining a filtered signal that mainly retains the range from the lower cutoff frequency to the upper cutoff frequency.

[0026] Next, to confirm the signal quality of the filtered signal, the system calculates the signal-to-noise ratio (SNR) and peak factor (CRF) of the filtered signal. The filtered signal sequence can be converted to the frequency domain using a Fast Fourier Transform (FFT) to obtain the amplitude spectrum of each frequency component. Within the effective frequency band from the lower to the upper cutoff frequency, the amplitudes of all frequency components within the band are squared and summed to obtain the signal power. The noise power is calculated in the remaining frequency bands, and the SNR is... ,in, For signal power, This represents the noise power. The peak factor is the ratio of the peak value in the filtered signal sequence to the root mean square value of the data in the filtered signal sequence.

[0027] Then, the signal-to-noise ratio (SNR) is compared with a preset SNR, and the peak factor is compared with a preset peak factor. If the SNR is greater than the preset SNR and the peak factor is greater than the preset peak factor, it indicates that effective collision energy dominates the signal, noise interference has been effectively suppressed, and there is obvious transient impact in the signal, confirming that the sensor has indeed captured real powder particle collision signals. In this case, the filtered signal is determined to be a pure emission signal. If the SNR is less than the preset SNR or the peak factor is less than the preset peak factor, it indicates that there is still significant noise interference in the filtered signal or a lack of effective collision information of metal powder. In this case, the system determines that the signal is invalid and re-acquires the acoustic emission signal.

[0028] The preset signal-to-noise ratio (SNR) was determined experimentally. At least 100 sets of background noise signals and 100 sets of effective collision signals were collected from the mixer during no-load operation and full-load operation. At multiple candidate thresholds (e.g., 35dB, 40dB, 45dB, 50dB), the effective signal throughput and background noise misclassification rate were calculated for each candidate threshold. For example, for a candidate threshold of 35dB, the SNR of each collected signal was compared with it. For effective collision signals, if the SNR ≥ 35dB, the system determined it as a valid signal; if the SNR < 35dB, the system determined it as an invalid signal, and the effective signal throughput was calculated by dividing the number of valid signals by the number of effective collision signals. For background noise signals, if the SNR ≥ 35dB, the system misclassified them as valid signals; if the SNR < 35dB, the system correctly classified them as invalid signals, and the background noise misclassification rate was calculated by dividing the number of misclassified valid signals by the number of background noise signals. Then, any candidate threshold with an effective signal throughput ≥ 95% and a background noise false positive rate ≤ 5% is selected as the preset signal-to-noise ratio. In this embodiment, the preset signal-to-noise ratio is 40dB. The value of the preset peak factor is determined based on the peak factor of noise signals under normal circumstances. Generally, the peak factor of noise signals such as Gaussian white noise and mechanical steady-state vibration is usually between 3.0 and 4.0. In order to effectively distinguish the transient impact signal generated by powder collision from the background noise, this embodiment sets the preset peak factor to 4.0.

[0029] In step S12, spectral feature transformation processing is performed on the pure transmitted signal to obtain a wave spectrum matrix, including: The pure transmitted signal is smoothed to obtain a smoothed signal; The smoothed signal is then transformed in the frequency domain to obtain a complex sequence in the frequency domain; The modulus of the frequency domain complex number sequence is extracted and squared to obtain the power spectrum sequence; Interpolation fitting is performed on the power spectrum sequence to obtain a continuous fluctuation spectrum curve; The continuous fluctuation spectrum curve is encapsulated as a fluctuation spectrum matrix.

[0030] It should be noted that the obtained clean transmitted signal is first subjected to weighted smoothing using a Hanning window. This is because if the clean transmitted signal is directly transformed in the frequency domain, it is prone to discontinuities at the truncation boundaries. The Hanning window can smoothly attenuate the sampling points at both ends of the clean transmitted signal, making the boundaries continuous. In this example, the Hanning window function is multiplied point by point with the clean transmitted signal to obtain the windowed signal sequence, i.e., the smoothed signal.

[0031] Next, since the smoothed signal does not directly reveal the frequency characteristics of powder collisions, a Fast Fourier Transform (FFT) is performed on it to transform it from the time domain to the frequency domain, obtaining a complex sequence in the frequency domain containing amplitude and phase information. This complex sequence allows for the analysis of the energy distribution of different frequency components. The FFT formula is as follows: ; in, The sequence number of the sampling point of the smoothed signal. The number of sampling points for the smoothed signal. For smoothing signals, imaginary unit , The index of the frequency domain output point. For the first The frequency domain complex number corresponding to each frequency point, each It contains a real part and an imaginary part, and can be represented as ,in The real part represents the amplitude of the cosine component in the signal. imaginary unit , The imaginary part represents the amplitude of the sinusoidal component in the signal. Together, they completely describe the amplitude and phase information of this frequency component. Next, the magnitude of the complex sequence in the frequency domain is extracted, as shown in the following formula: in, For the frequency domain complex sequence, the first... The modulus of a complex number in the frequency domain. For the frequency domain complex sequence, the first... The real part of a complex number in the frequency domain, For the frequency domain complex sequence, the first... The imaginary part of a complex number in the frequency domain.

[0032] After calculating all the moduli of the complex number sequence in the frequency domain, the moduli are squared to obtain the power spectrum sequence. The power spectrum sequence can intuitively reflect the energy distribution of powder collisions with frequency. Since the obtained power spectrum sequence consists of isolated points, interpolation fitting is required to transform it into a continuous fluctuating spectrum curve, so as to reflect the continuous change of energy with frequency and facilitate subsequent image processing.

[0033] In this embodiment, considering that the width of the fluctuation spectrum matrix of the subsequent input residual convolutional neural network is 256, the target number of points is set to 256 to ensure that each pixel column corresponds to an interpolation point during frequency axis mapping. The frequency range of the power spectrum sequence is... Its range is limited by the passband of the aforementioned filter, that is... The lower cutoff frequency is 10kHz. The upper cutoff frequency is set to 150kHz. The frequency axis of the power spectrum sequence is linearly mapped to the interval [0, 255], such that the minimum frequency point in the power spectrum sequence corresponds to index 0, the maximum frequency point corresponds to index 255, and the remaining integer indices are mapped to actual frequency values. The formula is as follows: .in, For the actual frequency corresponding to index i, , The minimum frequency is 10kHz and the maximum frequency is 150kHz, respectively.

[0034] Then, search for the match in the power spectrum sequence. Two adjacent frequency points are calculated using linear interpolation. The power spectral density at that location is given by the formula: .in, For the frequency value corresponding to index i, The power spectral density value of the index to be determined is... and For the power spectrum sequence and Two adjacent frequency points, and This represents the power spectral density value corresponding to the frequency point. Through the above processing, a continuous fluctuation spectrum curve consisting of 256 equally spaced points is finally obtained.

[0035] It's worth noting that encapsulating the continuous fluctuation spectrum curve into a fluctuation spectrum matrix involves first creating a zero matrix, the size of which must match the input size of the subsequent preset residual convolutional neural network. Then, the frequency x-axis of the continuous fluctuation spectrum curve is linearly mapped to the column index of the matrix, and the power spectrum y-axis is linearly mapped to the row index of the matrix. The power spectrum is then linearly mapped to grayscale values ​​from 0 to 255 as the corresponding coordinate values. The mapping formula is as follows. ; ; ; in, For column indexes, For a frequency in the continuous wave spectrum curve, , The minimum frequency of the continuous wave spectrum curve is 10kHz, and the maximum frequency is 150kHz, respectively. The width of the zero matrix. For row index, This is a power spectral value of a continuous fluctuation spectrum curve. , These are the minimum and maximum power spectral values ​​of the continuous wave spectrum curve, respectively. The height of the zero matrix is ​​256. Mapped coordinates The grayscale values ​​are obtained by mapping each point on the continuous wave spectrum curve. This yields the wave spectrum matrix.

[0036] Since the above mapping process only maps points on the wave spectrum curve to sparse single pixels, most areas in the resulting wave spectrum matrix may be blank backgrounds with low information density, which is not conducive to the subsequent extraction of spatial texture features by convolutional neural networks. Therefore, spatial diffusion processing can be performed on each mapped coordinate point. And its gray value s, with that point as the center, outwards with a radius of Gaussian function. The grayscale value is assigned within the circular neighborhood using the following formula: .

[0037] in, Coordinates in the wave spectrum matrix grayscale value at that location For this coordinate To the center point Euclidean distance, is the width parameter of the Gaussian kernel. The value of is not less than half the mapping distance between adjacent mapping points on the continuous wave spectrum curve in the image, ensuring that the diffusion regions of adjacent mapping points overlap to form continuous energy distribution characteristics, and the radius Take 3 In this embodiment, the mapping distance between adjacent mapping points in the image is approximately 2 pixels. 1.5 pixels, radius A value of 4.5 pixels is selected. For overlapping areas of different frequency diffusion regions, the maximum grayscale value is selected to retain the strongest energy response. Through the above spatial diffusion processing, the final wave spectrum matrix is ​​presented as a densely filled two-dimensional grayscale image.

[0038] It is worth noting that the fluctuation spectrum matrix is ​​formally equivalent to a grayscale image. The local continuity of the horizontal frequency axis corresponds to the proximity of signals with similar frequencies generated by collisions of different powder particles in the spectrum, while the local continuity of the vertical energy axis corresponds to the gradual change in collision intensity. Convolutional neural networks (CNNs) can effectively capture local gradients, peak shapes, and texture patterns through their convolutional kernels. These spectral patterns have an inherent physical correlation with the size distribution, collision velocity, and spatial motion state of the powder particles. Therefore, using the fluctuation spectrum matrix as input and leveraging the powerful structured feature extraction capabilities of CNNs to regress and predict the spatial distribution matrix of the powder has both physical basis and feasibility.

[0039] In step S13, the values ​​in the fluctuation spectrum matrix are first normalized, mapping their values ​​to the interval between 0.0 and 1.0. The entire fluctuation spectrum matrix is ​​traversed to find the maximum and minimum values ​​of all grayscale values. Then, for each element in the matrix, the normalized value is calculated according to the normalization formula. Finally, the normalized value is written back to the corresponding position in the matrix, resulting in a normalized two-dimensional data matrix. All values ​​in the normalized two-dimensional data matrix are unified to the [0,1] interval, which can eliminate the influence of dimensions on model inference. The normalization formula is: .in, The first in the wave spectrum matrix Line number The normalized gray values ​​of the column. The first in the wave spectrum matrix Line number The grayscale value of the column, The minimum gray value in the wave spectrum matrix. This represents the maximum grayscale value in the fluctuation spectrum matrix.

[0040] In one implementation, the pre-defined residual convolutional neural network includes an input layer, a convolutional layer, a pooling layer, four cascaded residual blocks, a flattening layer, a fully connected layer, and an output layer connected in sequence.

[0041] In this example, the normalized two-dimensional data matrix is ​​256×256 in size and is received by the input layer of the pre-defined residual convolutional neural network. The convolutional layer contains 64 3×3 convolutional kernels with a stride of 1 and uses same padding to extract local energy mutation features. After this convolutional layer, the output feature map size is 256×256×64. The pooling layer is 2×2 max pooling with a stride of 2 for spatial dimensionality reduction. After pooling, the feature map size is halved to 128×128×64. In the four cascaded residual blocks, the first residual block has 64 input channels and 128 output channels. The first 3×3 convolutional layer has a stride of 2, the second 3×3 convolutional layer has a stride of 1, and skip connections are introduced. After the first residual block, the feature map size is reduced from 128×128×64 to 64×64×128. The second residual block has 128 input channels and 256 output channels. The stride of the first 3×3 convolutional layer is 2, and the stride of the second 3×3 convolutional layer is 1. After the second residual block, the feature map size is reduced from 64×64×128 to 32×32×256.

[0042] The third residual block has 256 input channels and 512 output channels. The first 3×3 convolutional layer has a stride of 2, and the second 3×3 convolutional layer has a stride of 1. After the third residual block, the feature map size is reduced from 32×32×256 to 16×16×512. The fourth residual block has 512 input channels and 512 output channels. Both 3×3 convolutional layers have a stride of 1, maintaining the feature map size at 16×16×512. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function. The flattening layer is used to convert the 16×16×512 feature map output from the fourth residual block into a one-dimensional feature vector. The fully connected layer contains 512 nodes, reducing the one-dimensional feature vector to 512 dimensions and performing feature reorganization. The output layer contains 256 nodes and uses the Sigmoid activation function to compress the output value of each node to the [0,1] interval. The output is reshaped into a 16×16 powder spatial distribution matrix.

[0043] It is worth noting that the preset residual convolutional neural network is obtained through pre-training. During the training process, this invention constructs a three-dimensional geometric model of a mixer that is completely consistent with the real production environment in simulation software (such as the EDEM-Fluent coupling module), including the shape, size, rotational speed of the stirring blades, and the volume of the mixing cylinder. The physical parameters of the metal powder are set, including powder density, particle size distribution, coefficient of restitution, and coefficient of static friction. The total powder mass, mixing ratio, and initial random packing state are also set.

[0044] In this embodiment, the mixer's geometric model adopts a double-cone type mixer with a total cylinder volume of 50L. The stirring blades are three-bladed blades with a diameter of 200mm, and the rotation speed is set to 60r / min. For the metal powders, iron powder with a density of 7800kg / m³, a particle size distribution of 50-150μm, and a median diameter of 90μm is used to represent the heavy powder, while aluminum powder with a density of 2700kg / m³, a particle size distribution of 30-100μm, and a median diameter of 60μm is used to represent the light powder. The total filling mass is 15kg, and the mixing ratio (mass ratio) is set to 7 parts iron powder and 3 parts aluminum powder. Simultaneously, the coefficient of restitution between iron powder and aluminum powder is set to 0.6, the coefficient of restitution between iron powder and aluminum powder to 0.5, the coefficient of restitution between aluminum powder and aluminum powder to 0.4, and the coefficient of restitution between powder and cylinder wall collision to 0.7. The static friction coefficient between iron powder and iron powder is 0.4, between iron powder and aluminum powder is 0.35, between aluminum powder and aluminum powder is 0.3, and between powder and cylinder wall is 0.25. The mixing process is run in a simulation environment with a simulation time step of 1×10⁻⁶. -5 It outputs data every 0.1 seconds.

[0045] Simultaneously, data acquisition of input samples and label samples is carried out. In the EDEM-Fluent simulation software, at the sensor location on the outer wall of the simulated mixer, which coincides with the actual deployment location, a 1×10⁻⁶ sensor is used. -5 The system outputs contact force data at each sampling moment, with a fixed time step of seconds, for all metal powder particles colliding with the mixer drum wall. This data includes the magnitude of the contact force and the coordinates of the contact position. The instantaneous collision event between the particle and the drum wall is automatically identified by EDEM-Fluent's contact detection algorithm, which triggers the event when the distance between the particle and the drum wall surface is less than the particle radius.

[0046] Then, based on Hertzian contact theory, the instantaneous collision of each particle is equivalent to a point source excitation acting on the cylinder wall surface. Next, using the Green's function method as the elastic wave propagation model, the displacement response of this point source excitation propagating along the cylinder wall to the sensor installation position is calculated. First, the P-wave velocity and S-wave velocity are calculated according to the cylinder wall material, as shown in the following formulas. in, For P-wave velocity, For S-wave velocity, The density of the cylinder wall material, For elastic modulus, Let be Poisson's ratio. Next, the distance between the collision contact position coordinates and the sensor installation position coordinates is calculated using Euclidean algorithm. Then, the displacement response is calculated using the distance, P-wave velocity, and S-wave velocity, as shown in the formula. in, For displacement response, For P-wave velocity, Let S be the wave velocity of the S-wave, and F be the contact force during the collision. This represents the distance between the collision contact point coordinates and the sensor installation point coordinates. It should be noted that the cylinder wall material parameters are set according to the actual equipment; in this example, the carbon steel density is 7800 kg / m³, the elastic modulus is 200 GPa, and the Poisson's ratio is 0.3.

[0047] Finally, the displacement response at the sensor location is convolved in the time domain with the sensitivity response curve of the piezoelectric acoustic emission sensor to obtain the simulated sensor output voltage signal, which is the virtual acoustic emission signal. The sensitivity response curve adopts the standard parameters of the R15a type resonant acoustic emission sensor commonly used in this field, with a frequency response range of 50kHz to 400kHz and a peak sensitivity of 65dB re 1V / (m / s) at the center frequency of 150kHz. By performing an inverse Fourier transform on the frequency domain sensitivity curve, the time domain form of the sensitivity response curve is obtained.

[0048] The virtual acoustic emission signal is processed through steps S11 and S12 to obtain a wave spectrum matrix, which serves as the training input sample. Simultaneously, the entire internal space of the mixer is discretized into a 16×16 uniform grid. The proportion of the mass or volume of heavy powder within each grid cell to the total mass or volume of powder in that grid is calculated. This proportion is normalized to the [0,1] interval and then arranged according to the grid spatial position to directly form a 16×16 two-dimensional matrix, which is the true powder spatial distribution matrix and serves as the training label sample.

[0049] Next, the simulation was continuously run, covering the entire process from the initial segregation state to complete homogeneous mixing. At least 10,000 synchronized (input sample, label sample) data pairs were generated. These were divided into training, validation, and test sets in a 7:2:1 ratio. The Kaiming initialization method was used to initialize the pre-defined residual convolutional neural network. The training set samples were input into the network, and after calculation by each layer, the predicted powder spatial distribution matrix was output. The mean squared error was used as the loss function to calculate the error between the predicted value and the true label. The Adam optimizer was used, with an initial learning rate of 0.001 and a batch size of 32, for 200 iterations. After each iteration, the model performance was evaluated using the validation set. The model parameters with the minimum loss on the validation set were saved as the final pre-defined residual convolutional neural network. The model was considered successfully trained if the average structural similarity index between the predicted powder spatial distribution matrix and the true label was greater than 0.85, as verified by the test set.

[0050] In this embodiment, a preset residual convolutional neural network is input for feature mapping to obtain a powder spatial distribution matrix, including: After the normalized two-dimensional data matrix is ​​input through the input layer, the first feature map is obtained by sliding scan in the convolutional layer. The first feature map is input into the pooling layer for spatial dimensionality reduction to obtain the second feature map; The second feature map is input into four cascaded residual blocks for feature extraction to obtain a high-dimensional feature map. The high-dimensional feature map is input into the flattening layer for flattening to obtain the first feature vector; The first feature vector is input into a fully connected layer for feature recombination to obtain a second feature vector; The second feature vector output layer is subjected to regression calculation to obtain the powder spatial distribution matrix.

[0051] It should be noted that the normalized two-dimensional data matrix is ​​input into the input layer and then fed into the convolutional layer. The convolutional layer extracts local features of the spectral image through convolution operations, including edges, peaks, and gradient changes of the spectral curve, providing a basic local feature representation for subsequent deep feature extraction. Each pixel value in the first feature map obtained by the convolutional layer represents the intensity of energy abrupt changes in the corresponding local region in the input normalized two-dimensional data matrix. Next, the output first feature map is input into the pooling layer, where max pooling retains the strongest response in each local region, resulting in the second feature map.

[0052] Next, the output second feature map is sequentially input into four cascaded residual blocks. The first residual block combines shallow local features to form energy cluster features with a certain bandwidth; the second residual block begins to establish a rough correspondence between frequency domain energy distribution and spatial location; the third residual block fuses frequency domain features with the geometric structure information of the mixer to form a spatial distribution hypothesis; the fourth residual block outputs a high-dimensional abstract feature map, where each channel represents a latent attribute related to powder distribution. By extracting multi-level abstract features from simple to complex through a deep convolutional network, the latent mapping relationship between spectral energy distribution and powder spatial distribution is gradually established, resulting in a high-dimensional feature map.

[0053] Then, the high-dimensional feature map is input into a flattening layer. The flattening layer arranges all elements in the high-dimensional feature map into a one-dimensional vector, obtaining the first feature vector. The first feature vector is then input into a fully connected layer. Through linear transformation and nonlinear activation, the high-dimensional first feature vector is reduced and recombined to obtain a low-dimensional, compact second feature vector. The second feature vector is then input into the output layer. Linear regression is used to generate corresponding values ​​from the second feature vector, which are then reshaped into a two-dimensional matrix, namely the powder spatial distribution matrix. The value at each coordinate point in the powder spatial distribution matrix represents the relative concentration and aggregation degree of the powder in the corresponding spatial grid inside the closed mixer. A larger value indicates a higher powder concentration and stronger aggregation in that spatial region; a smaller value indicates a lower concentration and sparser distribution.

[0054] In step S14, the powder spatial distribution matrix is ​​uniformly divided into sub-regions of a preset size. The number of sub-regions in the row direction is equal to the ratio of the number of rows in the powder spatial distribution matrix to the number of rows in the sub-regions, and the number of sub-regions in the column direction is equal to the ratio of the number of columns in the powder spatial distribution matrix to the number of columns in the sub-regions. After division, the sub-regions do not overlap, and each sub-region represents the powder concentration distribution in a region within the mixer. The union of all sub-regions completely covers the entire powder spatial distribution matrix. In this embodiment, the powder spatial distribution matrix is ​​16×16 in size, and the sub-region size is 4×4, so the number of sub-regions in the row direction is 4, and the number of sub-regions in the column direction is 4.

[0055] The preset sub-region size must simultaneously meet the following requirements: the number of rows and columns of the sub-region must be a divisor of the number of rows and columns of the powder spatial distribution matrix; the number of sub-regions obtained by dividing according to the preset sub-region size must be between 4 and 64; and the number of points contained in each sub-region must be between 9 and 64. This is because too few sub-regions will result in an insufficient sample size for variance calculation, leading to unreliable statistical results; too many sub-regions will result in an excessively large local clustering matrix, increasing the computational burden. Furthermore, too few points in each sub-region will also lead to insufficient statistical stability; too many points in each sub-region will fail to effectively reflect local clustering characteristics. If multiple satisfactory values ​​exist, for ease of calculation, the value with the smallest total number of sub-regions after division can be selected as the preset value.

[0056] In one implementation, the segmentation threshold for each spatial sub-region is calculated, including: Take a spatial sub-region as an example: The values ​​in this spatial sub-region are used as the candidate threshold set; Choose any candidate threshold from the candidate threshold set. If the value in the spatial sub-region is less than or equal to the candidate threshold, the value is assigned to the low-density group. If the value in the spatial sub-region is greater than the candidate threshold, the value is assigned to the high-density group. Calculate the inter-class variance of the candidate threshold after classifying the values ​​into the high-density group and the low-density group; Iterate through all candidate thresholds, repeat the above steps of grouping and calculating inter-class variance, and obtain the inter-class variance for each candidate threshold. The candidate threshold with the largest inter-class variance is selected as the segmentation threshold for the spatial sub-region.

[0057] It should be noted that adaptive thresholding is used to independently calculate the segmentation threshold for each spatial sub-region. For a spatial sub-region, all values ​​are used as a candidate threshold set. If duplicate values ​​exist, they are removed, retaining only unique values ​​as candidate thresholds. For any candidate threshold in the candidate threshold set, all values ​​within the spatial sub-region are divided into two groups according to the following rules: values ​​less than or equal to the candidate threshold are assigned to the low-density group, representing light powder or low-concentration areas; values ​​greater than the candidate threshold are assigned to the high-density group, representing heavy powder or high-concentration areas. Next, the inter-class variance between the low-density and high-density groups is calculated. According to the OTSU algorithm, the standard definition of inter-class variance is... ,in, After algebraic simplification, the equivalent form can be obtained: in, To calculate the inter-class variance, This represents the number of values ​​in the low-density group. This represents the number of values ​​in the high-density group. This is the average of all values ​​in the low-density group. The value is the average of all values ​​in the high-density group. Then, all candidate thresholds are iterated over, and the steps of grouping and calculating inter-class variance are repeated to obtain the inter-class variance for each candidate threshold. After iteration, the inter-class variances corresponding to all candidate thresholds are compared, and the candidate threshold with the largest inter-class variance is selected as the segmentation threshold for that spatial sub-region. If multiple candidate thresholds have the same maximum inter-class variance, any one of them can be selected as the segmentation threshold. The largest inter-class variance means the greatest separation between the high-density and low-density groups, and the selected segmentation threshold best separates the heavy powder from the light powder. Since the powder concentration distribution may differ in different spatial sub-regions, the above process of calculating the segmentation threshold is performed independently for each spatial sub-region to obtain the segmentation threshold for each spatial sub-region. This adapts to the differences in powder concentration in different spatial regions and improves the accuracy of heavy powder extraction.

[0058] In this example, for each spatial sub-region, data above the corresponding segmentation threshold are extracted as representative data of heavy powder for spatial clustering analysis, resulting in a local clustering matrix, including: Within each of the spatial sub-regions, data exceeding the segmentation threshold are extracted and weighted summed to obtain the weighted sum value of that sub-region. The weighted sum of all spatial sub-regions is combined and output as a local clustering matrix.

[0059] It should be noted that in metal powder metallurgy, the mixture typically consists of heavy powders such as iron powder and tungsten powder, and light powders such as aluminum powder and graphite powder. The mechanical properties of the final product are mainly determined by the uniformity of the distribution of the heavy powder. However, due to their high density and inertia, heavy powders tend to agglomerate at the edges or bottom under centrifugal force during the mixing process, forming localized clusters. If the heavy powder is uniformly mixed, the light powder is usually also uniform. Therefore, it is necessary to extract and monitor the heavy powder in each spatial sub-region using a segmentation threshold.

[0060] It should be noted that for each spatial sub-region, areas above the segmentation threshold represent regions with high local powder concentration or aggregation. Heavy powders tend to agglomerate in these areas, typically significantly impacting mixing uniformity. Areas below the threshold represent sparsely distributed powder regions. Therefore, after obtaining the segmentation threshold for each spatial sub-region, data values ​​above the threshold are identified as representative of heavy powder and retained; data values ​​below or equal to the threshold are identified as interference data from light powder and filtered out. The retained data above the segmentation threshold are then weighted and summed. Because the powder at the center of the spatial sub-region is more thoroughly mixed and representative during the mixing process, the center coordinates are set to the highest weight of 1, with the weight smoothly decreasing towards the edges, and the lowest weight being 0.5. The Euclidean distance from each data point above the segmentation threshold within the spatial sub-region to the center is calculated, and the weight of each data point is calculated using a linear interpolation formula. The formula is as follows: .

[0061] in, Weights for data coordinate points that are above the segmentation threshold. This represents the Euclidean distance from the corresponding coordinate point to the center. This represents the maximum Euclidean distance from the coordinate point within the spatial sub-region to the center. The weighting coefficient of 0.5 is derived from a mixing dynamics simulation experiment. In a typical double-cone or V-type mixer, the flow field generated by the mixing blades results in a mixing effect difference of approximately 2:1 between the central region and the wall region, meaning that the mixing uniformity at the boundary is typically 50% of that in the central region.

[0062] Next, the data exceeding the segmentation threshold within each spatial sub-region are multiplied by their corresponding weights and then summed to obtain a weighted sum for that sub-region. This weighted sum represents the degree of aggregation of heavy powder within that spatial sub-region. The weighted sums calculated for each spatial sub-region are then combined into a new two-dimensional matrix, namely the local aggregation matrix, according to the spatial position of each sub-region in the original powder spatial distribution matrix.

[0063] In step S15, the variance of the local clustering matrix is ​​calculated as the current uniform value, including: Calculate the arithmetic mean of all elements in the local clustering matrix to obtain the global mean; Calculate the deviation of each element in the local clustering matrix from the global mean, and square each deviation to obtain the squared deviation of each element; Sum the squared deviations of all elements to obtain the total sum of squared deviations; The current uniform value is obtained by dividing the sum of squared total deviations by the total number of elements in the local clustering matrix.

[0064] It should be noted that calculating the variance of the local clustering matrix first involves summing the values ​​of all elements in the local clustering matrix and averaging them to obtain the global mean. Then, the deviation of each element in the local clustering matrix is ​​calculated by subtracting it from the global mean, and each deviation is squared to obtain the squared deviation of each element. Finally, the squared deviations of all elements are summed to obtain the total sum of squared deviations. This total sum of squared deviations is then divided by the total number of elements in the local clustering matrix to obtain the variance, which is the current uniformity. The variance reflects the dispersion of the degree of heavy powder aggregation among different spatial sub-regions. A smaller variance indicates that the weighted sum values ​​of each spatial sub-region are closer, representing a more uniform distribution of heavy powder; a larger variance indicates greater differences in the weighted sum values ​​of each sub-region, representing the presence of heavy powder aggregation. Therefore, the variance is used as the current uniformity value for subsequent mixing uniformity determination.

[0065] It should be noted that during the actual mixing process, due to random collisions and local agglomeration of powder particles, the current uniform value calculated from a single sampling may exhibit instantaneous fluctuations. For example, at a certain moment, local powder agglomeration may cause a sudden increase in variance, even though the overall mixing state is actually uniform. Directly using a single sampling value for judgment may lead to misjudgment. Therefore, it is necessary to smooth multiple current uniform values ​​over a continuous time period to eliminate transient fluctuation interference.

[0066] To eliminate transient fluctuations that may occur during a single sampling, the system calculates the average of multiple current uniform values ​​obtained over a continuous time period. In this example, the data acquisition engine continuously executes steps S11 to S14 once per second, obtaining one current uniform value each time and pushing it into a pre-allocated buffer queue with a capacity of 10. When the buffer queue is full, the arithmetic mean of the 10 current uniform values ​​in the buffer queue is calculated as the average uniform value. Each time a new current uniform value is obtained, it is pushed to the end of the queue, while automatically removing old data from the head of the queue, maintaining a queue length of 10, and the average uniform value of the buffer queue is recalculated. These operations effectively filter out transient high-frequency noise caused by localized powder agglomeration, random collisions, etc., ensuring that the average uniform value truly reflects the overall mixing uniformity and stability of the mixer during continuous operation.

[0067] Next, the system compares the average uniform value with the preset uniform value. If the average uniform value is less than or equal to the preset uniform value, it determines that the current metal powder has reached a uniform mixing state. At this time, the system generates an endpoint command to terminate the cyclic acquisition task of acoustic emission signals. The command is sent to the frequency converter driver of the mixer through the industrial control bus, so that the frequency converter driver starts the deceleration program and smoothly reduces from the rated frequency to zero at the set decay rate. At the same time, the mixing uniformity characteristic data of the current batch and the shutdown timestamp are packaged and stored in the distributed database, and an opening signal is sent to the discharge valve controller to automatically transport the uniformly mixed powder to the next process.

[0068] In the field of powder mixing, a relative standard deviation (RSD) of ≤5% typically meets the uniformity requirements of general industrial products. Therefore, the preset uniformity value can be determined by establishing a mapping relationship between the average uniformity value and the industry-standard RSD through experimental calibration. At least 100 sets of data pairs under different mixing states are collected. Each data pair includes the average uniformity value calculated by this invention and the RSD value obtained through offline sampling and testing. Regression analysis is performed on the average uniformity value and the corresponding RSD value to establish a mapping relationship. The average uniformity value with an RSD of 5% is selected as the preset uniformity value.

[0069] In one embodiment, the method further includes: If the average uniform value is greater than the preset uniform value, it is determined that the current metal powder has not reached a uniform mixing state, and the system continues to collect acoustic emission signals for continuous monitoring.

[0070] It should be noted that if the average uniform value is greater than the preset uniform value, it is determined that the current metal powder has not reached a uniform mixing state. At this time, the system intercepts the opening command of the discharge valve to prevent unqualified materials from flowing into the next process. At the same time, it continues to collect acoustic emission signals for continuous monitoring, and can also increase the sampling frequency to more accurately capture the collision characteristics of insufficiently mixed particles.

[0071] In summary, this invention discloses a sensor-based method for monitoring the uniformity of metal powder mixing. It acquires acoustic emission signals from metal powder collisions using sensors, filters and denoises them, and then uses a Fast Fourier Transform (FFT) to simultaneously visualize the signal, obtaining a fluctuation spectrum matrix reflecting energy distribution. This matrix is ​​input into a pre-trained residual convolutional neural network to map the powder spatial distribution matrix. Subsequently, the local aggregation matrix of heavy powder distribution is extracted and its variance is calculated. After mean processing, an average uniform value is obtained. By comparing this average uniform value with a preset uniform value, the operation and shutdown of the mixer are automatically controlled. This invention achieves a complete technical chain for effectively extracting powder spatial distribution characteristics from complex acoustic emission signals, quantifying the aggregation state of heavy powder, accurately judging mixing uniformity, and controlling the mixing endpoint. It effectively solves the problem of not being able to accurately determine whether metal powder in a closed mixing cylinder is mixed uniformly based on sensor signals, improving the reliability of mixing uniformity monitoring results and providing reliable technical support for intelligent monitoring and control of metal powder metallurgy processes.

[0072] This invention also provides a sensor-based metal powder mixing uniformity monitoring system, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the sensor-based metal powder mixing uniformity monitoring method described above.

[0073] It should be noted that the sensor-based metal powder mixing uniformity monitoring system provided in this embodiment of the invention is used to execute all the process steps of the sensor-based metal powder mixing uniformity monitoring method in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0074] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0075] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A sensor-based method for monitoring the uniformity of metal powder mixing, characterized in that, include: Acoustic emission sensors are deployed at equal intervals on the outer wall of the mixer to acquire acoustic emission signals generated by the tumbling of metal powder, and noise is filtered out to obtain pure emission signals; The pure transmitted signal is subjected to spectral feature transformation processing to obtain the wave spectrum matrix; The values ​​in the fluctuation spectrum matrix are normalized to obtain a normalized two-dimensional data matrix, which is then input into a preset residual convolutional neural network for feature mapping to obtain a powder spatial distribution matrix; wherein, the preset residual convolutional neural network is pre-trained under supervision using a training sample set generated by physical simulation. The powder spatial distribution matrix is ​​uniformly divided into spatial sub-regions, and the segmentation threshold of each spatial sub-region is calculated. For each spatial sub-region, data above the corresponding segmentation threshold is extracted as representative data of heavy powder for spatial clustering analysis to obtain a local clustering matrix. The variance of the local clustering matrix is ​​calculated as the current uniform value. The average uniform value is obtained by averaging the current uniform values ​​obtained by continuous monitoring. If the average uniform value is less than or equal to the preset uniform value, it is determined that the metal powder has reached a uniform mixing state, and an endpoint command is generated to control the metal powder mixer to stop operating.

2. The sensor-based method for monitoring the uniformity of metal powder mixing according to claim 1, characterized in that, The process of noise filtering to obtain a clean transmission signal includes: The acoustic emission signal is converted to obtain a digital time-domain signal; The digital time-domain signal is filtered and cleaned to obtain the filtered signal; Calculate the signal-to-noise ratio (SNR) and peak factor (CPF) of the filtered signal. If the SNR is greater than a preset SNR and the CPF is greater than a preset CPF, then the filtered signal is determined to be a clean transmission signal.

3. The sensor-based method for monitoring the uniformity of metal powder mixing according to claim 1, characterized in that, The step of performing spectral feature transformation processing on the pure transmitted signal to obtain the fluctuation spectrum matrix includes: The pure transmitted signal is smoothed to obtain a smoothed signal; The smoothed signal is then transformed in the frequency domain to obtain a complex sequence in the frequency domain; The modulus of the frequency domain complex number sequence is extracted and squared to obtain the power spectrum sequence; Interpolation fitting is performed on the power spectrum sequence to obtain a continuous fluctuation spectrum curve; The continuous fluctuation spectrum curve is encapsulated as a fluctuation spectrum matrix.

4. The sensor-based method for monitoring the uniformity of metal powder mixing according to claim 1, characterized in that, The preset residual convolutional neural network includes an input layer, a convolutional layer, a pooling layer, four cascaded residual blocks, a flattening layer, a fully connected layer, and an output layer connected in sequence.

5. The sensor-based method for monitoring the uniformity of metal powder mixing according to claim 1, characterized in that, The input preset residual convolutional neural network performs feature mapping to obtain the powder spatial distribution matrix, including: After the normalized two-dimensional data matrix is ​​input through the input layer, the first feature map is obtained by sliding scan in the convolutional layer. The first feature map is input into the pooling layer for spatial dimensionality reduction to obtain the second feature map; The second feature map is input into four cascaded residual blocks for feature extraction to obtain a high-dimensional feature map. The high-dimensional feature map is input into the flattening layer for flattening to obtain the first feature vector; The first feature vector is input into a fully connected layer for feature recombination to obtain a second feature vector; The second feature vector output layer is subjected to regression calculation to obtain the powder spatial distribution matrix.

6. The sensor-based method for monitoring the uniformity of metal powder mixing according to claim 1, characterized in that, The calculation of the segmentation threshold for each spatial sub-region includes: Take a spatial sub-region as an example: The values ​​in this spatial sub-region are used as the candidate threshold set; Choose any candidate threshold from the candidate threshold set. If the value in the spatial sub-region is less than or equal to the candidate threshold, the value is assigned to the low-density group. If the value in the spatial sub-region is greater than the candidate threshold, the value is assigned to the high-density group. Calculate the inter-class variance of the candidate threshold after classifying the values ​​into the high-density group and the low-density group; Iterate through all candidate thresholds, repeat the above steps of grouping and calculating inter-class variance, and obtain the inter-class variance for each candidate threshold. The candidate threshold with the largest inter-class variance is selected as the segmentation threshold for the spatial sub-region.

7. The sensor-based method for monitoring the uniformity of metal powder mixing according to claim 1, characterized in that, For each spatial sub-region, data above the corresponding segmentation threshold are extracted as representative data of heavy powder for spatial clustering analysis to obtain a local clustering matrix, including: Within each of the spatial sub-regions, data exceeding the segmentation threshold are extracted and weighted summed to obtain the weighted sum value of that sub-region. The weighted sum of all spatial sub-regions is combined and output as a local clustering matrix.

8. The sensor-based method for monitoring the uniformity of metal powder mixing according to claim 1, characterized in that, The calculation of the variance of the local clustering matrix as the current uniform value includes: Calculate the arithmetic mean of all elements in the local clustering matrix to obtain the global mean; Calculate the deviation of each element in the local clustering matrix from the global mean, and square each deviation to obtain the squared deviation of each element; Sum the squared deviations of all elements to obtain the total sum of squared deviations; The current uniform value is obtained by dividing the sum of squared total deviations by the total number of elements in the local clustering matrix.

9. The sensor-based method for monitoring the uniformity of metal powder mixing according to claim 1, characterized in that, The method further includes: If the average uniform value is greater than the preset uniform value, it is determined that the current metal powder has not reached a uniform mixing state, and the system continues to collect acoustic emission signals for continuous monitoring.

10. A sensor-based metal powder mixing uniformity monitoring system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the sensor-based metal powder mixing uniformity monitoring method according to any one of claims 1 to 9.

Images