Powder compact density gradient detection method based on terahertz spectral imaging

Abstract

The invention relates to the technical field of powder compact density gradient detection, and particularly discloses a terahertz spectral imaging-based powder compact density gradient detection method, which comprises the following steps of: pre-establishing an environment interference compensation model and a signal-density correlation model, and transmitting multi-angle terahertz detection pulses to a compact preset point during detection, so as to detect the density gradient of a compact; synchronously applying high-frequency micro-amplitude mechanical vibration to mark a signal, and generating an original multi-dimensional feature packet in combination with real-time environmental parameters; then, environment interference is quantized by using a preset compensation lookup table, a pure signal feature set is constructed, and a three-dimensional density gradient map is generated through spatial interpolation; according to the invention, through an active marking and compensation mechanism, the stability of detection under complex working conditions is improved, a data processing link from original signal acquisition to density value output is established, and a three-dimensional density map is reconstructed by combining multi-angle detection with spatial interpolation. And an effective technical means is provided for quality control and process improvement of powder metallurgy manufacturing.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of powder compact density gradient detection, in particular to a powder compact density gradient detection method based on terahertz spectral imaging. BACKGROUND

[0002] The powder compact is a key intermediate product in the powder metallurgy process, and the uniformity of its internal density distribution directly determines the mechanical properties and service life of the final sintered part. In actual production process, due to the poor flowability of powder, uneven transmission of pressing force and mold wear and other factors, significant density gradient often forms inside the compact. This density non-uniformity will be amplified in the subsequent sintering process, resulting in cracks, deformation or even scrap of the product. Traditional quality control mainly relies on destructive sampling detection, but this method cannot achieve full detection and has a large detection blind area, making it difficult to find batch quality problems in time and causing a large number of unqualified products to flow into the subsequent process.

[0003] At present, the evaluation of powder compact density distribution in the industry mainly adopts X-ray CT scanning, ultrasonic detection or overall density measurement based on Archimedes principle. Although X-ray CT can obtain three-dimensional density information, the device is expensive, the detection speed is slow and there is a radiation safety problem, making it difficult to apply to online detection. The ultrasonic method is limited by the porous structure of the powder compact and the discontinuity of acoustic impedance, and the signal attenuation is serious and the resolution is low, making it difficult to accurately identify local density changes. The Archimedes method can only measure the overall average density and cannot reflect the internal density gradient information at all.

[0004] The common disadvantage of these traditional methods is that they cannot meet the requirements of rapid, accurate, three-dimensional and non-destructive detection at the same time. Although the X-ray method is accurate, the detection period is long and the cost is high, the ultrasonic method is non-destructive but has insufficient reliability in porous materials, and the simple mass density method completely ignores the internal structure information. More importantly, existing methods generally ignore the influence of environmental factors and equipment vibration on measurement results during detection, and have poor repeatability and stability in actual production environment, making it difficult to establish reliable quality judgment standards and meet the needs of real-time online quality monitoring in modern intelligent manufacturing. SUMMARY

[0005] In view of this, in order to solve the problems raised in the background art, a powder compact density gradient detection method based on terahertz spectral imaging is proposed.

[0006] The objective of this invention can be achieved through the following technical solution: This invention provides a method for detecting the density gradient of powder compacts based on terahertz spectral imaging. Before formal testing, this method first performs offline calibration to establish a basic database, specifically including: S1, establishing a compensation lookup table, by measuring a reference body with known physical properties under different combinations of environmental parameters and mechanical disturbances, establishing the correspondence between environmental interference and signal distortion modes to generate a compensation lookup table; S2, establishing a density feature association table, by measuring multiple standard samples with known densities, establishing the correspondence between signal features and density values ​​to generate a density feature association table.

[0007] After completing the above calibration, online testing includes the following steps:

[0008] S3. Generation of original multidimensional feature package: A set of multi-angle terahertz detection pulses are emitted to the preset scanning point of the powder compact, and a preset high-frequency micro-amplitude mechanical vibration is applied simultaneously to superimpose predictable disturbance features into the received echo signal. At the same time, real-time environmental parameters are collected, and the echo signal and real-time environmental parameters are packaged to generate the original multidimensional feature package.

[0009] S4. Construction of environmental interference feature vector: Based on the real-time environmental parameters and mechanical disturbance features in the original multidimensional feature package, query the compensation lookup table generated in step S1 to identify and quantify signal distortion, thereby constructing an environmental interference feature vector.

[0010] S5. Generating a pure signal feature set: Perform differential operations on the echo signal and environmental interference feature vector contained in the original multidimensional feature package, and extract features from the signal data after differential operations to generate a pure signal feature set that characterizes the true terahertz response of the billet.

[0011] S6. Pure signal stream generation: Based on the transmission angle sequence of the terahertz antenna array and the preset fixed geometric position relationship, the pure signal feature set is analyzed and assigned a unique position coordinate to generate a position-coded pure signal stream.

[0012] S7. Generation of Discrete Density Point Cloud: Match the signal features in the location-encoded clean signal stream with the density feature association table generated in step S2, and assign a density level value to each location-encoded signal point to generate a discrete density point cloud.

[0013] S8. Construction of 3D density gradient map: Spatial interpolation is performed on the 3D coordinates of all data points in the discrete density point cloud to calculate and fill the density prediction values ​​of unknown nodes in the 3D spatial grid in order to construct a 3D density gradient map.

[0014] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects: (1) The present invention actively applies predictable mechanical disturbances during terahertz detection and embeds the disturbance characteristics as signal markers into the original data. Combined with a pre-calibrated compensation lookup table, it can identify and compensate for environmental interference. This active marking and compensation mechanism can distinguish between the intrinsic response of materials and external disturbances, which helps to improve the stability and repeatability of measurement results under complex working conditions such as press vibration and temperature fluctuations.

[0015] (2) This invention establishes a processing link from the original signal to the density value. Through a dual calibration mechanism of compensation lookup table and density feature association table, the complex signal processing process is partially transformed into a lookup table matching operation. This design improves the efficiency and consistency of data processing and makes the measurement results traceable, providing data support for subsequent process optimization and quality traceability.

[0016] (3) This invention employs a strategy of multi-angle terahertz detection combined with spatial interpolation reconstruction, which can obtain the continuous three-dimensional density distribution inside the compact based on a limited number of actual measurement points. By expanding the discrete density point cloud into a three-dimensional gradient map, the spatial variation trend of density can be intuitively displayed, density uniformity can be quantitatively assessed, and feedback can be provided for the optimization of pressing process parameters. This data reconstruction method helps to shorten the data acquisition time while ensuring detection accuracy.

[0017] (4) This invention achieves modular integration of detection hardware, compensation algorithm, and visualization output. The evaluation system can be embedded in existing press equipment and is compatible with the production process. The modular design gives the system good scalability and maintainability. At the same time, the three-dimensional density gradient map output by the system can be correlated with the process parameters of the pressing process, which helps to achieve closed-loop control of quality inspection and process improvement. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the method steps of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figure 1 Before performing online detection, the method of this invention requires offline calibration to establish a basic database. S1. Compensation Lookup Table Establishment: This step aims to establish a database to compensate for the effects of environmental factors and actively applied mechanical disturbances on the signal during subsequent online detection. Specifically, a physically stable and known reference body, such as a high-density ceramic block, is selected and placed in an experimental environment that simulates different working conditions. Terahertz signals are systematically acquired by changing the ambient temperature, humidity, and background vibration level, and combining different active mechanical disturbance parameters, such as vibrations of different frequencies and amplitudes. The signals measured under each combination of conditions are compared with the reference signal under ideal conditions to extract the signal distortion modes, such as amplitude attenuation, phase drift, and time delay. These distortion modes are then combined with the corresponding "environmental parameters + mechanical disturbance" as key-value pairs and stored to form a compensation lookup table.

[0022] S2. Density Feature Correlation Table Establishment: This step aims to establish a mapping relationship between terahertz signal characteristics and the actual density values ​​of the material. Specifically, a series of standard samples are prepared, identical to the material to be tested but with different densities, and whose uniform density has been accurately measured using destructive methods (such as the Archimedes method). For each standard sample, steps S3 to S6, the same as those for subsequent online detection, are performed: acquiring its terahertz signal, performing environmental compensation, and extracting a pure signal feature set. Finally, the pure signal feature set corresponding to each standard sample, such as dielectric constant and absorption coefficient values, is mapped one-to-one with its known density value and stored in a database to form a density feature correlation table.

[0023] After completing the above calibration, perform online testing:

[0024] This invention provides a method for detecting the density gradient of powder compact based on terahertz spectral imaging, including: S3, generating the original multidimensional feature package: transmitting a set of multi-angle terahertz probe pulses to the preset scanning point of the powder compact, and simultaneously applying a preset high-frequency micro-amplitude mechanical vibration to superimpose predictable disturbance features into the received echo signal, while simultaneously acquiring real-time environmental parameters, and packaging the echo signal and real-time environmental parameters to generate the original multidimensional feature package.

[0025] In a specific embodiment of the present invention, the specific steps of packaging the echo signal and real-time environmental parameters to generate the original multi-dimensional feature package include: driving the terahertz antenna array to emit terahertz detection pulses at multiple angles toward a preset scanning point.

[0026] When transmitting a terahertz probe pulse, a miniature piezoelectric ceramic actuator rigidly coupled to the terahertz antenna array is activated to generate high-frequency micro-amplitude mechanical vibrations, thereby superimposing predictable disturbance characteristics onto the received echo signal.

[0027] Collect real-time environmental parameters provided by environmental reference sensors.

[0028] The echo signal containing predictable disturbance characteristics is packaged with real-time environmental parameters to generate the original multidimensional feature package.

[0029] It should be noted that the operation control system first activates the active probe module to perform the data acquisition task. This process begins by driving a terahertz antenna array integrated into a designated location inside the press. This terahertz antenna array aligns sequentially with a series of preset scanning points on the powder compact according to a pre-defined scanning path. For each preset scanning point, the terahertz antenna array emits a set of multi-angle terahertz detection pulses, for example, pulse signals with a center frequency of 0.5 terahertz emitted sequentially from angles of -15 degrees, 0 degrees, and +15 degrees. After the signal penetrates the compact, reflections occur at the interfaces of different density regions within it, and the antenna array synchronously receives these echo signals at the same emission location. At the precise moment of each terahertz detection pulse emission, the control system synchronously outputs a trigger signal to a miniature piezoelectric ceramic actuator that is rigidly physically connected to the terahertz antenna array. Once activated, this miniature piezoelectric ceramic actuator generates a preset high-frequency micro-amplitude mechanical vibration. This high-frequency micro-amplitude mechanical vibration is transmitted to the antenna through a rigid structure, thereby modulating and superimposing the known, predictable disturbance characteristics onto the received echo signal. Meanwhile, an environmental reference sensor integrated with the active probe module continuously operates, measuring and outputting real-time environmental parameters within the current compressor operating environment. Specific data includes the compressor's internal temperature, humidity, and background vibration data. Finally, the data processing unit integrates the received echo signal, already superimposed with predictable disturbance characteristics, with the real-time environmental parameters acquired at the same time. These two data sets are encapsulated into a unified data structure, forming a raw multidimensional feature package. This feature package fully records all the raw information of a single scan point under specific detection angles, specific mechanical disturbances, and specific environmental conditions, providing unprocessed input data for subsequent analysis.

[0030] The active probe module is a composite hardware unit integrating signal transmission, reception, vibration application, and environmental sensing functions. Its function is to actively probe the compact and collect multi-dimensional raw data. The terahertz antenna array is the core component of the active probe module, composed of multiple miniature terahertz transceiver antennas arranged in a specific geometric configuration, used to transmit and receive terahertz waves at different angles. The compact refers to an unsintered solid blank formed by pressing metal or other material powder in a mold. The preset scanning points are a series of coordinate points pre-planned on the compact's three-dimensional model to ensure that the detection path fully covers the critical areas of the compact. Their setting is based on the distribution of typical density non-uniform areas determined after destructive testing of several similar compacts. Multi-angle terahertz detection pulses refer to terahertz beams emitted from multiple different incident angles, used to probe the same area from different paths to obtain richer structural information. The echo signal is the signal reflected back when the terahertz detection pulse encounters interfaces with different dielectric constants inside the compact; its waveform and delay carry information about the compact's internal structure and density. A miniature piezoelectric ceramic actuator is a micro-actuator that converts electrical signals into precise mechanical displacements, used here to generate controlled micro-vibrations. The preset high-frequency micro-amplitude mechanical vibration refers to a specific vibration mode generated by the miniature piezoelectric ceramic actuator, with a frequency set at 50 kHz and an amplitude of 10 micrometers. These parameters are experimentally optimized to ensure that the vibration produces clearly identifiable characteristics in the echo signal without damaging the compact structure. Predictable disturbance characteristics are specific imprints produced by the preset high-frequency micro-amplitude mechanical vibration in the time or frequency domain of the echo signal. Since the vibration source is known, the pattern of this imprint is also determined. An environmental reference sensor is a collection of sensors used to measure physical quantities in the equipment's operating environment, providing environmental baseline data for signal compensation. Real-time environmental parameters are a set of data collected by the environmental reference sensor describing the current environmental conditions; the data structure is a vector containing specific numerical values. The press internal temperature refers to the temperature of the gas or solid near the mold of the pressing equipment. Humidity refers to the relative humidity of the air inside the press. Background vibration data refers to environmental vibration signals not preset by the press operation or other external factors.

[0031] For example, the operator activates the evaluation system, and the active probe module is excited. A terahertz antenna array located inside the press is aligned with the first preset scanning point on the compact, i.e., coordinates (X1, Y1, Z1), and emits a terahertz probe pulse from a negative 15-degree angle. Almost simultaneously, a miniature piezoelectric ceramic actuator rigidly coupled to the antenna array is activated, generating a preset high-frequency micro-amplitude mechanical vibration with a frequency of 50 kHz and an amplitude of 10 micrometers. The antenna array receives the echo signal, which is reflected and superimposed with predictable disturbance characteristics due to the internal structure of the compact. Simultaneously, an environmental reference sensor acquires the current real-time environmental parameters, specifically the press internal temperature of 28.5 degrees Celsius, humidity of 55%, and a set of acceleration sequences characterizing background vibration data. The data processing unit packages this echo signal data containing predictable disturbance characteristics, along with the real-time environmental parameters containing temperature, humidity, and vibration values, to generate a raw multidimensional feature package, marked as the data acquisition of point (X1, Y1, Z1) at a negative 15-degree angle. The system will then automatically repeat this process to complete the data acquisition at 0 degrees and positive 15 degrees for the scanning point, and perform the same operation on all preset scanning points in sequence.

[0032] In a specific embodiment of the present invention, after generating the original multidimensional feature package, the method further includes: comparing the predictable disturbance features contained in the echo signal with preset reference disturbance features to generate data quality assessment indicators.

[0033] It's important to note that, firstly, once the original multidimensional feature packet is successfully generated, the system doesn't immediately pass it to the next step. Instead, the system immediately performs a rapid "health check" on this newly acquired data packet. Next, the system extracts the predictable perturbation feature applied by the miniature piezoelectric ceramic actuator from the original echo signal contained in the feature packet. This predictable perturbation feature, because its frequency and amplitude are pre-set, should exhibit a known and well-defined pattern in the frequency or time domain of the signal. Then, the system compares this perturbation feature extracted from the actual echo signal with a reference perturbation feature pre-stored in the system. This reference perturbation feature is the gold-standard signal pattern recorded under ideal, interference-free laboratory conditions when the same mechanical vibration is applied to a standard reference object. The comparison process typically involves calculating their cross-correlation coefficient. This comparison process generates a quantified value, namely the data quality assessment index. This data quality assessment metric is a value between -1 and 1. The closer it is to 1, the higher the similarity, indicating that the actively applied perturbation labels have been clearly and completely recorded.

[0034] It should also be noted that, in order to obtain a standardized data quality assessment index that is independent of signal energy and ranges from -1 to 1, we typically use the cross-correlation coefficient. The formula for calculating the cross-correlation coefficient is: ,in, The reference disturbance characteristic signal is in integer form. For discrete sequences indexed, The disturbance characteristic signal extracted from the actual echo is in integer form. A discrete sequence indexed by; It is time shift; and These are signals and signal The average value.

[0035] When the data quality assessment index is lower than the preset quality threshold, a remeasurement of the preset scan points is triggered.

[0036] It should be noted that the system compares the generated data quality assessment index with a preset quality threshold. If the data quality assessment index is higher than or equal to the preset quality threshold, it indicates that the data acquisition quality is qualified, and the original multidimensional feature package is verified as valid. Conversely, if the data quality assessment index is lower than the preset quality threshold, it indicates that during this measurement, the actively applied disturbance markers may have been severely overwhelmed due to sudden strong external interference or momentary equipment failure. The system will determine that the acquired data is invalid, discard it, and immediately and automatically trigger a remeasurement at the same preset scan point until a qualified data packet is obtained.

[0037] The preset quality threshold is a pre-defined critical value used to determine whether the data quality assessment indicators meet the standards. This preset quality threshold is usually determined based on a large amount of experimental data to achieve a balance between ensuring data quality and avoiding unnecessary retesting. In a specific embodiment of the present invention, the cross-correlation coefficient threshold can be 0.9.

[0038] S4. Construction of environmental interference feature vector: Based on the real-time environmental parameters and mechanical disturbance features in the original multidimensional feature package, query the compensation lookup table generated in step S1 to identify and quantify signal distortion, thereby constructing an environmental interference feature vector.

[0039] In a specific embodiment of the present invention, the specific steps for constructing the environmental interference feature vector include: accessing the compensation lookup table generated in step S1, which records standard echo signal disturbance patterns under different combinations of environmental parameters and mechanical disturbance features.

[0040] The real-time environmental parameters and mechanical disturbance features in the original multidimensional feature package are used as a combined index for matching queries in the compensation lookup table.

[0041] Based on the standard echo signal disturbance pattern obtained from the matching query, the signal distortion caused by the current environment and mechanical vibration is quantified and constructed as an environmental interference feature vector.

[0042] It should be noted that, firstly, the operator accesses a pre-established and stored compensation lookup table. This compensation lookup table is a data structure that records in detail the disturbance patterns exhibited by the standard echo signal generated by a reference body with known physical characteristics under specific conditions. This record is generated after experimental calibration based on different combinations of environmental parameters and specific mechanical disturbances. Next, the system uses the information in the original multidimensional feature package generated in the previous step as an index to perform a matching query in the accessed compensation lookup table. Specifically, the "real-time environmental parameters" directly contained in the original multidimensional feature package, such as the collected temperature, humidity, and background vibration values, are used as the first type of index. At the same time, since the applied mechanical disturbance has preset parameters and is predictable, these preset parameters used to apply the mechanical disturbance, such as a vibration frequency of 50 kHz, are converted into a kind of "mechanical disturbance feature" data, serving as the second type of index. The system searches for the most matching or closest entry in the compensation lookup table based on these two types of index values. Based on the matching query results, the system can accurately identify and quantify the signal distortion caused by the combined effects of the current actual measurement environment and controlled mechanical vibration on the acquired echo signal. This identification and quantification process is derived from the comparison and analysis between the "standard echo signal disturbance pattern" found in the compensation lookup table and the "actual measurement data". Finally, the identified and quantified signal distortion is constructed in a structured data form. This constitutes an "environmental interference feature vector", which represents how much undesirable change exists in the signal caused by these external factors under the current specific environmental and mechanical disturbance conditions, and serves as the basis for subsequent data purification.

[0043] The compensation lookup table is a crucial data resource, storing the expected distortion patterns of the system response under different input conditions. When measurements are taken at specific scanning points on the compact, the "real-time environmental parameters" carried in the original multidimensional feature package are the internal temperature, humidity, and background vibration data of the press collected by sensors, set based on measured data within the standard industrial press operating range. The "mechanical disturbance characteristics" refer to the vibration parameters set when activating the micro-piezoelectric ceramic actuator to generate high-frequency micro-amplitude mechanical vibration, such as its fixed operating frequency of 50 kHz and amplitude of 10 micrometers. These parameters are set based on experimental optimization, aiming to generate significant but non-damaging echo signal disturbances to the compact. The standard echo signal disturbance pattern is an experimentally calibrated typical signal change shape or characteristic produced on a known reference object for a specific combination of environmental and mechanical disturbances; its data carrier is a series of numerical sequences representing signal amplitude or phase changes.

[0044] For example, the system receives a raw multidimensional feature packet containing echo signals acquired at angle A for a specific scanning point on the compact. This feature packet also records real-time environmental parameters: a temperature of 30 degrees Celsius, humidity of 60%, and background vibration data represented as a set of acceleration readings with an amplitude not exceeding 0.2g. Simultaneously, the system recognizes that the measurement utilized mechanical disturbances generated by a piezoelectric ceramic actuator with a frequency of 50 kHz and an amplitude of 10 μm. The system operator uses a pre-calibrated compensation lookup table, employing "temperature 30 degrees Celsius," "humidity 60%," "background vibration 0.2g," and "mechanical disturbance feature 50 kHz / 10 μm" as combined indices for a matching query. The entry in the compensation lookup table corresponding to this combined index indicates that this combination introduces a specific energy attenuation of approximately 15% and a system delay of approximately 10 microseconds in the echo signal. Based on this, the system determines that approximately 15% of the energy loss and 10 microsecond delay in the acquired echo signal are due to the current environmental and mechanical disturbances. The system integrates this loss magnitude and delay value into an environmental disturbance feature vector, for example, represented as [0.15, 1.0e-5], where the first component represents the energy loss ratio and the second component represents the time delay.

[0045] S5. Generating a pure signal feature set: Perform differential operations on the echo signal and environmental interference feature vector contained in the original multidimensional feature package, and extract features from the signal data after differential operations to generate a pure signal feature set that characterizes the true terahertz response of the billet.

[0046] In a specific embodiment of the present invention, the specific steps for generating a pure signal feature set characterizing the true terahertz response of the billet include: performing a difference operation by subtracting the numerical values ​​of the echo signal and the environmental interference feature vector to remove noise components, thereby generating signal data.

[0047] The generated signal data is analyzed to extract the dielectric constant and absorption coefficient, which reflect the physical properties of the pressed material. The dielectric constant and absorption coefficient are then combined to generate a pure signal feature set.

[0048] It should be noted that, firstly, the system accesses and reads the generated and encapsulated original multidimensional feature package. From this package, the original echo signal data is extracted. This data consists of the original measurement results received after the interaction between the terahertz probe pulse and the compact, containing both the material's true response and information about disturbances introduced from the outside. Next, the system performs a differential operation between the calculated and constructed environmental interference feature vector and the extracted original echo signal. The core of this operation lies in numerical subtraction, specifically subtracting the quantized noise component from the environmental interference feature vector point-by-point from the original echo signal value. This precise numerical calculation aims to remove interference components superimposed on the signal by environmental factors and pre-set mechanical disturbances, thus purifying the signal data. After the differential operation, a set of purified signal data is obtained. Then, the system performs feature extraction on this purified signal data. The purpose of this process is to identify and calculate parameters from the purified signal that directly reflect the inherent physical properties of the compact material. These parameters include, for example, the material's dielectric constant and absorption coefficient; these are intrinsic properties of the material and are unaffected by external environment and operational disturbances. Finally, all parameters reflecting the physical properties of the material obtained through feature extraction are compiled to form a complete dataset. This dataset is a pure signal feature set characterizing the true terahertz response of the compact at a specific scanning point. It excludes external interference and retains only the inherent information of the material.

[0049] The original echo signal data refers to the numerical representation of the original electromagnetic wave signal returned after the terahertz probe pulse encounters a reflecting surface inside the compact. The environmental interference feature vector is a vector whose value represents the quantified impact of preset mechanical vibration and environmental factors on the original echo signal; its data is based on the specific quantized distortion value obtained from a compensation lookup table. Differential operation is a mathematical subtraction operation used to subtract one set of values ​​from another. Numerical subtraction refers to performing subtraction calculations one by one. The noise component is the part of the original echo signal introduced by environmental factors and mechanical vibration, which is not a physical characteristic of the material itself. Feature extraction is the process of separating and calculating parameters that represent the essential properties of the material from the processed useful signal. The compact material refers to the powder-pressed compact material being evaluated. Physical properties refer to the inherent properties of the material itself, such as its ability to interact with electromagnetic waves. The dielectric constant is a physical quantity describing the polarization ability of a material under the action of an electric field; it affects the propagation speed and impedance of light waves. The absorption coefficient is a physical quantity describing the degree to which a material absorbs electromagnetic energy during propagation. A dataset is a structured container used to store related data. A clean signal feature set is a set of parameters, after denoising, that accurately characterizes the terahertz response of a compact at a specific scanning point, determined by its own material properties.

[0050] For example, the original echo signal data of a certain scan point is extracted from the output original multidimensional feature packet and denoted as a numerical sequence containing 1024 data points. Simultaneously, the output environmental interference feature vector is denoted as another numerical sequence of the same length (1024), representing the overall impact of the combination of environmental and mechanical disturbances on the signal. The system performs a differential operation, subtracting the corresponding value from the environmental interference feature vector from the value of each data point in the original echo signal sequence. For example, if the i-th data point of the original signal is `R_i` and the i-th value of the interference vector is `I_i`, then the purified signal data `C_i` is equal to `R_i - I_i`, resulting in a new 1024-point sequence. Subsequently, the system performs feature extraction on these 1024 purified signal data points, calculating that the dielectric constant of the material corresponding to the scan point is approximately 4.5 and the absorption coefficient is approximately 0.1 cm⁻¹. Finally, these two values ​​(4.5 and 0.1 cm⁻¹) are combined into a dataset, which is the pure signal feature set of that scan point.

[0051] S6. Pure signal stream generation: Based on the transmission angle sequence of the terahertz antenna array and the preset fixed geometric position relationship, the pure signal feature set is analyzed and assigned a unique position coordinate to generate a position-coded pure signal stream.

[0052] In a specific embodiment of the present invention, the specific steps of parsing the pure signal feature set and assigning unique position coordinates to generate a position-coded pure signal stream include: based on the transmission angle sequence of the terahertz antenna array and combined with the preset fixed geometric position relationship of the probe module relative to the center of the press mold, the unique position coordinates of each scanning point in the three-dimensional spatial coordinate system of the press blank are parsed in the calculation unit.

[0053] The clean signal feature set is bound to its corresponding unique location coordinates one by one to form a location-coded clean signal stream.

[0054] It should be noted that, firstly, in the computing unit, the system uses pre-collected data to determine the precise position of each scanning point in the three-dimensional spatial coordinate system (X,Y,Z) of the pressed billet. This determination process is based on two key pieces of information: first, the transmission angle sequence used by the terahertz antenna array when acquiring data, indicating the directions from which the signal is emitted to probe a specific area; and second, the fixed geometric positional relationship of the probe module relative to the center of the press die, established by pre-calibration work. This geometric relationship details the precise installation position, orientation, and offset of the probe module within the press, as well as its offset relative to the die center. By combining the known fixed position and orientation of the probe module, and the detection angle from the probe to a point on the pressed billet surface, the computing unit can automatically analyze and determine the unique position coordinates of the probe point in the three-dimensional spatial coordinate system (X,Y,Z) of the pressed billet itself, with the die center as the reference, much like performing triangulation. This process ensures that each acquired data point can be accurately located within the three-dimensional space of the pressed billet. Subsequently, the system correlates this with the generated clean signal feature set. Each clean signal feature set represents the actual physical property response of the compact at a specific scan point. The system then binds each of these clean signal feature sets, representing "what it is," to a unique location coordinate (X, Y, Z) determined earlier in the computational unit, representing "where it is." The final output is a data structure called a "location-encoded clean signal stream." This data stream contains dual information: both the three-dimensional spatial coordinates of a specific point within the compact and the parameters extracted from that point through cleansing, reflecting the material's true physical properties. This allows subsequent analysis to correlate material properties with their specific spatial location within the compact.

[0055] In this context, a computing unit refers to a hardware device used to perform data processing and computation tasks, such as a high-performance computer. A scanning point refers to a series of discrete positions on the surface of the press blank that require data acquisition and detection. The transmission angle sequence of the terahertz antenna array refers to the angular data record formed by the terahertz antenna array sequentially transmitting detection pulses from different angles when acquiring data at specific scanning points. The pre-calibrated fixed geometric position relationship of the probe module relative to the center of the press die is determined through precise measurements, including the fixed installation position and orientation of the probe module, as well as its offset and rotation relationship with the reference point of the press die. The three-dimensional spatial coordinate system (X, Y, Z) of the press blank refers to a three-dimensional rectangular coordinate system, whose origin is usually set at the geometric center of the press die, and the X, Y, and Z axes define the three-dimensional directions in space. A unique position coordinate refers to the (X, Y, Z) numerical combination that can uniquely identify a point in the aforementioned three-dimensional spatial coordinate system.

[0056] For example, for a generated pure signal feature set, it includes the dielectric constant of 3.2 and the absorption coefficient of 0.05 cm⁻¹ measured at the scanning point. At this time, the computing unit, based on the fact that the terahertz antenna array emitted probe pulses from three angles (-20 degrees, 0 degrees, and +20 degrees) when acquiring data at this point (i.e., the emission angle sequence), and knowing that the probe module's position is fixed at an upward offset of 0.5 meters (Z-axis direction) and a forward offset of 0.2 meters (X-axis direction) relative to the mold center (0,0,0), and parallel to the mold plane, uses this information to accurately calculate the unique three-dimensional position coordinates of this scanning point inside the compact (X=0.1 m, Y=0.05 m, Z=0.02 m). The system then binds the coordinates (0.1, 0.05, 0.02) to a set of clean signal features, namely a dielectric constant of 3.2 and an absorption coefficient of 0.05 cm⁻¹, forming a recording point, for example, represented as {X: 0.1, Y: 0.05, Z: 0.02, dielectric constant: 3.2, absorption coefficient: 0.05}. All such recording points are collected to form the position-encoded clean signal stream.

[0057] S7. Generation of Discrete Density Point Cloud: Match the signal features in the location-encoded clean signal stream with the density feature association table generated in step S2, and assign a density level value to each location-encoded signal point to generate a discrete density point cloud.

[0058] In a specific embodiment of the present invention, the specific steps for generating discrete density point clouds include: calling the density feature association table generated in step S2, which stores typical patterns of pure signal streams with position codes corresponding to standard density samples.

[0059] The signal characteristics of each point in the location-encoded clean signal stream are matched against typical patterns in the density feature association table.

[0060] The signal point encoded at each location is matched and assigned the closest density level value, thereby transforming the signal stream data into a set of data points containing three-dimensional coordinates and discrete density values, generating a discrete density point cloud.

[0061] It's important to note that, firstly, the system invokes a pre-established density feature association table stored in the database. The creation of this table is a crucial offline calibration process: operators prepare a series of standard density samples with known densities, whose densities are precisely measured using traditional, accurate destructive testing methods. Then, for each standard density sample, steps S3 to S6 are performed identically to those in the actual test: terahertz echo signals are acquired at specific locations, undergo the same compensation and purification processes, and the clean signal feature set is extracted and associated with its spatial location. Finally, each set of clean signal feature sets is mapped one-to-one with its known density value and stored in the table, forming an authoritative lookup dictionary of signal features to density values. Next, the system processes the location-encoded clean signal stream obtained in the previous step. For each data point in this signal stream, the system extracts its contained clean signal feature set. Then, the system uses this set of feature values ​​as query input and performs a rapid matching search in the density feature association table. This matching process aims to find which standard density sample's signal features, recorded in the table, are closest to or most similar to the signal features of the current measurement point. This is typically achieved by calculating the Euclidean distance in the feature space; the entry with the smallest distance is the best match. Finally, once the best-matching typical pattern is found for the current signal point, the system assigns the density level value corresponding to that pattern to the signal point. Since the signal point itself already carries unique three-dimensional coordinates (X, Y, Z), this assignment process completes the data transformation from {coordinates, signal features} to {coordinates, density value}. After this process has been performed on all signal points in the position-encoded clean signal stream, the original signal stream data has been completely transformed into a completely new set of data points. Each point in this set contains three-dimensional coordinates and discrete density values, thus generating a discrete density point cloud.

[0062] The density feature association table is a pre-calibrated database or lookup table. It stores a series of key-value pairs, where the "key" is the purified signal feature obtained after processing a standard density sample through steps S3 to S6, and the "value" is the known precise density of that standard density sample. The standard density sample is a series of reference blocks with the same physical and chemical composition as the test compact, but with different and precisely measured uniform densities through destructive testing. The typical pattern of the position-encoded purified signal stream refers to the standard signal feature stored in the association table, representing a specific density. It is a set of purified signal features measured and calculated on the standard density sample using the exact same procedure as the actual test. The signal feature is a parameter extracted from the purified terahertz signal, characterizing the physical properties of the material; in this invention, it mainly refers to the dielectric constant and absorption coefficient. The density level value is a discrete density value stored in the association table associated with a specific typical pattern.

[0063] For example, suppose there is a point in the position-encoded clean signal stream with the following data: {coordinates: (10mm, 15mm, 20mm), signal characteristics: (dielectric constant: 4.8, absorption coefficient: 0.12 cm⁻¹)}. The system then calls the density feature association table, which may contain records such as: {signal characteristics: (dielectric constant: 4.6, absorption coefficient: 0.10 cm⁻¹), density value: 7.2 g / cm³}, {signal characteristics: (dielectric constant: 4.9, absorption coefficient: 0.13 cm⁻¹), density value: 7.4 g / cm³}, {signal characteristics: (dielectric constant: 5.2, absorption coefficient: 0.17 cm⁻¹), density value: 7.6 g / cm³}. The system compares the signal characteristics (4.8, 0.12) of the point to be measured with the signal characteristics of each record in the table. Calculations revealed that (4.8, 0.12) is closest in feature space to (4.9, 0.13) in the second record. Therefore, the system determined that the test point best matches the second record. The system then retrieved the density value of 7.4 g / cm³ corresponding to the second record and assigned it to the test point. The original signal point data was transformed into new density point data: {coordinates: (10mm, 15mm, 20mm), density value: 7.4 g / cm³}. This process was repeated for all signal points, ultimately yielding tens of thousands of such density points, which together constitute a discrete density point cloud characterizing the density distribution within the entire compact.

[0064] S8. Construction of 3D density gradient map: Spatial interpolation is performed on the 3D coordinates of all data points in the discrete density point cloud to calculate and fill the density prediction values ​​of unknown nodes in the 3D spatial grid in order to construct a 3D density gradient map.

[0065] In a specific embodiment of the present invention, the specific steps of calculating and filling the density prediction values ​​of unknown nodes in the three-dimensional spatial grid to construct a three-dimensional density gradient map include: taking the three-dimensional coordinates of all data points in the discrete density point cloud as input to construct a three-dimensional spatial grid covering the entire compact volume.

[0066] It should be noted that, firstly, the system receives and processes the discrete density point cloud output from the previous step S5. This discrete density point cloud consists of a set of data points containing three-dimensional spatial coordinates (X, Y, Z) and corresponding discrete density values. These points represent the density information of known locations within the compact. Next, based on the three-dimensional coordinates (X, Y, Z) of all these discrete density points, the system automatically constructs a three-dimensional spatial grid covering the entire compact volume within the computer's computing unit. The grid settings include determining the grid boundaries, defined by the minimum and maximum X, Y, Z coordinates of the discrete point cloud (typically encompassing the entire compact area), and the density of the grid nodes, such as setting the grid node spacing to 1 mm. Each node in the grid represents a point with a known location in three-dimensional space.

[0067] Using a spatial interpolation algorithm, the density prediction value of each unknown node in the three-dimensional spatial grid is calculated and filled based on the density values ​​of the discrete density point cloud around the grid nodes of the three-dimensional spatial grid.

[0068] It should be noted that the system then performs spatial interpolation on each node in the constructed 3D spatial mesh. The purpose of this operation is to predict the density value at that node. During implementation, the system identifies known discrete density points within a certain radius around each mesh node. Based on the 3D coordinates and density values ​​of these known points, a spatial interpolation algorithm is used to calculate the predicted density value of the mesh node. The core principle of the interpolation algorithm is to perform a weighted average of the density values ​​of the known points based on the spatial distance or spatial correlation between the known points and the mesh nodes, thereby estimating the density at the mesh node. According to the specific implementation of the spatial interpolation algorithm, the formula can be expressed as: ,in, It is the grid node density value to be predicted; It is the first The values ​​of a known discrete density point; The first is determined based on spatial distance. The weighting factor for a known discrete density point is usually inversely proportional to the distance, such as... ,in It is the grid node and the first Euclidean distance between known discrete density points This is an adjustment parameter, typically set to 2. This formula calculates the estimated density at a given grid node by weighted summation of known point density values ​​and normalization of the weights.

[0069] The filled 3D spatial mesh is visualized and rendered as a color cloud map to construct a 3D density gradient map.

[0070] It should be noted that after the predicted density values ​​of all grid nodes have been calculated, the 3D spatial grid becomes a continuous density field. Finally, the system visualizes and renders this filled 3D spatial grid. Visualization can be achieved using color cloud maps, where different colors represent different density values ​​within the entire 3D internal structure of the compact. Through these visualization methods, the continuous trend of density variation within the compact from the center to the edges and from top to bottom is intuitively presented, thus constructing a directly interpretable 3D density gradient map.

[0071] Discrete density point cloud refers to the set of data points composed of three-dimensional coordinates (X, Y, Z) and corresponding discrete density values, representing the density information of a finite number of sampling points inside the compact. A three-dimensional spatial grid refers to dividing the space where the compact is located into a series of regularly arranged cubic or hexahedral units with fixed spacing; the center point or vertex of each unit can be considered a grid node. The density prediction value of a grid node is calculated using an interpolation algorithm based on the values ​​of surrounding known discrete density points, representing an estimate of the continuous density distribution of that grid node in space. Spatial interpolation is a mathematical method used to estimate the value of an unknown point based on the values ​​of a set of known data points, assuming that spatially close points have similar attribute values. The weighting factor is a coefficient used in the interpolation algorithm to determine the degree of influence of each known point on the unknown point; its magnitude typically depends on the spatial distance between the known and unknown points. Weighting parameters... This is an adjustment parameter in the inverse distance weighted interpolation method, affecting the degree to which distance influences the weight; it is usually set to 2. Visualization rendering refers to the conversion of a 3D data model into an image or other visual form that is recognizable to the human eye using computer graphics technology. A color cloud map is a 3D visualization technique that uses different colors to represent different ranges of data values. A 3D density gradient map is an image that visually displays the distribution and trend of density values ​​inside a compact as a function of spatial location.

[0072] For example, suppose a discrete density point cloud contains 1000 data points, each with an (X, Y, Z) coordinate and a density value. For instance, point 1 has coordinates of (0.01m, 0.02m, 0.03m) and a density of 7500 kg / m³; point 2 has coordinates of (0.015m, 0.02m, 0.03m) and a density of 7650 kg / m³. The system constructs a three-dimensional mesh, for example, with 100 nodes on each of the X, Y, and Z axes, for a total of 1 million mesh nodes, covering the entire compact volume. The system selects a mesh node located at (0.012m, 0.02m, 0.03m). Using the interpolation formula (p=2), the system calculates the distance from this node to all nearby known discrete points (point 1, point 2, etc.). For example, calculate the distance to point 1. Distance to point 2 Then calculate the weights. , . All known points corresponding to ( )and The values ​​are added together and normalized to obtain the predicted density value for that grid node, for example, 7580 kg / m³. The system performs this calculation for all 1 million grid nodes. Finally, these grid data filled with predicted density values ​​are used to generate a color cloud map using volume rendering technology. Areas with a density near 7500 kg / m³ are displayed in green, near 7600 kg / m³ in blue, and near 7700 kg / m³ in red, thus visually demonstrating the continuous variation in density within the compact.

[0073] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, and all such modifications and additions should fall within the protection scope of the present invention.

Claims

1. A method for detecting the density gradient of powder compacts based on terahertz spectral imaging, characterized in that, Includes the following steps: S1. Compensation lookup table establishment: By measuring a reference body with known physical characteristics under different combinations of environmental parameters and mechanical disturbances in advance, the correspondence between environmental interference and signal distortion modes is established to generate a compensation lookup table. S2. Density feature correlation table establishment: By measuring multiple standard samples with known densities in advance and processing them according to steps S3 to S6 below, a correspondence between the pure signal feature set and density values ​​is established to generate a density feature correlation table. S3. Generation of original multidimensional feature package: A set of multi-angle terahertz detection pulses are emitted to the preset scanning point of the powder compact, and a preset high-frequency micro-amplitude mechanical vibration is applied simultaneously to superimpose predictable disturbance features into the received echo signal. At the same time, real-time environmental parameters are collected, and the echo signal and real-time environmental parameters are packaged to generate the original multidimensional feature package. S4. Construction of environmental interference feature vector: Based on the real-time environmental parameters and mechanical disturbance characteristics in the original multidimensional feature package, query the compensation lookup table generated in step S1 to identify and quantify signal distortion, thereby constructing an environmental interference feature vector. S5. Pure signal feature set generation: The echo signal contained in the original multidimensional feature package is differentially processed with the environmental interference feature vector, and the signal data after differential processing is feature extracted to generate a pure signal feature set that characterizes the true terahertz response of the billet. S6. Pure signal stream generation: Based on the transmission angle sequence of the terahertz antenna array and the preset fixed geometric position relationship, the pure signal feature set is analyzed and assigned a unique position coordinate to generate a position-coded pure signal stream. S7. Generation of Discrete Density Point Cloud: Match the signal features in the location-encoded clean signal stream with the density feature association table generated in step S2, and assign a density level value to each location-encoded signal point to generate a discrete density point cloud. S8. Construction of 3D density gradient map: Spatial interpolation is performed on the 3D coordinates of all data points in the discrete density point cloud to calculate and fill the density prediction values ​​of unknown nodes in the 3D spatial grid in order to construct a 3D density gradient map.

2. The method for detecting the density gradient of powder compacts based on terahertz spectral imaging according to claim 1, characterized in that, The specific steps for packaging the echo signal and real-time environmental parameters to generate the original multi-dimensional feature package include: Drive the terahertz antenna array to transmit terahertz detection pulses at multiple angles toward a preset scanning point; When transmitting a terahertz detection pulse, a miniature piezoelectric ceramic actuator rigidly coupled to the terahertz antenna array is activated to generate high-frequency micro-amplitude mechanical vibrations, thereby superimposing predictable disturbance characteristics onto the received echo signal. Collect real-time environmental parameters provided by environmental reference sensors; The echo signal containing predictable disturbance characteristics is packaged with real-time environmental parameters to generate the original multidimensional feature package.

3. The method for detecting the density gradient of powder compacts based on terahertz spectral imaging according to claim 2, characterized in that, After generating the original multidimensional feature package, the following is also included: The predictable disturbance features contained in the echo signal are compared with the preset reference disturbance features to generate data quality assessment indicators. When the data quality assessment index is lower than the preset quality threshold, a remeasurement of the preset scan points is triggered.

4. The method for detecting the density gradient of powder compacts based on terahertz spectral imaging according to claim 1, characterized in that, The specific steps for constructing the environmental interference feature vector include: The access record contains a compensation lookup table generated in step S1 for standard echo signal disturbance patterns under different combinations of environmental parameters and mechanical disturbance characteristics. The real-time environmental parameters and mechanical disturbance features in the original multidimensional feature package are used as a combined index for matching and querying in the compensation lookup table. Based on the standard echo signal disturbance pattern obtained from the matching query, the signal distortion caused by the current environment and mechanical vibration is quantified and constructed as an environmental interference feature vector.

5. The method for detecting the density gradient of powder compacts based on terahertz spectral imaging according to claim 1, characterized in that, The specific steps for generating the pure signal feature set characterizing the true terahertz response of the compact include: A differential operation is performed by subtracting the numerical values ​​of the echo signal from the environmental interference feature vector to remove the noise component, thereby generating signal data; The generated signal data is analyzed to extract the dielectric constant and absorption coefficient, which reflect the physical properties of the pressed material. The dielectric constant and absorption coefficient are then combined to generate a pure signal feature set.

6. The method for detecting the density gradient of powder compacts based on terahertz spectral imaging according to claim 1, characterized in that, The specific steps for parsing the clean signal feature set and assigning unique position coordinates to generate a position-coded clean signal stream include: Based on the transmission angle sequence of the terahertz antenna array, and combined with the preset fixed geometric position relationship of the probe module relative to the center of the press mold, the unique position coordinates of each scanning point in the three-dimensional spatial coordinate system of the press blank are resolved in the calculation unit. The clean signal feature set is bound to its corresponding unique location coordinates one by one to form a location-coded clean signal stream.

7. The method for detecting the density gradient of powder compacts based on terahertz spectral imaging according to claim 1, characterized in that, The specific steps for generating discrete density point clouds include: The density feature association table generated in step S2 is invoked, which stores typical patterns of pure signal streams with position codes corresponding to standard density samples. Match and search typical patterns in the correlation table between the signal features and density features of each point in the location-encoded clean signal stream; The signal point encoded at each location is matched and assigned the closest density level value, thereby transforming the signal stream data into a set of data points containing three-dimensional coordinates and discrete density values, generating a discrete density point cloud.

8. The method for detecting the density gradient of powder compacts based on terahertz spectral imaging according to claim 1, characterized in that, The specific steps for calculating and filling the density prediction values ​​of unknown nodes in the three-dimensional spatial grid to construct a three-dimensional density gradient map include: Using the three-dimensional coordinates of all data points in the discrete density point cloud as input, a three-dimensional spatial mesh covering the entire compact volume is constructed. Using a spatial interpolation algorithm, the density prediction value of each unknown node in the three-dimensional spatial grid is calculated and filled based on the density values ​​of the discrete density point cloud around the grid nodes of the three-dimensional spatial grid. The filled 3D spatial mesh is visualized and rendered as a color cloud map to construct a 3D density gradient map.

Images