Amorphous alloy powder cold pressing flow intelligent sensing method and system

Abstract

This invention relates to an intelligent flow sensing method and system for cold-pressing amorphous alloy powder, belonging to the field of flow measurement technology. The method includes constructing an acoustic emission feature vector; inputting the original capacitance measurement tensor and acoustic emission feature vector into a capacitance tomography reconstruction algorithm to construct a multi-frequency dielectric-mechanical coupled Gaussian mixture model; performing soft classification decoupling on the dielectric constant distribution matrix to generate a shear band density field and a porosity distribution field; introducing the shear band density field to perform free volume correction on the powder's apparent density distribution field, and solving to obtain a real-time flow characteristic scalar; inputting this time-series graph structure data into a graph neural network-long short-term memory coupled model; comparing the shear band network connectivity with a threshold in real time; and constructing a multi-objective cost function to achieve coordinated control of high-precision flow tracking and active suppression of shear band instability during cold pressing. This invention improves the flow sensing accuracy, instability early warning timeliness, and process control coordination of cold-pressing amorphous alloy powder.

Description

Technical Field

[0001] This invention relates to the field of flow measurement technology, and more specifically, to an intelligent sensing method and system for the flow rate of cold-pressed amorphous alloy powder. Background Technology

[0002] Cold pressing of amorphous alloy powder is a key process for preparing bulk amorphous alloy components, and its forming quality is highly dependent on the precise control of powder filling flow rate and the real-time monitoring of shear band instability. Existing technologies mainly use weighing flow sensors or single-point capacitance sensors to measure powder flow rate, judge abnormal states in the pressing process through empirical thresholds, and rely on manual adjustment of powder supply rate and pressing pressure to achieve process control.

[0003] However, weighing sensors suffer from response lag, making it difficult to capture the dynamic transient characteristics of powder filling during cold pressing; single-point capacitive sensors lack spatial resolution, failing to distinguish powder density distribution, shear band initiation regions, and pore aggregation locations; empirical threshold methods lack physical understanding of the evolution mechanism of the shear transition zone in amorphous alloys, leading to delayed early warning of shear band network instability; and manual adjustment control has low precision, making it difficult to achieve synergistic optimization of flow tracking and instability suppression. Therefore, existing technologies have significant limitations in measurement accuracy, spatial resolution, instability early warning, and control synergy, failing to meet the high-quality, high-efficiency, and high-reliability process requirements of amorphous alloy powder cold pressing. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent sensing method and system for the cold pressing flow rate of amorphous alloy powder, so as to improve the above-mentioned problems. To achieve the above objective, the technical solution adopted by this invention is as follows:

[0005] In a first aspect, this application provides a method for intelligent sensing of the cold pressing flow rate of amorphous alloy powder, including:

[0006] A miniature capacitance sensor array and a broadband acoustic emission sensor array are embedded in the cavity wall of a cold-pressing mold to collect capacitance tomography signals and acoustic emission signals of amorphous alloy powder during the cold pressing process. Then, the original capacitance measurement tensor and acoustic emission feature vector are constructed. Wavelet packet decomposition is performed on the acoustic emission signal to extract the shear band initiation energy features and construct the acoustic emission feature vector. The original capacitance measurement tensor is composed of a three-dimensional stack of multi-frequency capacitance measurement value sequences according to electrode number, frequency number, and time series points.

[0007] The original capacitance measurement tensor and acoustic emission feature vector are input into the capacitance tomography reconstruction algorithm to perform adaptive iterative reconstruction based on acoustic emission features, outputting a spatiotemporally continuous dielectric constant distribution matrix. The prior dielectric dispersion characteristics obtained from offline calibration experiments are called to construct a multi-frequency dielectric-mechanical coupled Gaussian mixture model. The equiprobability surface of the equivalent dielectric constant probability distribution of each phase output by the model is used as the decision boundary. The fixed dielectric threshold is replaced with a frequency-energy related nonlinear dielectric classification surface. Soft classification decoupling is performed on the dielectric constant distribution matrix to generate a shear band density field and a porosity distribution field.

[0008] The dielectric constant-density mapping relationship established by offline calibration experiments is used to convert the dielectric constant distribution matrix after soft classification decoupling into a powder apparent density distribution field. A shear band density field is introduced to correct the free volume of this apparent density distribution field, generating a free volume corrected density field. The effective flow area field is calculated based on the porosity distribution field, and a Darcy flow model is established by combining the free volume corrected density field to obtain the real-time flow characteristic scalar. Simultaneously, fractal dimension analysis is performed on the shear band density field to extract the shear band network connectivity. The real-time flow characteristic scalar, the shear band network fractal dimension, the shear band network connectivity, and the acoustic emission energy change rate in the acoustic emission characteristic vector are combined to construct a powder flow state characteristic vector.

[0009] The powder flow state feature vector is expanded into a time-series graph structure, with local state variables of each sensor sub-region as node features and spatial adjacency relationships of shear zone regions in the shear zone density field as edge features. This time-series graph structure is input into a graph neural network-long short-term memory coupled model. The graph neural network extracts spatial topological correlation features of the shear zone through message passing and outputs a spatial embedding feature vector. The long short-term memory network performs time-series dynamic modeling on this spatial embedding feature vector and outputs a time-series prediction feature vector. After decoding, a flow prediction sequence and instability risk probability are generated. The percolation critical connectivity threshold determined by offline calibration experiments is called, and the shear zone network connectivity is compared with the threshold in real time. If the threshold is exceeded, an instability warning signal is generated; if it is not exceeded, the normal state is maintained.

[0010] Using the flow prediction sequence, instability risk probability, and instability early warning signal as inputs, a multi-objective cost function is constructed. This multi-objective cost function includes a flow tracking accuracy term, an instability risk suppression term, a pressing energy consumption term, and an emergency safety protection term. A rolling time-domain optimization solution is performed on the multi-objective cost function to obtain the optimal control sequence, which in turn generates the powder supply rate adjustment, pressing pressure adjustment, and emergency pressure relief valve opening command. The powder supply rate adjustment is output to the powder supply driver, the pressing pressure adjustment is output to the servo pressing unit, and the emergency pressure relief valve opening command is output to the safety pressure relief valve, achieving coordinated control of high-precision flow tracking and active suppression of shear band instability during cold pressing.

[0011] Preferably, a miniature capacitance sensor array and a broadband acoustic emission sensor array are embedded in the cavity wall of the cold-pressing mold to collect capacitance tomography signals and acoustic emission signals of amorphous alloy powder during the cold-pressing process. This allows for the construction of the original capacitance measurement tensor and acoustic emission feature vector. Wavelet packet decomposition is then performed on the acoustic emission signal to extract shear band initiation energy features, and the acoustic emission feature vector is constructed, including:

[0012] A ring-shaped micro-capacitive sensor array is embedded in the cavity wall of a cold-pressing mold. The array consists of excitation electrodes and measurement electrodes arranged alternately. A multi-frequency excitation mode is adopted, with low frequency used to penetrate the powder gap and high frequency used to distinguish the microstructure of the shear band. The acquisition frequency is not less than 10,000 frames per second to obtain a sequence of multi-frequency capacitance measurement values ​​on the cross section.

[0013] The multi-frequency capacitance measurement value sequence is stacked in three dimensions according to electrode number, frequency number and time series points to construct the original capacitance measurement tensor.

[0014] A broadband acoustic emission sensor array was arranged in the shear stress concentration region of the cavity wall. The center frequency of the sensor was matched with the characteristic frequency of the shear band of the amorphous alloy, and the sampling rate was set to 5 times the highest frequency of the acoustic emission signal. Acoustic emission signal sequences of friction, collision and shear band initiation between amorphous alloy powder particles were collected. Wavelet packet decomposition was performed on the acoustic emission signal sequence with 4 decomposition layers to obtain wavelet coefficients of 16 frequency bands. Wavelet energy of the characteristic frequency range of shear band from 300 kHz to 500 kHz corresponding to the 9th to 12th frequency bands was extracted to construct the shear band initiation energy feature. The shear band initiation energy features of each sensor node were arranged according to spatial position to construct the acoustic emission feature vector.

[0015] Preferably, the original capacitance measurement tensor and acoustic emission feature vector are input into the capacitance tomography reconstruction algorithm to perform adaptive iterative reconstruction based on acoustic emission features, outputting a spatiotemporally continuous dielectric constant distribution matrix; prior dielectric dispersion characteristics obtained from offline calibration experiments are called to construct a multi-frequency dielectric-mechanical coupled Gaussian mixture model; using the equiprobability surface of the equivalent dielectric constant probability distribution of each phase output by the model as the decision boundary, the fixed dielectric threshold is replaced with a frequency-energy related nonlinear dielectric classification surface, including:

[0016] The original capacitance measurement tensor and acoustic emission feature vector are input into the capacitance tomography reconstruction algorithm to construct an adaptive iterative reconstruction operator based on acoustic emission features. The iteration step size is dynamically adjusted according to the shear band initiation energy.

[0017] Perform iterative reconstruction until convergence, and output a spatiotemporally continuous dielectric constant distribution matrix; call the prior dielectric dispersion characteristics obtained from the offline calibration experiment, which measures the complex dielectric constant versus frequency curves of undeformed amorphous alloy powder, pre-compression induced shear band powder, and residual gas in the mold under standard temperature and humidity conditions; extract the static dielectric constant, optical frequency dielectric constant, and dielectric relaxation time of each phase from the complex dielectric constant versus frequency curves. These three parameters are the characteristic parameters of the Debye relaxation model and together constitute the prior dielectric dispersion characteristics;

[0018] Based on the prior dielectric dispersion characteristics, a conditional probability distribution of the dielectric constant of each phase state with respect to frequency is established. Using the conditional probability distribution of each phase state as components and the phase volume fraction as the mixing coefficient, a multi-frequency dielectric-mechanical coupled Gaussian mixture model is constructed. The input of this model is frequency and acoustic emission energy, and the output is the probability distribution of the equivalent dielectric constant of each phase state. Using the equiprobability surface of the probability distribution of the equivalent dielectric constant of each phase state output by this model as the decision boundary, the fixed dielectric threshold is replaced with a frequency-energy related nonlinear dielectric classification surface.

[0019] Preferably, the soft classification decoupling of the dielectric constant distribution matrix to generate a shear band density field and a porosity distribution field includes:

[0020] Soft classification decoupling is performed on the dielectric constant distribution matrix, the posterior probability of each spatial location belonging to each phase state is calculated, and a three-phase probability density field is generated.

[0021] Shear band phase probability components are extracted from the three-phase probability density field. After Gaussian kernel spatial smoothing and adaptive threshold binarization, isolated regions with areas smaller than the minimum cluster area are removed, and connected component labeling is performed. The shear band phase probability values ​​of each connected component are averaged by area to generate the shear band density field. At the same time, the gas phase probability components are extracted and processed in the same way to serve as the porosity distribution field.

[0022] Preferably, the step of calling the dielectric constant-density mapping relationship established by the offline calibration experiment converts the dielectric constant distribution matrix after soft classification decoupling into a powder apparent density distribution field; the shear band density field is introduced to perform free volume correction on the powder apparent density distribution field to generate a free volume corrected density field, including:

[0023] The dielectric constant-density mapping relationship established by the offline calibration experiment is invoked. This offline calibration experiment uses a precision balance to measure the ratio of powder mass to capacitance chromatography reconstruction volume under different compaction states to establish a quadratic polynomial mapping between dielectric constant and powder apparent density. The dielectric constant distribution matrix is ​​then substituted into this mapping relationship to convert it into the powder apparent density distribution field.

[0024] A shear band density field is introduced to correct the apparent density distribution field of the powder for free volume. Considering the volume expansion effect in the shear transition region, the shear band region is regarded as a loose structure containing free volume, thus generating a free volume corrected density field.

[0025] Preferably, the coefficient of the volume expansion effect in the shear transition zone ranges from 0.02 to 0.05, and the minimum cluster area is determined by the ratio of the average particle size of the powder particles to the imaging resolution.

[0026] Preferably, the effective flow area field is calculated based on the porosity distribution field, and a Darcy flow model is established by combining the free volume corrected density field to obtain the real-time flow characteristic scalar; simultaneously, fractal dimension analysis is performed on the shear band density field to extract the shear band network connectivity; the real-time flow characteristic scalar, the shear band network fractal dimension, the shear band network connectivity, and the acoustic emission energy change rate in the acoustic emission characteristic vector are combined to construct a powder flow state characteristic vector, including:

[0027] The effective flow area is calculated based on the porosity distribution field. The cross-sectional area of ​​the cavity is multiplied by the local porosity complement to obtain the effective flow area distribution, generating the effective flow area field. A Darcy flow model is established by combining the free volume corrected density field. The nonlinear resistance characteristics of powder deformation are considered. The permeability adopts the Kozeny-Carman correction form and introduces the shear band tortuosity correction factor. A tortuosity corrected permeability field is generated. The local velocity field is calculated by the pressure gradient, the tortuosity corrected permeability field and the effective viscosity. The product of the local velocity field and the free volume corrected density field is integraled on the cross-section of the cavity to solve for the real-time flow characteristic scalar.

[0028] Fractal dimension analysis was performed on the shear band density field. The fractal dimension of the shear band network was calculated using the box counting method. The box size sequence was arranged in a geometric progression. The fractal dimension was determined by the slope of the linear fitting of the double logarithmic coordinates. The shear band network connectivity was extracted and defined as the ratio of the area of ​​the largest connected cluster to the area of ​​the total shear band region. The shear band network connectivity was then generated. The real-time flow characteristic scalar, the fractal dimension of the shear band network, the shear band network connectivity, and the acoustic emission energy change rate were combined to construct a powder flow state feature vector.

[0029] Preferably, the powder flow state feature vector is expanded into a time-series graph structure data, with the local state variables of each sensor sub-region as node features and the spatial adjacency relationship of the shear band region in the shear band density field as edge features; this time-series graph structure data is input into a graph neural network-long short-term memory coupled model; the graph neural network extracts the spatial topological correlation features of the shear band through message passing and outputs a spatial embedding feature vector; the long short-term memory network performs time-series dynamic modeling on this spatial embedding feature vector and outputs a time-series prediction feature vector; after decoding, a flow prediction sequence and instability risk probability are generated, including:

[0030] The powder flow state feature vector is spatially divided according to the sensor sub-regions. Each sub-region corresponds to a node in the graph. The node feature vector consists of the real-time flow feature scalar of the sub-region, the fractal dimension scalar of the shear band network, the connectivity scalar of the shear band network, and the acoustic emission energy change rate, generating a node feature matrix. The edge set of the graph is determined by the spatial adjacency relationship of the shear band regions in the shear band density field. When the spatial distance between the corresponding sub-regions of two nodes in the shear band density field is less than the connectivity threshold, an edge connection is established. The edge weight is proportional to the spatial correlation coefficient of the shear band density between adjacent nodes, generating a weighted adjacency matrix. Using the current time as the endpoint, historical data of a fixed time length is extracted to construct a sliding window time series graph sequence with a time span, generating time series graph sequence data.

[0031] The time-series graph sequence data is input into the graph neural network-long short-term memory coupled model. The graph neural network takes the node feature matrix and weighted adjacency matrix as input, calculates the attention coefficient between nodes through the graph attention mechanism, aggregates neighbor node information through message passing, extracts the spatial topological correlation features of the shear band, and outputs a sequence of spatially embedded feature vectors. The long short-term memory network takes the sequence of spatially embedded feature vectors as input, captures the temporal dependencies through the gating mechanism, models the flow evolution trend and shear band network dynamics, and outputs a temporal prediction feature vector.

[0032] The time-series prediction feature vector is input into the fully connected decoding layer, and after linear transformation and activation function mapping, a traffic prediction sequence and instability risk probability for multiple future time steps are generated.

[0033] Preferably, the rolling time-domain optimization solution of the multi-objective cost function is performed to obtain the optimal control sequence, and then the powder supply rate adjustment, pressing pressure adjustment, and emergency pressure relief valve opening commands are generated, including:

[0034] A rolling time-domain optimization solution is performed on the multi-objective cost function, with the prediction time domain taking 100 to 200 time steps and the control time domain taking half of the prediction time domain. The constraints include: the upper limit of the pressing pressure is the product of the yield strength of the amorphous alloy and the correction value of the average porosity; the powder supply rate is limited to between zero and the maximum value; and the emergency pressure relief response time does not exceed 50 milliseconds. A sequential quadratic programming algorithm is used to solve the problem, obtaining the optimal control sequence, and generating the optimal powder supply rate sequence, the optimal pressing pressure sequence, and the optimal emergency pressure sequence.

[0035] The control commands for the current moment are generated from the optimal control sequence: the powder supply rate adjustment is the difference between the optimal supply rate and the actual supply rate at the previous moment; the pressing pressure adjustment is the difference between the optimal pressing pressure and the actual pressing pressure at the previous moment; and the emergency pressure relief valve opening command is the ratio of the optimal emergency pressure to the maximum allowable pressure converted into a percentage. The powder supply rate adjustment control command, the pressing pressure adjustment control command, and the emergency pressure relief valve opening command are generated.

[0036] Secondly, this application also provides an intelligent sensing system for the cold pressing flow rate of amorphous alloy powder, comprising:

[0037] The construction module is used to embed a miniature capacitance sensor array and a broadband acoustic emission sensor array on the cavity wall of a cold-pressing mold. It collects capacitance tomography signals and acoustic emission signals of amorphous alloy powder during the cold pressing process, and then constructs the original capacitance measurement tensor and acoustic emission feature vector. Wavelet packet decomposition is performed on the acoustic emission signal to extract the shear band initiation energy features and construct the acoustic emission feature vector. The original capacitance measurement tensor is composed of a three-dimensional stack of multi-frequency capacitance measurement value sequences according to electrode number, frequency number, and time series points.

[0038] The decoupling generation module is used to input the original capacitance measurement tensor and acoustic emission feature vector into the capacitance tomography reconstruction algorithm, perform adaptive iterative reconstruction based on acoustic emission features, and output a spatiotemporally continuous dielectric constant distribution matrix; it calls the prior dielectric dispersion characteristics obtained from offline calibration experiments to construct a multi-frequency dielectric-mechanical coupled Gaussian mixture model; using the equiprobability surface of the equivalent dielectric constant probability distribution of each phase output by the model as the decision boundary, it replaces the fixed dielectric threshold with a frequency-energy related nonlinear dielectric classification surface, performs soft classification decoupling on the dielectric constant distribution matrix, and generates a shear band density field and porosity distribution field;

[0039] Extraction and Combination Module: This module is used to call the dielectric constant-density mapping relationship established by the offline calibration experiment, convert the dielectric constant distribution matrix after soft classification decoupling into a powder apparent density distribution field; introduce the shear band density field to perform free volume correction on the powder apparent density distribution field, generating a free volume corrected density field; calculate the effective flow area field based on the porosity distribution field, and establish a Darcy flow model in combination with the free volume corrected density field to obtain the real-time flow characteristic scalar; simultaneously perform fractal dimension analysis on the shear band density field to extract the shear band network connectivity; combine the real-time flow characteristic scalar, the shear band network fractal dimension, the shear band network connectivity, and the acoustic emission energy change rate in the acoustic emission characteristic vector to construct a powder flow state characteristic vector;

[0040] The judgment module expands the powder flow state feature vector into a time-series graph structure, using the local state variables of each sensor sub-region as node features and the spatial adjacency relationships of the shear zone region in the shear zone density field as edge features. This time-series graph structure is input into a graph neural network-long short-term memory coupled model. The graph neural network extracts the spatial topological correlation features of the shear zone through message passing and outputs a spatial embedding feature vector. The long short-term memory network performs time-series dynamic modeling on this spatial embedding feature vector and outputs a time-series prediction feature vector. After decoding, a flow prediction sequence and instability risk probability are generated. The percolation critical connectivity threshold determined by offline calibration experiments is called, and the shear zone network connectivity is compared with this threshold in real time. If the threshold is exceeded, an instability warning signal is generated; otherwise, the normal state is maintained.

[0041] The perception and control module is used to construct a multi-objective cost function based on the flow prediction sequence, instability risk probability, and instability early warning signal. This multi-objective cost function includes a flow tracking accuracy term, an instability risk suppression term, a pressing energy consumption term, and an emergency safety protection term. A rolling time-domain optimization solution is performed on the multi-objective cost function to obtain the optimal control sequence, which in turn generates the powder supply rate adjustment, pressing pressure adjustment, and emergency pressure relief valve opening command. The powder supply rate adjustment is output to the powder supply driver, the pressing pressure adjustment is output to the servo pressing unit, and the emergency pressure relief valve opening command is output to the safety pressure relief valve, achieving coordinated control of high-precision flow tracking and active suppression of shear band instability during cold pressing.

[0042] Thirdly, this application also provides an intelligent sensing device for the cold pressing flow rate of amorphous alloy powder, comprising:

[0043] Memory, used to store computer programs;

[0044] A processor is used to implement the steps of the intelligent sensing method for cold-pressed flow rate of amorphous alloy powder when executing the computer program.

[0045] Fourthly, this application also provides a readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described intelligent sensing method for cold-pressed flow rate of amorphous alloy powder.

[0046] The beneficial effects of this invention are as follows:

[0047] This invention constructs an original capacitance measurement tensor and acoustic emission feature vector to achieve multimodal fusion sensing of capacitance tomography signals and acoustic emission signals, achieving high spatiotemporal resolution monitoring of the powder filling process. It solves the problems of fixed dielectric thresholds and low phase decoupling accuracy in existing technologies, achieving accurate identification of shear band and pore regions. Furthermore, by calling the dielectric constant-density mapping relationship established through offline calibration experiments, it solves the real-time flow characteristic scalar, addressing the problems of not considering the volume expansion effect of the shear transition zone and the lack of coupling of flow calculation with pore seepage in existing technologies, achieving accurate characterization of effective powder flow and quantification of shear band network topology. It extracts spatial topological correlation features of the shear band through a graph neural network to generate instability warning signals or normal state indicators, solving the problems of lack of physical basis for empirical thresholds and delayed instability warnings in existing technologies. Finally, it constructs a multi-objective cost function, outputting it to the powder supply driver, servo pressing unit, and safety pressure relief valve, achieving coordinated control of high-precision flow tracking and active suppression of shear band instability during cold pressing, solving the problems of low precision in manual adjustment and the inability to coordinate flow tracking and instability suppression in existing technologies, achieving automated and intelligent control of the cold pressing process.

[0048] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0049] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is a schematic diagram of the intelligent sensing method for cold pressing flow of amorphous alloy powder as described in this embodiment of the invention;

[0051] Figure 2 This is a schematic diagram of the intelligent sensing system for cold pressing flow of amorphous alloy powder as described in an embodiment of the present invention;

[0052] Figure 3 This is a schematic diagram of the intelligent sensing device for cold-pressed flow rate of amorphous alloy powder as described in an embodiment of the present invention.

[0053] In the diagram: 701, Construction Module; 702, Decoupling Generation Module; 703, Extraction and Combination Module; 704, Judgment Module; 705, Sensing and Control Module; 800, Intelligent Sensing Device for Cold Pressing Flow of Amorphous Alloy Powder; 801, Processor; 802, Memory; 803, Multimedia Component; 804, I / O Interface; 805, Communication Component. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0055] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0056] Example 1:

[0057] This embodiment provides a method for intelligent sensing of the flow rate of cold-pressed amorphous alloy powder.

[0058] See Figure 1 The figure shows that the method includes steps S100, S200, S300, S400 and S500.

[0059] S100: A miniature capacitance sensor array and a broadband acoustic emission sensor array are embedded in the cavity wall of the cold pressing mold to collect capacitance tomography signals and acoustic emission signals of amorphous alloy powder during the cold pressing process. Then, the original capacitance measurement tensor and acoustic emission feature vector are constructed. Wavelet packet decomposition is performed on the acoustic emission signal to extract the shear band initiation energy features and construct the acoustic emission feature vector. The original capacitance measurement tensor is composed of a three-dimensional stack of multi-frequency capacitance measurement value sequences according to electrode number, frequency number and time series points.

[0060] It is understood that step S100 includes S101, S102, and S103, wherein:

[0061] S101. An array of ring-shaped micro-capacitive sensors is embedded in the wall of the cold-pressing mold cavity. The array consists of excitation electrodes and measurement electrodes arranged alternately. A multi-frequency excitation mode is adopted, with low frequency used to penetrate the powder gap and high frequency used to distinguish the microstructure of the shear band. The acquisition frequency is not less than 10,000 frames per second to obtain a sequence of multi-frequency capacitance measurement values ​​on the cross section.

[0062] It should be noted that a multi-frequency excitation mode is used in this step, with the excitation frequency range set from 100 kHz to 10 MHz. The low-frequency band of 100 kHz to 1 MHz is used to penetrate the powder gaps, and the high-frequency band of 1 MHz to 10 MHz is used to distinguish the microstructure of the shear band.

[0063] S102. Stack the multi-frequency capacitance measurement value sequence in three dimensions according to electrode number, frequency number, and time series points to construct the original capacitance measurement tensor. The calculation formula is as follows:

[0064]

[0065] In the formula, for Time of the first At the excitation frequency, the first The excitation electrode and the first The capacitance measurement value between the measuring electrodes This represents the induced charge between the corresponding electrode pairs. for Time of the first The excitation electrode in the first... Excitation voltage at a given frequency Number the excitation electrodes. Here, k is the measurement electrode number, k is the frequency number, and t is the discrete time sequence number;

[0066] S103. A broadband acoustic emission sensor array is arranged in the shear stress concentration area of ​​the cavity wall. The center frequency of the sensor is matched with the characteristic frequency of the shear band of the amorphous alloy. The sampling rate is set to 5 times the highest frequency of the acoustic emission signal, i.e., not less than 2.5 MHz. Acoustic emission signal sequences of friction, collision and shear band initiation between amorphous alloy powder particles are collected. Wavelet packet decomposition is performed on the acoustic emission signal sequence. The decomposition level is 4 layers to obtain wavelet coefficients of 16 frequency bands. Wavelet energy of the characteristic frequency range of shear band from 300 kHz to 500 kHz corresponding to the 9th to 12th frequency bands is extracted to construct the shear band initiation energy feature. The shear band initiation energy features of each sensor node are arranged according to spatial position to construct the acoustic emission feature vector.

[0067] It should be noted that the formula for calculating the energy characteristic of shear band initiation is as follows:

[0068]

[0069] In the formula, The energy characteristics of shear band initiation at time t. Let be the acoustic emission signal at time t′, and WT[·] be the continuous wavelet transform. , For the first The upper and lower boundary frequencies of each frequency band For the first The frequency band weights are determined by the energy concentration of the shear band. For frequency band numbering, It is a frequency integral infinitesimal element.

[0070] Specifically, a spatiotemporal registration mechanism for capacitance tomography signals and acoustic emission signals is established. The geometric center of the capacitance sensor array is used as the spatial reference origin, and the coordinates of the acoustic emission sensor are mapped to the capacitance measurement coordinate system. The registration transformation matrix is ​​determined by least squares calibration. The acoustic emission signal is resampled and aligned using the capacitance acquisition trigger signal as the time reference to eliminate the clock deviation between the two types of signals, with a time synchronization accuracy of not less than 10 microseconds. The original capacitance measurement tensor and the acoustic emission feature vector after registration are timestamped to construct a multimodal sensing data frame.

[0071] The multimodal sensing data frame contains the original capacitance measurement tensor and acoustic emission feature vector. Step S200 directly calls these two components from the data frame as input to perform capacitance tomography reconstruction and phase-field decoupling. That is, the output of step one, the "multimodal sensing data frame," is the source of input data for step S200, from which step S200 extracts the original capacitance measurement tensor and acoustic emission feature vector for use.

[0072] S200: Input the original capacitance measurement tensor and acoustic emission feature vector into the capacitance tomography reconstruction algorithm, perform adaptive iterative reconstruction based on acoustic emission features, and output a spatiotemporally continuous dielectric constant distribution matrix; call the prior dielectric dispersion characteristics obtained from the offline calibration experiment to construct a multi-frequency dielectric-mechanical coupled Gaussian mixture model; use the equiprobability surface of the equivalent dielectric constant probability distribution of each phase output by the model as the decision boundary, replace the fixed dielectric threshold with a frequency-energy related nonlinear dielectric classification surface, perform soft classification decoupling on the dielectric constant distribution matrix, and generate a shear band density field and porosity distribution field.

[0073] It is understood that step S200 includes S201, S202, and S203, wherein:

[0074] S201. Input the original capacitance measurement tensor and acoustic emission feature vector into the capacitance tomography reconstruction algorithm to construct an adaptive iterative reconstruction operator based on acoustic emission features. The iteration step size is dynamically adjusted according to the shear band initiation energy, and its calculation formula is as follows:

[0075]

[0076] In the formula, Let t be the adaptive iteration step size. As the reference step size, The acoustic emission coupling coefficient is... The energy characteristics of shear band initiation at time t. The average acoustic emission energy. The standard deviation of acoustic emission energy is used; when the shear band initiation energy is higher than the mean, the iteration step size is reduced to avoid overfitting, and when it is lower than the mean, the step size is increased to accelerate convergence.

[0077] S202. Perform iterative reconstruction until convergence, and output a spatiotemporally continuous dielectric constant distribution matrix; call the prior dielectric dispersion characteristics obtained from the offline calibration experiment. The offline calibration experiment measures the complex dielectric constant versus frequency curves of undeformed amorphous alloy powder, pre-compression induced shear band powder, and residual gas in the mold under standard temperature and humidity conditions; extract the static dielectric constant, optical frequency dielectric constant, and dielectric relaxation time of each phase from the complex dielectric constant versus frequency curves. These three parameters are the characteristic parameters of the Debye relaxation model and together constitute the prior dielectric dispersion characteristics.

[0078] It should be noted that broadband dielectric spectroscopy measurements were performed on the three types of standard samples, with the frequency scan range covering the operating frequency band of the capacitive sensor array, typically from the kilohertz to megahertz level. The measured complex dielectric constant versus frequency curves for each sample exhibit typical Debye relaxation characteristics: the dielectric constant tends to the static dielectric constant in the low-frequency range, the dielectric constant tends to the optical frequency dielectric constant in the high-frequency range, and dispersion characteristics are observed in the mid-frequency range. The Debye relaxation model is a classic theory describing the polarization relaxation of polar molecules or interfaces.

[0079] S203. Based on the prior dielectric dispersion characteristics, establish the conditional probability distribution of the dielectric constant of each phase state with respect to frequency. Using the conditional probability distribution of each phase state as components and the phase volume fraction as the mixing coefficient, construct a multi-frequency dielectric-mechanical coupled Gaussian mixture model. The input of this model is frequency and acoustic emission energy, and the output is the probability distribution of the equivalent dielectric constant of each phase state. Using the equiprobability surface of the probability distribution of the equivalent dielectric constant of each phase state output by this model as the decision boundary, replace the fixed dielectric threshold with a frequency-energy related nonlinear dielectric classification surface.

[0080] It should be noted that, based on the prior dielectric dispersion characteristics obtained from S202, a conditional probability distribution of the dielectric constant of each phase state with respect to frequency is established. Specifically, for a given frequency, the dielectric constant of each phase state follows a Gaussian distribution with the mean of the Debye model predictions and the standard deviation of the measurement uncertainty. This conditional probability distribution integrates the dielectric physics model and measurement statistics, making it more robust than simply using a deterministic model.

[0081] A multi-frequency dielectric-mechanical coupled Gaussian mixture model is constructed, using the conditional probability distribution of each phase state as components and the phase volume fraction as mixing coefficients. The Gaussian mixture model is a classic probabilistic model describing the distribution of multimodal data, and its mixing coefficients characterize the volume proportion of each phase state in the measurement voxel. The model inputs are the excitation frequency and acoustic emission energy. The introduction of acoustic emission energy couples the mechanical state with the dielectric response: higher shear band initiation energy indicates a more active shear transition region, resulting in a corresponding increase in the mixing coefficient of the shear band phase and a shift in the model output towards the dielectric characteristics of the shear band phase. The classification surface exhibits a curved shape in the dielectric constant-frequency-energy three-dimensional space: in the high-frequency, high-energy region, the dielectric difference between the shear band phase and the undeformed powder phase is amplified, and the classification surface tends to be steep; in the low-frequency, low-energy region, the dielectric characteristics of the two phases are similar, and the classification surface tends to be gentle. This nonlinear classification strategy fully considers the dielectric-mechanical coupling characteristics during the cold pressing process of amorphous alloy powder, significantly improving the identification accuracy between shear band regions and undeformed powder regions, and laying the foundation for the accurate extraction of the shear band density field in the future.

[0082] Furthermore, step S200 also includes:

[0083] Soft classification decoupling is performed on the dielectric constant distribution matrix, and the posterior probability of each spatial location belonging to each phase state is calculated to generate a three-phase probability density field. The calculation formula is as follows:

[0084]

[0085] In the formula, Let be the posterior probability that the position (x, y) at time t belongs to the m-th phase. Let be the element of the dielectric constant distribution matrix at position (x, y) at time t. Let be the mixing coefficient of the m-th phase. For the m-th element, the excitation frequency f and the shear band volume fraction are relevant. The mean function, It follows a Gaussian distribution. Let M be the variance of the m-th phase, M be the total number of phase states, and m be the phase state number. For the summation index, Let f be the spatial coordinates and f be the excitation frequency. This represents the volume fraction of the shear band. For the first The mixing coefficient of the phase, For the first Related to excitation frequency f and shear band volume fraction The mean function, For the first The variance of the phase;

[0086] Shear band phase probability components are extracted from the three-phase probability density field. After Gaussian kernel spatial smoothing and adaptive threshold binarization, isolated regions with areas smaller than the minimum cluster area are removed, and connected component labeling is performed. The shear band phase probability values ​​of each connected component are averaged by area to generate the shear band density field. At the same time, the gas phase probability components are extracted and processed in the same way to serve as the porosity distribution field. The shear band probability density field is a dimensionless probability value with a value range of [0,1].

[0087] It should be noted that the shear band phase probability component is extracted from the three-phase probability density field. This component represents the posterior probability value of each spatial location belonging to the shear band phase, reflecting the likelihood of shear transition zone formation. Gaussian kernel spatial smoothing is performed, with the kernel window size matched to the spatial resolution of the capacitive sensor, suppressing probability fluctuations caused by reconstruction noise while preserving the true spatial structure of the shear band network. Adaptive threshold binarization is employed, with the threshold being an adaptive proportion of the average shear band phase probability to accommodate differences in shear band development at different cold-pressing stages. Then, isolated regions with areas smaller than the minimum cluster area, determined by the ratio of the average powder particle size to the imaging resolution, are removed to eliminate single-particle pseudo-shear band noise. Eight-connected domain labeling is performed, and the shear band phase probability values ​​of each connected domain are averaged by area to generate a shear band density field. This density field characterizes the spatial distribution and development intensity of the shear band network.

[0088] The gas phase probability component is extracted synchronously and subjected to the same Gaussian kernel spatial smoothing, adaptive threshold binarization, connected component labeling, and small region removal processing to serve as the porosity distribution field, reflecting the gas interstitial distribution in the powder pack. To address the spatial overlap between the shear band density field and the porosity distribution field, a priority rule is set, prioritizing the shear band phase over the gas phase. Pixels in overlapping regions are assigned to the shear band phase, ensuring the physical consistency of phase-field decoupling. This step, through a probability field post-processing chain, achieves high-precision spatial quantization of the shear band network and porosity distribution, providing a reliable phase field input for subsequent free volume correction and flow rate calculation.

[0089] S300: The dielectric constant-density mapping relationship established through offline calibration experiments is invoked to convert the dielectric constant distribution matrix after soft classification decoupling into a powder apparent density distribution field. A shear band density field is introduced to perform free volume correction on this powder apparent density distribution field, generating a free volume corrected density field. Based on the porosity distribution field, the effective flow area field is calculated. A Darcy flow model is established in conjunction with the free volume corrected density field, and the real-time flow characteristic scalar is obtained. Simultaneously, fractal dimension analysis is performed on the shear band density field to extract the shear band network connectivity. The real-time flow characteristic scalar, the shear band network fractal dimension, the shear band network connectivity, and the acoustic emission energy change rate in the acoustic emission characteristic vector are combined to construct a powder flow state characteristic vector.

[0090] It is understood that step S300 includes S301 and S302, wherein:

[0091] S301. Call the dielectric constant-density mapping relationship established by the offline calibration experiment. The offline calibration experiment uses a precision balance to measure the ratio of powder mass to capacitance chromatography reconstruction volume under different compaction states to establish a quadratic polynomial mapping between dielectric constant and powder apparent density. Substitute the dielectric constant distribution matrix into the mapping relationship to convert it into the powder apparent density distribution field.

[0092] It should be noted that this offline calibration experiment prepared a series of amorphous alloy powder standard samples with different compaction degrees. The compaction degree was achieved by controlling the pressing pressure and holding time, covering the entire process range from loose packing to near-dense packing. The mass of each standard sample was measured using a precision balance, and the volume of each sample was reconstructed simultaneously using a capacitance tomography system to calculate the apparent density as the true value. The voxel values ​​of the dielectric constant distribution matrix were matched with the corresponding apparent density true values ​​to establish a dielectric constant-density data pair set. A quadratic polynomial fitting was performed on this data pair set, with the fitting objective being to minimize the mean square error between the predicted dielectric constant density and the true density, resulting in a quadratic polynomial mapping relationship between the dielectric constant and the powder apparent density. This mapping relationship captures the nonlinear characteristic that the dielectric constant tends to saturate with increasing density. The dielectric constant distribution matrix output from step two was substituted into this mapping relationship, and the apparent density distribution field of the powder was calculated voxel by voxel. This apparent density distribution field reflects the macroscopic density distribution of powder packing during cold pressing, but it does not consider the free volume effect in the shear band region and needs further correction to restore the true flow characteristics.

[0093] S302. A shear band density field is introduced to correct the apparent density distribution field of the powder for free volume. Considering the volume expansion effect in the shear transition region, the shear band region is regarded as a loose structure containing free volume, generating a free volume corrected density field. The calculation formula is as follows:

[0094]

[0095] In the formula, Let be the free volume corrected density at position (x, y) at time t. Let be the field element representing the apparent density distribution of the powder at position (x, y) at time t. Let be the shear band density field element at position (x, y) at time t. The shear band saturation density, The coefficient of volume expansion in the shear transition region is... Let t be the spatial coordinates and t be the time.

[0096] It is understandable that step S300 also includes S303 and S304, wherein:

[0097] S303. Calculate the effective flow area based on the porosity distribution field. Multiply the cross-sectional area of ​​the cavity by the local porosity complement to obtain the effective flow area distribution, generating the effective flow area field. Establish the Darcy flow model by combining the free volume corrected density field, considering the nonlinear resistance characteristics of powder deformation. The permeability adopts the Kozeny-Carman correction form and introduces the shear band tortuosity correction factor. Generate the tortuosity corrected permeability field. Calculate the local velocity field from the pressure gradient, the tortuosity corrected permeability field, and the effective viscosity. Perform area integration on the cross-section of the cavity by multiplying the local velocity field and the free volume corrected density field to solve for the real-time flow characteristic scalar.

[0098] It should be noted that this calculation originates from the effective medium approximation in porous media flow theory, where the porosity complement, i.e., the solid volume fraction, directly determines the openness of the flow channels. In actual cold pressing, the pore distribution is non-uniform, and local pore aggregation leads to a sharp reduction in the flow area, forming a flow bottleneck. The effective flow area field accurately captures this spatial heterogeneity. While the traditional Darcy law applies to linear laminar flow at low Reynolds numbers, the dynamic reorganization of contact force chains between powder particles during cold pressing results in significantly nonlinear flow. This model introduces the Forchheimer equation to describe the inertial effect, while also considering the nonlinear drag characteristics of powder deformation. Permeability is expressed in the Kozeny-Carman modified form, which represents permeability as a function of the ratio of the cube of porosity to the square of the specific surface area, reflecting the geometric constraint of the pore structure on the flow.

[0099] A shear band tortuosity correction factor is further introduced. The shear band network acts as a weak path within the powder, weakening the interparticle bonding and causing the flow channels to bend and circumvent, macroscopically manifesting as reduced permeability. The tortuosity correction factor is calculated from the gradient field of the shear band density field; a larger gradient indicates steeper shear band boundaries, more intense flow circumvention, and higher tortuosity. This correction factor is multiplied by the Kozeny-Carman permeability to generate a tortuosity-corrected permeability field, achieving spatial modulation of seepage characteristics by the shear band network.

[0100] The local velocity field is calculated using the pressure gradient of the cold pressing process, the tortuosity-corrected permeability field, and the effective viscosity of the powder. The pressure gradient is calculated based on real-time pressure feedback from the servo pressing unit and cavity geometry. The effective viscosity is described using the Herschel-Bulkley model, which includes three parameters: yield stress, consistency coefficient, and flow index, and is determined offline using a rotational rheometer. The local velocity field is calculated using a generalized form of Darcy's law, i.e., velocity equals permeability multiplied by the pressure gradient and then divided by the effective viscosity. The product of the local velocity field and the free volume-corrected density field is integrally applied across the cavity cross-section. The physical meaning of this product is the local mass flux. The density field corrects for the free volume effect in the shear band region, and the velocity field reflects the flow resistance of the pore structure and shear band network. The area integration operation condenses the distributed field quantities into a global scalar, solving for the real-time flow characteristic scalar. This scalar characterizes the powder mass flow rate through the cavity cross-section per unit time and is the core feedback variable for cold pressing process control.

[0101] S304. Perform fractal dimension analysis on the shear band density field. Calculate the fractal dimension of the shear band network using the box counting method. The box size sequence is arranged in a geometric progression. The fractal dimension is determined by the slope of a linear fit to a double logarithmic coordinate system, generating the shear band network fractal dimension. Extract the shear band network connectivity, defined as the ratio of the area of ​​the largest connected cluster to the total shear band region area, thus generating the shear band network connectivity. Combine the real-time flow characteristic scalar, the shear band network fractal dimension, the shear band network connectivity, and the acoustic emission energy change rate to construct a powder flow state feature vector. The calculation formula is as follows:

[0102]

[0103] In the formula, Let be the characteristic vector of the powder flow state at time t. Let be the real-time traffic characteristic scalar at time t. Let be the fractal dimension scalar of the shear band network at time t. Let be the scalar of the connectivity of the cut-band network at time t. Let dt be the rate of change of acoustic emission energy at time t, T be the transpose, and dt be the time derivative.

[0104] It should be noted that the box size sequence is arranged in a geometric progression, covering a scale range from the spatial resolution of the capacitive sensor to the feature size of the cavity, ensuring sufficient sampling in the scale-free interval. In a double logarithmic coordinate system, the number of covered boxes is linearly related to the box size, and the negative value of the slope is the fractal dimension, which generates the fractal dimension of the shear band network.

[0105] The physical implications of fractal dimension are rich: a dimension close to 1 indicates a sparse, linear distribution of shear bands, suggesting an underdeveloped network; a dimension close to 2 indicates densely interwoven, planar distribution of shear bands, indicating a highly mature network. This dimension quantifies the spatial filling complexity of the shear band network and, independently of connectivity, characterizes the network structure. Combining both allows for a more comprehensive assessment of the cold-pressing stage. In practical research, a dynamic model of shear band network development can be established by comparing the evolution of fractal dimension at different cold-pressing stages.

[0106] Among them, the 8-connected neighborhood includes orthogonal and diagonal adjacency relationships of pixels, which is better at capturing the oblique connectivity of shear bands than the 4-connected neighborhood. The area of ​​each cluster is marked, and the largest connected cluster is selected. The ratio of the area of ​​the largest connected cluster to the total area of ​​the shear band region is defined as the shear band network connectivity, generating the shear band network connectivity. When the connectivity is below the percolation threshold, the shear band clusters are isolated and dispersed, and the powder maintains its load-bearing capacity; when the connectivity exceeds the threshold, the largest cluster penetrates the system, forming a continuous weak path, and the powder becomes unstable. The determination of the percolation threshold requires combination of discrete element simulation or in-situ experimental statistics and is material and process dependent. Real-time monitoring of the shear band network connectivity provides the physical basis for early warning of cold-pressing instability.

[0107] A powder flow state feature vector is constructed by combining real-time flow characteristic scalars, shear band network fractal dimension, shear band network connectivity, and acoustic emission energy change rate. The four components of this feature vector correspond to flow transport characteristics, network structure complexity, network connectivity, and dynamic activity intensity, respectively, forming a multidimensional state description of the cold pressing process. Before constructing the feature vector, each component needs to be normalized to eliminate dimensional differences. Z-score normalization is suitable for components with approximately normal distributions, while Min-Max scaling is suitable for bounded components. This powder flow state feature vector serves as the input feature for subsequent spatiotemporal prediction by a graph neural network, balancing model accuracy and computational efficiency in terms of compactness and completeness.

[0108] S400: Expand the powder flow state feature vector into a time-series graph structure data, using the local state variables of each sensor sub-region as node features and the spatial adjacency relationship of the shear zone region in the shear zone density field as edge features; input this time-series graph structure data into a graph neural network-long short-term memory coupled model; the graph neural network extracts the spatial topological correlation features of the shear zone through message passing and outputs a spatial embedding feature vector; the long short-term memory network performs time-series dynamic modeling on this spatial embedding feature vector and outputs a time-series prediction feature vector; after decoding, a flow prediction sequence and instability risk probability are generated; the percolation critical connectivity threshold determined by offline calibration experiments is called, and the shear zone network connectivity is compared with the threshold in real time; if the threshold is exceeded, an instability warning signal is generated; if the threshold is not exceeded, the normal state is maintained.

[0109] It is understood that in this step, S400 includes S401, S402, and S403, wherein:

[0110] S401. The powder flow state feature vector is spatially divided according to the sensor sub-regions. Each sub-region corresponds to a node in the graph. The node feature vector is composed of the real-time flow feature scalar of the sub-region, the fractal dimension scalar of the shear band network, the connectivity scalar of the shear band network, and the acoustic emission energy change rate, generating a node feature matrix. The edge set of the graph is determined by the spatial adjacency relationship of the shear band regions in the shear band density field. When the spatial distance between the corresponding sub-regions of two nodes in the shear band density field is less than the connectivity threshold, an edge connection is established. The edge weight is proportional to the spatial correlation coefficient of the shear band density between adjacent nodes, generating a weighted adjacency matrix. Taking the current time as the endpoint, historical data of a fixed time length is extracted to construct a sliding window time series graph sequence with a time span, generating time series graph sequence data.

[0111] It should be noted that the feature selection of the node feature matrix is ​​based on the physical causal chain: flow rate represents the current filling state, fractal dimension represents the shear band development stage, connectivity represents the degree of instability approach, and acoustic emission energy change rate represents the intensity of shear activity. Together, these four constitute a complete state description of the cold pressing process.

[0112] It should be noted that the powder flow state feature vector is spatially divided into sub-regions corresponding to the corresponding electrodes of the capacitive sensor array. Each sub-region corresponds to a node in the graph. The node feature vector consists of the real-time flow characteristic scalar of the sub-region, the fractal dimension scalar of the shear band network, the connectivity scalar of the shear band network, and the acoustic emission energy change rate, generating a node feature matrix. When the spatial distance between the corresponding sub-regions of two nodes in the shear band density field is less than a connectivity threshold, an edge connection is established. This connectivity threshold is determined by a combination of the typical width of the shear band and the spatial resolution of the sensor, ensuring that the edge connection only captures physically relevant shear band associations. The edge weight is proportional to the spatial correlation coefficient of the shear band density between adjacent nodes. This correlation coefficient uses the Pearson correlation coefficient or mutual information metric to reflect the statistical dependence of the shear band development intensity between the two sub-regions, generating a weighted adjacency matrix. The asymmetry of the weighted adjacency matrix stems from the directionality of shear band propagation; the influence of high-density regions on low-density regions is stronger than the reverse. This characteristic is further characterized by directed graphs or attention mechanisms. Using the current moment as the endpoint, historical data of a fixed time length is extracted to construct a sliding window time series graph sequence, generating time series graph sequence data. The sliding window mechanism ensures a balance between the model's real-time performance and historical dependency; an excessively long window increases computational burden and includes outdated information, while an excessively short window loses key dynamic features. The construction of the time series graph sequence data marks a paradigm shift from static field analysis to dynamic network analysis, providing structured input for spatiotemporal feature extraction in graph neural networks.

[0113] S402. Input the time-series graph sequence data into the graph neural network-long short-term memory coupled model; the graph neural network takes the node feature matrix and weighted adjacency matrix as input, calculates the attention coefficient between nodes through the graph attention mechanism, aggregates neighbor node information through message passing, extracts the spatial topological correlation features of the shear band, and outputs a spatial embedding feature vector sequence; the long short-term memory network takes the spatial embedding feature vector sequence as input, captures the temporal dependency relationship through the gating mechanism, models the flow evolution trend and shear band network dynamics, and outputs the temporal prediction feature vector;

[0114] It should be noted that the design of this coupled architecture stems from the dual characteristics of the cold-pressing process prediction task: the spatial topology of the shear band network needs to be extracted by a graph neural network, while the temporal dependence of flow evolution and network dynamics requires long short-term memory network modeling. The coupling of these two aspects enables integrated learning of spatiotemporal features. Meanwhile, the node feature matrix carries multi-dimensional state information of each sub-region, and the weighted adjacency matrix encodes the spatial correlation strength of the shear band network. Attention coefficients between nodes are calculated through a graph attention mechanism, which assigns adaptive neighbor weights to each node, unlike the fixed-weight averaging of traditional graph convolution. The calculation of the attention coefficients integrates node feature similarity with prior edge weights; the more similar the features and the higher the edge weights, the larger the attention coefficient.

[0115] In this step, message passing follows an iterative update rule: each node receives messages from its neighbors, fuses them with its own features, and outputs an updated embedding vector. Multi-layer message passing expands the receptive field, enabling distant nodes to interact indirectly through intermediate nodes, ultimately capturing the global topology. This process outputs a sequence of spatial embedding feature vectors, each vector encoding the topological context information of its corresponding sub-region. The Long Short-Term Memory (LSTM) network, through a triple gating mechanism of input, forget, and output gates, selectively memorizes long-term states and absorbs short-term inputs, solving the gradient vanishing problem of traditional recurrent networks. Using the spatial embedding feature vector sequence as input, the gating mechanism captures temporal dependencies, modeling the flow evolution trend and shear band network dynamics. The flow evolution trend reflects the macroscopic transport laws of the filling process, while the shear band network dynamics reflect the growth, connection, and saturation processes of the microstructure; their coupling determines the future state of the cold-pressing system. This process outputs a temporal prediction feature vector, fusing spatial topological information and temporal evolution information to provide a complete hidden state representation for subsequent decoding and prediction results.

[0116] S403. Input the time-series prediction feature vector into the fully connected decoding layer, and generate the traffic prediction sequence and instability risk probability for multiple future time steps through linear transformation and activation function mapping.

[0117] It should be noted that the formulas for calculating the flow prediction sequence and the probability of instability risk are as follows:

[0118]

[0119] in,

[0120] In the formula, The predicted output sequence for the next n time steps. Let be the predicted output vector at time t+jΔt. Let be the predicted flow rate at time t+jΔt. Let be the probability of instability risk at time t+jΔt, n be the number of prediction steps in the time domain, s be the step index in the prediction time domain, and T be the transpose.

[0121] The S500 uses the flow prediction sequence, instability risk probability, and instability early warning signal as inputs to construct a multi-objective cost function, which includes a flow tracking accuracy term, an instability risk suppression term, a pressing energy consumption term, and an emergency safety protection term. A rolling time-domain optimization solution is performed on the multi-objective cost function to obtain the optimal control sequence, which in turn generates the powder supply rate adjustment, pressing pressure adjustment, and emergency pressure relief valve opening command. The powder supply rate adjustment is output to the powder supply driver, the pressing pressure adjustment is output to the servo pressing unit, and the emergency pressure relief valve opening command is output to the safety pressure relief valve, achieving coordinated control of high-precision flow tracking and active suppression of shear band instability during cold pressing.

[0122] It is understood that in this step, S500 includes S501 and S502, wherein:

[0123] S501. Perform rolling time-domain optimization to solve the multi-objective cost function. The prediction time domain is taken as 100 to 200 time steps, and the control time domain is taken as half of the prediction time domain. The constraints include: the upper limit of the pressing pressure is the product of the yield strength of the amorphous alloy and the correction value of the average porosity; the powder supply rate is limited to between zero and the maximum value; and the emergency pressure relief response time does not exceed 50 milliseconds. Use a sequential quadratic programming algorithm to solve the problem, obtain the optimal control sequence, and generate the optimal powder supply rate sequence, the optimal pressing pressure sequence, and the optimal emergency pressure sequence.

[0124] It should be noted that rolling time-domain optimization is performed on the multi-objective cost function. Rolling time-domain optimization is a model predictive control strategy. Its core idea is to predict future dynamics based on the current state at each sampling time, solve for the optimal control sequence in a finite time domain, implement only the first step of control, and start rolling optimization again at the next time step. This strategy balances prediction accuracy and computational feasibility and is suitable for real-time control of cold pressing processes.

[0125] Its prediction time domain spans 100 to 200 time steps, covering the typical evolution cycle of shear band networks from initiation to percolation, ensuring that the prediction span includes key dynamic characteristics. The control time domain is half of the prediction time domain, reducing the dimensionality of optimization variables, accelerating solution convergence, while retaining sufficient control degrees of freedom. The constraint conditions are set by integrating material physics and process safety: the upper limit of the pressing pressure is the product of the yield strength of the amorphous alloy and the average porosity correction value. The higher the porosity, the smaller the effective bearing area, and the lower the allowable pressure, preventing powder crushing; the powder supply rate is limited between zero and the maximum value, with zero rate corresponding to powder supply interruption protection, and the maximum rate being the mechanical limit of the equipment; the emergency depressurization response time does not exceed 50 milliseconds, ensuring rapid depressurization when instability is triggered, avoiding mold damage.

[0126] In this step, a sequential quadratic programming algorithm is used to solve the problem. This algorithm transforms the nonlinear constrained optimization problem into a series of quadratic programming subproblems that are solved iteratively. Utilizing the Hessian matrix approximation of the Lagrange function, it achieves fast convergence and is suitable for medium-scale constrained optimization. The solution yields the optimal control sequence, generating the optimal powder supply rate sequence, the optimal pressing pressure sequence, and the optimal emergency pressure sequence. These three sequences correspond to the future action plans of the three actuators, respectively.

[0127] S502. Generate control commands for the current moment from the optimal control sequence: the powder supply rate adjustment amount is the difference between the optimal supply rate and the actual supply rate at the previous moment; the pressing pressure adjustment amount is the difference between the optimal pressing pressure and the actual pressing pressure at the previous moment; the emergency pressure relief valve opening command is the ratio of the optimal emergency pressure to the maximum allowable pressure converted into a percentage; generate control commands for powder supply rate adjustment amount, pressing pressure adjustment amount, and emergency pressure relief valve opening amount.

[0128] It should be noted that a multi-objective cost function is constructed using the flow prediction sequence, the probability of instability risk, and the instability early warning signal as inputs. The calculation formula is as follows:

[0129]

[0130] In the formula, J is the value of the multi-objective cost function, and n is the number of prediction time-domain steps. for Traffic flow forecast at any given time For target traffic, for The probability of instability at any given moment. To suppress the pressure adjustment, Alarm(t) is the instability warning signal at time t. For emergency pressure relief, To track precision weights, Weights to suppress instability As energy consumption weight, For safety protection weights, s is the step index in the prediction time domain, t is the current time, and Δt is the sampling time interval.

[0131] It should be noted that the powder supply rate adjustment is the difference between the optimal supply rate and the actual supply rate at the previous moment. This differential form directly drives the servo motor speed adjustment, realizing incremental adjustment of the powder supply. The pressing pressure adjustment is the difference between the optimal pressing pressure and the actual pressing pressure at the previous moment. This differential form drives the servo hydraulic system pressure adjustment, realizing a smooth transition of pressing force. The emergency pressure relief valve opening command is the ratio of the optimal emergency pressure to the maximum allowable pressure converted into a percentage. This normalized form is directly mapped to the proportional valve opening, realizing linear control of the pressure relief. The generation of powder supply rate adjustment control command, pressing pressure adjustment control command, and emergency pressure relief valve opening control command means that the three commands are converted from digital to analog or transmitted via fieldbus and output to the powder supply driver, servo pressing unit, and safety pressure relief valve respectively, completing the closed-loop control of perception-prediction-decision-execution, realizing the coordinated control of high-precision flow tracking and active suppression of shear band instability during cold pressing.

[0132] The cost function includes four terms: the tracking accuracy term of the deviation between the flow prediction sequence and the target flow, the suppression term of the probability of instability risk, the energy consumption term of suppressing the pressure adjustment, and the emergency safety protection term triggered by the instability warning signal; among them, the safety protection term has a much larger weight than the other three terms.

[0133] Example 2:

[0134] like Figure 2 As shown, this embodiment provides an intelligent sensing system for the cold pressing flow rate of amorphous alloy powder. (See attached image.) Figure 2 The system includes:

[0135] Module 701 is used to embed a miniature capacitance sensor array and a broadband acoustic emission sensor array on the cavity wall of a cold-pressing mold. It collects capacitance tomography signals and acoustic emission signals of amorphous alloy powder during the cold pressing process, and then constructs the original capacitance measurement tensor and acoustic emission feature vector. Wavelet packet decomposition is performed on the acoustic emission signal to extract shear band initiation energy features and construct the acoustic emission feature vector. The original capacitance measurement tensor is composed of a three-dimensional stack of multi-frequency capacitance measurement value sequences according to electrode number, frequency number, and time series points.

[0136] Decoupling Generation Module 702: This module is used to input the original capacitance measurement tensor and acoustic emission feature vector into the capacitance tomography reconstruction algorithm, perform adaptive iterative reconstruction based on acoustic emission features, and output a spatiotemporally continuous dielectric constant distribution matrix; it calls the prior dielectric dispersion characteristics obtained from the offline calibration experiment to construct a multi-frequency dielectric-mechanical coupled Gaussian mixture model; using the equiprobability surface of the equivalent dielectric constant probability distribution of each phase output by the model as the decision boundary, it replaces the fixed dielectric threshold with a frequency-energy related nonlinear dielectric classification surface, performs soft classification decoupling on the dielectric constant distribution matrix, and generates a shear band density field and porosity distribution field;

[0137] Extraction and Combination Module 703: This module is used to call the dielectric constant-density mapping relationship established by the offline calibration experiment, convert the dielectric constant distribution matrix after soft classification decoupling into the powder apparent density distribution field; introduce the shear band density field to perform free volume correction on the powder apparent density distribution field, generating a free volume corrected density field; calculate the effective flow area field based on the porosity distribution field, and establish the Darcy flow model in combination with the free volume corrected density field to obtain the real-time flow characteristic scalar; simultaneously perform fractal dimension analysis on the shear band density field to extract the shear band network connectivity; combine the real-time flow characteristic scalar, the shear band network fractal dimension, the shear band network connectivity, and the acoustic emission energy change rate in the acoustic emission characteristic vector to construct the powder flow state characteristic vector;

[0138] Judgment module 704: This module expands the powder flow state feature vector into a time-series graph structure, using the local state variables of each sensor sub-region as node features and the spatial adjacency relationships of the shear zone region in the shear zone density field as edge features. This time-series graph structure data is input into a graph neural network-long short-term memory coupled model. The graph neural network extracts the spatial topological correlation features of the shear zone through message passing and outputs a spatial embedding feature vector. The long short-term memory network performs time-series dynamic modeling on this spatial embedding feature vector and outputs a time-series prediction feature vector. After decoding, a flow prediction sequence and instability risk probability are generated. The percolation critical connectivity threshold determined by offline calibration experiments is called, and the shear zone network connectivity is compared with this threshold in real time. If the threshold is exceeded, an instability warning signal is generated; otherwise, the normal state is maintained.

[0139] The sensing and control module 705 is used to construct a multi-objective cost function based on the flow prediction sequence, instability risk probability, and instability early warning signal. The multi-objective cost function includes a flow tracking accuracy term, an instability risk suppression term, a pressing energy consumption term, and an emergency safety protection term. The module performs rolling time-domain optimization on the multi-objective cost function to obtain the optimal control sequence, and then generates the powder supply rate adjustment amount, the pressing pressure adjustment amount, and the emergency pressure relief valve opening command. The module outputs the powder supply rate adjustment amount to the powder supply driver, the pressing pressure adjustment amount to the servo pressing unit, and the emergency pressure relief valve opening command to the safety pressure relief valve, thereby realizing the coordinated control of high-precision flow tracking and active suppression of shear band instability during cold pressing.

[0140] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0141] Example 3:

[0142] Corresponding to the above method embodiments, this embodiment also provides an intelligent sensing device for the cold-pressed flow rate of amorphous alloy powder. The intelligent sensing device for the cold-pressed flow rate of amorphous alloy powder described below and the intelligent sensing method for the cold-pressed flow rate of amorphous alloy powder described above can be referred to in correspondence.

[0143] Figure 3 This is a block diagram illustrating an intelligent sensing device 800 for the cold pressing flow rate of amorphous alloy powder, according to an exemplary embodiment. Figure 3 As shown, the amorphous alloy powder cold-pressing flow rate intelligent sensing device 800 includes a processor 801 and a memory 802. The amorphous alloy powder cold-pressing flow rate intelligent sensing device 800 also includes one or more of a multimedia component 803, an I / O interface 804, and a communication component 805.

[0144] The processor 801 controls the overall operation of the amorphous alloy powder cold-pressed flow rate intelligent sensing device 800 to complete all or part of the steps in the aforementioned amorphous alloy powder cold-pressed flow rate intelligent sensing method. The memory 802 stores various types of data to support the operation of the amorphous alloy powder cold-pressed flow rate intelligent sensing device 800. This data may include, for example, instructions for any application or method operating on the amorphous alloy powder cold-pressed flow rate intelligent sensing device 800, as well as application-related data such as contact data, sent and received messages, images, audio, video, etc. The memory 802 can be implemented using any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The multimedia component 803 may include a screen and an audio component. The screen may be, for example, a touchscreen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signals may be further stored in the memory 802 or transmitted via the communication component 805. The audio component also includes at least one speaker for outputting audio signals. I / O interface 804 provides an interface between processor 801 and other interface modules, such as keyboards, mice, or buttons. These buttons can be virtual or physical. Communication component 805 is used for wired or wireless communication between the amorphous alloy powder cold-pressed flow intelligent sensing device 800 and other devices. Wireless communication includes, for example, Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, or 4G, or one or more combinations thereof. Therefore, the corresponding communication component 805 may include a Wi-Fi module, a Bluetooth module, or an NFC module.

[0145] In an exemplary embodiment, the amorphous alloy powder cold-pressing flow rate intelligent sensing device 800 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above-described amorphous alloy powder cold-pressing flow rate intelligent sensing method.

[0146] In another exemplary embodiment, a computer-readable storage medium including program instructions is also provided, which, when executed by a processor, implement the steps of the above-described intelligent sensing method for cold-pressed flow rate of amorphous alloy powder. For example, the computer-readable storage medium may be the memory 802 including program instructions, which may be executed by the processor 801 of the intelligent sensing device 800 for cold-pressed flow rate of amorphous alloy powder to complete the above-described intelligent sensing method for cold-pressed flow rate of amorphous alloy powder.

[0147] Example 4:

[0148] Corresponding to the above method embodiments, this embodiment also provides a readable storage medium. The readable storage medium described below can be referred to in conjunction with the intelligent sensing method for cold pressing flow of amorphous alloy powder described above.

[0149] A computer program is stored on a readable storage medium, and when the computer program is executed by a processor, it implements the steps of the intelligent sensing method for cold-pressed flow rate of amorphous alloy powder in the above method embodiments.

[0150] Specifically, the readable storage medium can be a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, or any other readable storage medium capable of storing program code.

[0151] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An amorphous alloy powder cold pressing flow intelligent sensing method, characterized in that, include: A miniature capacitance sensor array and a broadband acoustic emission sensor array are embedded in the cavity wall of a cold-pressing mold to collect capacitance tomography signals and acoustic emission signals of amorphous alloy powder during the cold pressing process. Then, the original capacitance measurement tensor and acoustic emission feature vector are constructed. Wavelet packet decomposition is performed on the acoustic emission signal to extract the shear band initiation energy features and construct the acoustic emission feature vector. The original capacitance measurement tensor is composed of a three-dimensional stack of multi-frequency capacitance measurement value sequences according to electrode number, frequency number, and time series points. The original capacitance measurement tensor and acoustic emission feature vector are input into the capacitance tomography reconstruction algorithm to perform adaptive iterative reconstruction based on acoustic emission features, outputting a spatiotemporally continuous dielectric constant distribution matrix. The prior dielectric dispersion characteristics obtained from offline calibration experiments are called to construct a multi-frequency dielectric-mechanical coupled Gaussian mixture model. The equiprobability surface of the equivalent dielectric constant probability distribution of each phase output by the model is used as the decision boundary. The fixed dielectric threshold is replaced with a frequency-energy related nonlinear dielectric classification surface. Soft classification decoupling is performed on the dielectric constant distribution matrix to generate a shear band density field and a porosity distribution field. The dielectric constant-density mapping relationship established by the offline calibration experiment is called to convert the dielectric constant distribution matrix after soft classification decoupling into the powder apparent density distribution field; the shear band density field is introduced to correct the free volume of the powder apparent density distribution field to generate the free volume corrected density field; the effective flow area field is calculated based on the porosity distribution field, and the Darcy flow model is established in combination with the free volume corrected density field to obtain the real-time flow characteristic scalar. Simultaneously, fractal dimension analysis was performed on the shear band density field to extract the shear band network connectivity; the real-time flow characteristic scalar, the shear band network fractal dimension, the shear band network connectivity, and the acoustic emission energy change rate in the acoustic emission characteristic vector were combined to construct the powder flow state characteristic vector; The powder flow state feature vector is expanded into a time-series graph structure data, with the local state variables of each sensor sub-region as node features and the spatial adjacency relationship of the shear zone region in the shear zone density field as edge features. The time series graph structure data is input into a graph neural network-long short-term memory coupled model; the graph neural network extracts the spatial topological correlation features of the shear band through message passing and outputs a spatial embedding feature vector; the long short-term memory network performs temporal dynamic modeling on the spatial embedding feature vector and outputs a temporal prediction feature vector. The flow prediction sequence and instability risk probability are generated after decoding; Call the percolation critical connectivity threshold determined by the offline calibration experiment and compare the shear band network connectivity with the threshold in real time; If the threshold is exceeded, an instability warning signal will be generated; If the limit is not exceeded, the normal status indicator will be maintained. Using the flow prediction sequence, instability risk probability, and instability early warning signal as inputs, a multi-objective cost function is constructed. This multi-objective cost function includes a flow tracking accuracy term, an instability risk suppression term, a pressing energy consumption term, and an emergency safety protection term. A rolling time-domain optimization solution is performed on the multi-objective cost function to obtain the optimal control sequence, which in turn generates the powder supply rate adjustment, pressing pressure adjustment, and emergency pressure relief valve opening command. The powder supply rate adjustment is output to the powder supply driver, the pressing pressure adjustment is output to the servo pressing unit, and the emergency pressure relief valve opening command is output to the safety pressure relief valve, achieving coordinated control of high-precision flow tracking and active suppression of shear band instability during cold pressing.

2. The method of claim 1, wherein the method comprises: The method involves embedding a micro-capacitive sensor array and a broadband acoustic emission sensor array into the cavity wall of the cold-pressing mold to collect capacitance tomography signals and acoustic emission signals of amorphous alloy powder during the cold-pressing process. This allows for the construction of the original capacitance measurement tensor and acoustic emission feature vector. Wavelet packet decomposition is then performed on the acoustic emission signal to extract shear band initiation energy features, and the acoustic emission feature vector is constructed, including: A ring-shaped micro-capacitive sensor array is embedded in the cavity wall of a cold-pressing mold. The array consists of excitation electrodes and measurement electrodes arranged alternately. A multi-frequency excitation mode is adopted, with low frequency used to penetrate the powder gap and high frequency used to distinguish the microstructure of the shear band. The acquisition frequency is not less than 10,000 frames per second to obtain a sequence of multi-frequency capacitance measurement values ​​on the cross section. The multi-frequency capacitance measurement sequence is stacked in three dimensions according to electrode number, frequency number, and time series points to construct the original capacitance measurement tensor, and its calculation formula is as follows: In the formula, for Time of the first At the excitation frequency, the first The excitation electrode and the first The capacitance measurement value between the measuring electrodes This represents the induced charge between the corresponding electrode pairs. for Time of the first The excitation electrode in the first... Excitation voltage at a given frequency Number the excitation electrodes. Here, k is the measurement electrode number, k is the frequency number, and t is the discrete time sequence number; A broadband acoustic emission sensor array was arranged in the shear stress concentration region of the cavity wall. The center frequency of the sensor was matched with the characteristic frequency of the shear band of the amorphous alloy, and the sampling rate was set to 5 times the highest frequency of the acoustic emission signal. Acoustic emission signal sequences of friction, collision and shear band initiation between amorphous alloy powder particles were collected. Wavelet packet decomposition was performed on the acoustic emission signal sequence with 4 decomposition layers to obtain wavelet coefficients of 16 frequency bands. Wavelet energy of the characteristic frequency range of shear band from 300 kHz to 500 kHz corresponding to the 9th to 12th frequency bands was extracted to construct the shear band initiation energy feature. The shear band initiation energy features of each sensor node were arranged according to spatial position to construct the acoustic emission feature vector.

3. The method of claim 1, wherein the method comprises: The original capacitance measurement tensor and acoustic emission feature vector are input into the capacitance tomography reconstruction algorithm to perform adaptive iterative reconstruction based on acoustic emission features, outputting a spatiotemporally continuous dielectric constant distribution matrix; prior dielectric dispersion characteristics obtained from offline calibration experiments are used to construct a multi-frequency dielectric-mechanical coupled Gaussian mixture model; using the equiprobability surface of the equivalent dielectric constant probability distribution of each phase output by this model as the decision boundary, the fixed dielectric threshold is replaced with a frequency-energy related nonlinear dielectric classification surface, including: The original capacitance measurement tensor and acoustic emission feature vector are input into the capacitance tomography reconstruction algorithm to construct an adaptive iterative reconstruction operator based on acoustic emission features. The iteration step size is dynamically adjusted according to the shear band initiation energy, and its calculation formula is as follows: In the formula, is the adaptive iteration step length at time t, is the reference step length, is the acoustic emission coupling coefficient, is the shear band initiation energy feature at time t, is the acoustic emission energy mean value, is the acoustic emission energy standard deviation; Perform iterative reconstruction until convergence, and output a spatiotemporally continuous dielectric constant distribution matrix; call the prior dielectric dispersion characteristics obtained from the offline calibration experiment, which measures the complex dielectric constant versus frequency curves of undeformed amorphous alloy powder, pre-compression induced shear band powder, and residual gas in the mold under standard temperature and humidity conditions; extract the static dielectric constant, optical frequency dielectric constant, and dielectric relaxation time of each phase from the complex dielectric constant versus frequency curves, where the static dielectric constant, optical frequency dielectric constant, and dielectric relaxation time are the three characteristic parameters of the Debye relaxation model, which together constitute the prior dielectric dispersion characteristics; Based on the prior dielectric dispersion characteristics, a conditional probability distribution of the dielectric constant of each phase state with respect to frequency is established. Using the conditional probability distribution of each phase state as components and the phase volume fraction as the mixing coefficient, a multi-frequency dielectric-mechanical coupled Gaussian mixture model is constructed. The input of this model is frequency and acoustic emission energy, and the output is the probability distribution of the equivalent dielectric constant of each phase state. Using the equiprobability surface of the probability distribution of the equivalent dielectric constant of each phase state output by this model as the decision boundary, the fixed dielectric threshold is replaced with a frequency-energy related nonlinear dielectric classification surface.

4. The method of claim 3, wherein the method comprises: The soft classification decoupling of the dielectric constant distribution matrix generates a shear band density field and a porosity distribution field, including: Soft classification decoupling is performed on the dielectric constant distribution matrix, and the posterior probability of each spatial location belonging to each phase state is calculated to generate a three-phase probability density field. The calculation formula is as follows: In the formula, Let be the posterior probability that the position (x, y) at time t belongs to the m-th phase. Let be the element of the dielectric constant distribution matrix at position (x, y) at time t. Let be the mixing coefficient of the m-th phase. For the m-th element, the excitation frequency f and the shear band volume fraction are relevant. The mean function, It follows a Gaussian distribution. Let M be the variance of the m-th phase, M be the total number of phase states, and m be the phase state number. For the summation index, Let f be the spatial coordinates and f be the excitation frequency. This represents the volume fraction of the shear band. For the first The mixing coefficient of the phase, For the first Related to excitation frequency f and shear band volume fraction The mean function, For the first The variance of the phase; Shear band phase probability components are extracted from the three-phase probability density field. After Gaussian kernel spatial smoothing and adaptive threshold binarization, isolated regions with areas smaller than the minimum cluster area are removed, and connected component labeling is performed. The shear band phase probability values ​​of each connected component are averaged by area to generate the shear band density field. At the same time, the gas phase probability components are extracted and processed in the same way to serve as the porosity distribution field.

5. The intelligent sensing method for cold-pressed flow rate of amorphous alloy powder according to claim 4, characterized in that, The dielectric constant-density mapping relationship established by calling the offline calibration experiment is used to convert the dielectric constant distribution matrix after soft classification decoupling into a powder apparent density distribution field; a shear band density field is introduced to perform free volume correction on this powder apparent density distribution field, generating a free volume corrected density field, including: The dielectric constant-density mapping relationship established by the offline calibration experiment is invoked. This offline calibration experiment uses a precision balance to measure the ratio of powder mass to capacitance chromatography reconstruction volume under different compaction states to establish a quadratic polynomial mapping between dielectric constant and powder apparent density. The dielectric constant distribution matrix is ​​then substituted into this mapping relationship to convert it into the powder apparent density distribution field. A shear band density field is introduced to correct the apparent density distribution field of the powder for free volume. Considering the volume expansion effect in the shear transition region, the shear band region is regarded as a loose structure containing free volume, generating a free volume corrected density field. The calculation formula is as follows: In the formula, Let be the free volume corrected density at position (x, y) at time t. Let be the field element representing the apparent density distribution of the powder at position (x, y) at time t. Let be the shear band density field element at position (x, y) at time t. The shear band saturation density, The coefficient of volume expansion in the shear transition region is... Let t be the spatial coordinates and t be the time.

6. The intelligent sensing method for cold-pressed flow rate of amorphous alloy powder according to claim 5, characterized in that, The coefficient of the volume expansion effect in the shear transition zone ranges from 0.02 to 0.05, and the minimum cluster area is determined by the ratio of the average particle size of the powder particles to the imaging resolution.

7. The intelligent sensing method for cold-pressed flow rate of amorphous alloy powder according to claim 1, characterized in that, The effective flow area field is calculated based on the porosity distribution field, and the Darcy flow model is established by combining the free volume corrected density field to obtain the real-time flow characteristic scalar. Simultaneously, fractal dimension analysis was performed on the shear band density field to extract the shear band network connectivity. The real-time flow characteristic scalar, the shear band network fractal dimension, the shear band network connectivity, and the acoustic emission energy change rate in the acoustic emission characteristic vector were combined to construct a powder flow state characteristic vector, which includes: The effective flow area is calculated based on the porosity distribution field. The cross-sectional area of ​​the cavity is multiplied by the local porosity complement to obtain the effective flow area distribution, generating the effective flow area field. A Darcy flow model is established by combining the free volume corrected density field. The nonlinear resistance characteristics of powder deformation are considered. The permeability adopts the Kozeny-Carman correction form and introduces the shear band tortuosity correction factor. A tortuosity corrected permeability field is generated. The local velocity field is calculated by the pressure gradient, the tortuosity corrected permeability field and the effective viscosity. The product of the local velocity field and the free volume corrected density field is integraled on the cross-section of the cavity to solve for the real-time flow characteristic scalar. Fractal dimension analysis was performed on the shear band density field. The fractal dimension of the shear band network was calculated using the box counting method, with the box size sequence arranged in a geometric progression. The fractal dimension was determined by the slope of a linear fit to a double logarithmic coordinate system. The shear band network connectivity was extracted and defined as the ratio of the area of ​​the largest connected cluster to the total shear band region area. The real-time flow characteristic scalar, the shear band network fractal dimension, the shear band network connectivity, and the acoustic emission energy change rate were combined to construct a powder flow state feature vector. The calculation formula is as follows: In the formula, Let be the characteristic vector of the powder flow state at time t. Let be the real-time traffic characteristic scalar at time t. Let be the fractal dimension scalar of the shear band network at time t. Let be the scalar of the connectivity of the cut-band network at time t. Let dt be the rate of change of acoustic emission energy at time t, T be the transpose, and dt be the time derivative.

8. The intelligent sensing method for cold-pressed flow rate of amorphous alloy powder according to claim 1, characterized in that, The powder flow state feature vector is expanded into a time-series graph structure data, with the local state quantities of each sensor sub-region as node features and the spatial adjacency relationship of the shear band region in the shear band density field as edge features. The time series graph structure data is input into a graph neural network-long short-term memory coupled model; the graph neural network extracts the spatial topological correlation features of the shear band through message passing and outputs a spatial embedding feature vector. Long Short-Term Memory (LSTM) networks perform temporal dynamic modeling of the spatially embedded feature vectors and output temporal predicted feature vectors. The decoded sequence generates a flow prediction sequence and an instability risk probability, including: The powder flow state feature vector is spatially divided according to the sensor sub-regions. Each sub-region corresponds to a node in the graph. The node feature vector consists of the real-time flow feature scalar of the sub-region, the fractal dimension scalar of the shear band network, the connectivity scalar of the shear band network, and the acoustic emission energy change rate, generating a node feature matrix. The edge set of the graph is determined by the spatial adjacency relationship of the shear band regions in the shear band density field. When the spatial distance between the corresponding sub-regions of two nodes in the shear band density field is less than the connectivity threshold, an edge connection is established. The edge weight is proportional to the spatial correlation coefficient of the shear band density between adjacent nodes, generating a weighted adjacency matrix. Using the current time as the endpoint, historical data of a fixed time length is extracted to construct a sliding window time series graph sequence with a time span, generating time series graph sequence data. The time-series graph sequence data is input into the graph neural network-long short-term memory coupled model. The graph neural network takes the node feature matrix and weighted adjacency matrix as input, calculates the attention coefficient between nodes through the graph attention mechanism, aggregates neighbor node information through message passing, extracts the spatial topological correlation features of the shear band, and outputs a sequence of spatially embedded feature vectors. The long short-term memory network takes the sequence of spatially embedded feature vectors as input, captures the temporal dependencies through the gating mechanism, models the flow evolution trend and shear band network dynamics, and outputs a temporal prediction feature vector. The time-series prediction feature vector is input into the fully connected decoding layer, and after linear transformation and activation function mapping, a traffic prediction sequence and instability risk probability for multiple future time steps are generated.

9. The intelligent sensing method for cold-pressing flow rate of amorphous alloy powder according to claim 1, characterized in that, The rolling time-domain optimization solution of the multi-objective cost function is performed to obtain the optimal control sequence, which in turn generates commands for powder supply rate adjustment, pressing pressure adjustment, and emergency pressure relief valve opening, including: A rolling time-domain optimization solution is performed on the multi-objective cost function, with the prediction time domain taking 100 to 200 time steps and the control time domain taking half of the prediction time domain. The constraints include: the upper limit of the pressing pressure is the product of the yield strength of the amorphous alloy and the correction value of the average porosity; the powder supply rate is limited to between zero and the maximum value; and the emergency pressure relief response time does not exceed 50 milliseconds. A sequential quadratic programming algorithm is used to solve the problem, obtaining the optimal control sequence, and generating the optimal powder supply rate sequence, the optimal pressing pressure sequence, and the optimal emergency pressure sequence. The control commands for the current moment are generated from the optimal control sequence: the powder supply rate adjustment is the difference between the optimal supply rate and the actual supply rate at the previous moment; the pressing pressure adjustment is the difference between the optimal pressing pressure and the actual pressing pressure at the previous moment; and the emergency pressure relief valve opening command is the ratio of the optimal emergency pressure to the maximum allowable pressure converted into a percentage. The powder supply rate adjustment control command, the pressing pressure adjustment control command, and the emergency pressure relief valve opening command are generated.

10. A smart sensing system for cold-pressed flow rate of amorphous alloy powder, based on the smart sensing method for cold-pressed flow rate of amorphous alloy powder as described in claim 1, characterized in that, include: The construction module is used to embed a miniature capacitance sensor array and a broadband acoustic emission sensor array on the cavity wall of a cold-pressing mold. It collects capacitance tomography signals and acoustic emission signals of amorphous alloy powder during the cold pressing process, and then constructs the original capacitance measurement tensor and acoustic emission feature vector. Wavelet packet decomposition is performed on the acoustic emission signal to extract the shear band initiation energy features and construct the acoustic emission feature vector. The original capacitance measurement tensor is composed of a three-dimensional stack of multi-frequency capacitance measurement value sequences according to electrode number, frequency number, and time series points. The decoupling generation module is used to input the original capacitance measurement tensor and acoustic emission feature vector into the capacitance tomography reconstruction algorithm, perform adaptive iterative reconstruction based on acoustic emission features, and output a spatiotemporally continuous dielectric constant distribution matrix; it calls the prior dielectric dispersion characteristics obtained from offline calibration experiments to construct a multi-frequency dielectric-mechanical coupled Gaussian mixture model; using the equiprobability surface of the equivalent dielectric constant probability distribution of each phase output by the model as the decision boundary, it replaces the fixed dielectric threshold with a frequency-energy related nonlinear dielectric classification surface, performs soft classification decoupling on the dielectric constant distribution matrix, and generates a shear band density field and porosity distribution field; Extraction and Combination Module: This module is used to call the dielectric constant-density mapping relationship established by the offline calibration experiment, convert the dielectric constant distribution matrix after soft classification decoupling into the powder apparent density distribution field; introduce the shear band density field to perform free volume correction on the powder apparent density distribution field, generating a free volume corrected density field; calculate the effective flow area field based on the porosity distribution field, and establish the Darcy flow model in combination with the free volume corrected density field to obtain the real-time flow characteristic scalar; Simultaneously, fractal dimension analysis was performed on the shear band density field to extract the shear band network connectivity; the real-time flow characteristic scalar, the shear band network fractal dimension, the shear band network connectivity, and the acoustic emission energy change rate in the acoustic emission characteristic vector were combined to construct the powder flow state characteristic vector; Judgment module: used to expand the powder flow state feature vector into time series diagram structure data, with the local state variables of each sensor sub-region as node features and the spatial adjacency relationship of the shear band region in the shear band density field as edge features; The time series graph structure data is input into a graph neural network-long short-term memory coupled model; the graph neural network extracts the spatial topological correlation features of the shear band through message passing and outputs a spatial embedding feature vector; the long short-term memory network performs temporal dynamic modeling on the spatial embedding feature vector and outputs a temporal prediction feature vector. The flow prediction sequence and instability risk probability are generated after decoding; Call the percolation critical connectivity threshold determined by the offline calibration experiment and compare the shear band network connectivity with the threshold in real time; If the threshold is exceeded, an instability warning signal will be generated; If the limit is not exceeded, the normal status indicator will be maintained. The perception and control module is used to construct a multi-objective cost function based on the flow prediction sequence, instability risk probability, and instability early warning signal. This multi-objective cost function includes a flow tracking accuracy term, an instability risk suppression term, a pressing energy consumption term, and an emergency safety protection term. A rolling time-domain optimization solution is performed on the multi-objective cost function to obtain the optimal control sequence, which in turn generates the powder supply rate adjustment, pressing pressure adjustment, and emergency pressure relief valve opening command. The powder supply rate adjustment is output to the powder supply driver, the pressing pressure adjustment is output to the servo pressing unit, and the emergency pressure relief valve opening command is output to the safety pressure relief valve, achieving coordinated control of high-precision flow tracking and active suppression of shear band instability during cold pressing.

Images