Method and system for automatically detecting boundary of color band of separation column based on image recognition

Abstract

The invention provides a separation column color band boundary automatic detection method and system based on image recognition, and relates to the technical field of data processing, and the method comprises the steps: inputting enhanced ultrasonic feature data into a pre-trained deep convolutional neural network model, the pre-trained deep convolutional neural network model comprises an input layer, a convolutional layer, a pooling layer, an up-sampling layer, a boundary prediction convolutional layer, an activation function layer, a boundary positioning layer and a smooth constraint unit, and obtaining a color band boundary feature response; the boundary positioning layer is used for carrying out extremum search on the color band boundary feature response to obtain boundary candidate points, the confidence coefficient is calculated based on the response intensity and neighborhood distribution features of the boundary candidate points, and then a color band boundary detection result is obtained through coordinate fitting, consistency correction of the smooth constraint unit and connection processing. According to the invention, the accuracy, the stability and the automation degree of color tape boundary identification are effectively improved.

Description

An Automatic Detection Method and System for Separation Column Color Band Boundaries Based on Image Recognition Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an automatic detection method and system for the boundary of a color band in a separation column based on image recognition. Background Technology

[0002] In practical applications, metal separation columns are often used for the separation and purification of target components in chemical raw materials. While metal columns (such as stainless steel and titanium alloy separation columns) have good sealing and structural strength, their inherent non-transparent nature makes it difficult for operators to directly observe the specific separation interface, band width, and dynamic migration process of the color band inside the column with the naked eye. This, to some extent, brings inconvenience to the real-time monitoring of the separation process and is not conducive to the accurate control of the separation process.

[0003] Traditional detection methods used to address the aforementioned observation challenges mainly fall into two categories, both of which have certain limitations: One category relies on pre-set, limited-sized viewing windows at both ends of the column, allowing operators to observe directly with the naked eye or with the aid of a simple magnifying glass. This method is not only limited by the size of the viewing window (usually only 1 / 5 to 1 / 4 of the column's cross-section) and the observation angle, but is also greatly affected by factors such as operator visual fatigue and subjective judgment differences. It can only roughly obtain the outline information of the color band at both ends of the column, and cannot effectively observe key information such as the color band separation state and boundary clarity in the middle section of the column (especially the core area of ​​1 / 2 to 2 / 3 of the column's length). The other category uses optical sensors for fixed-point detection. Optical sensors are usually deployed at both ends of the column and a few locations near the ends. The position of the color band is indirectly determined by the light signals received by the sensors. However, due to the blocking effect of the metal column on the light signals and the limited number of sensors, it may not be possible to achieve full coverage of the column's length, resulting in an inevitable obvious detection blind spot in the middle section of the column, making it difficult to accurately reflect key parameters such as the dynamic migration speed and boundary diffusion of the color band during the separation process. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide an automatic detection method and system for the boundary of the color band of the separation column based on image recognition, which effectively improves the accuracy, stability and automation of the color band boundary recognition.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] A first aspect is an automatic detection method for the boundary of a color band in a separation column based on image recognition, the method comprising:

[0007] Step 1: Emit ultrasonic waves to the separation column and receive the reflected signals to obtain the raw ultrasonic echo data; and set three ultrasonic reference nuclei at the center of the sample inlet end face, the center of the sample outlet end face, and the center of the cross section of the middle section of the column to obtain the spatial coordinate vector of the ultrasonic reference nuclei.

[0008] Step 2: Construct an ultrasonic spatial measurement field based on the spatial coordinate vectors of the three ultrasonic reference cores; perform acoustic field unit mosaicking on the ultrasonic spatial measurement field to obtain the acoustic field unit topology, which contains multiple triangular acoustic field units; obtain the coordinates of each vertex of the triangular acoustic field unit, accumulate the product of the x-coordinate of each vertex and the y-coordinate of the next vertex to obtain the first accumulated value, accumulate the product of the y-coordinate of each vertex and the x-coordinate of the next vertex to obtain the second accumulated value, and take half of the absolute value of the difference between the first accumulated value and the second accumulated value as the area of ​​the triangular acoustic field unit; perform deviation feature extraction and analysis on the area of ​​all triangular acoustic field units to generate a dynamic deviation correction operator;

[0009] Step 3: Perform ultrasonic feature enhancement processing based on the original ultrasonic echo data and dynamic deviation correction operator to obtain enhanced ultrasonic feature data;

[0010] Step 4: Input the enhanced ultrasound feature data into a pre-trained deep convolutional neural network model. The pre-trained deep convolutional neural network model includes an input layer, a convolutional layer, a pooling layer, an upsampling layer, a boundary prediction convolutional layer, an activation function layer, a boundary localization layer, and a smoothing constraint unit to obtain the color band boundary feature response.

[0011] Step 5: Use the boundary localization layer to perform extreme value search on the color band boundary feature response to obtain boundary candidate points. Calculate the confidence level based on the response intensity and neighborhood distribution characteristics of the boundary candidate points. Then, obtain the color band boundary detection result through coordinate fitting, consistency correction of smooth constraint units, and connection processing.

[0012] Secondly, an automatic detection system for the boundary of color bands in a separation column based on image recognition includes:

[0013] The acquisition module is used to transmit ultrasonic waves to the separation column and receive reflected signals to obtain raw ultrasonic echo data; and three ultrasonic reference nuclei are set at the center of the sample inlet end face, the center of the sample outlet end face, and the center of the cross section of the middle section of the column to obtain the spatial coordinate vector of the ultrasonic reference nuclei.

[0014] The deviation feature extraction module is used to construct an ultrasonic spatial measurement field based on the spatial coordinate vectors of three ultrasonic reference kernels; perform acoustic field unit mosaicking on the ultrasonic spatial measurement field to obtain the acoustic field unit topology, which contains multiple triangular acoustic field units; obtain the coordinates of each vertex of the triangular acoustic field unit, accumulate the product of the x-coordinate of each vertex and the y-coordinate of the next vertex to obtain the first accumulated value, accumulate the product of the y-coordinate of each vertex and the x-coordinate of the next vertex to obtain the second accumulated value, and take half the absolute value of the difference between the first accumulated value and the second accumulated value as the area of ​​the triangular acoustic field unit; perform deviation feature extraction analysis on the area of ​​all triangular acoustic field units to generate a dynamic deviation correction operator;

[0015] The ultrasonic feature processing module is used to perform ultrasonic feature enhancement processing based on the original ultrasonic echo data and dynamic deviation correction operator to obtain enhanced ultrasonic feature data.

[0016] The feature response acquisition module is used to input enhanced ultrasound feature data into a pre-trained deep convolutional neural network model. The pre-trained deep convolutional neural network model includes an input layer, a convolutional layer, a pooling layer, an upsampling layer, a boundary prediction convolutional layer, an activation function layer, a boundary localization layer, and a smoothing constraint unit to obtain the color band boundary feature response.

[0017] The color band boundary detection module is used to perform extreme value search on the color band boundary feature response using the boundary localization layer to obtain boundary candidate points. The confidence level is calculated based on the response intensity and neighborhood distribution characteristics of the boundary candidate points. Then, through coordinate fitting, consistency correction of smooth constraint units, and connection processing, the color band boundary detection result is obtained.

[0018] Thirdly, a computing device, comprising:

[0019] One or more processors;

[0020] A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0021] The above-described solution of the present invention has at least the following beneficial effects:

[0022] By performing sound field unit mosaicking on the ultrasonic spatial measurement field, calculating the area of ​​each triangular sound field unit and extracting deviation features, a dynamic deviation correction operator is generated. This process enhances the original ultrasonic echo data, effectively filtering out interference noise in the ultrasonic signal and improving the clarity and integrity of the ultrasonic feature data. Simultaneously, the enhanced ultrasonic feature data is input into a pre-trained deep convolutional neural network model. The boundary localization layer of the model automatically regresses to obtain the color band boundary detection result. The entire process eliminates the need for subjective judgment by operators, avoiding errors caused by visual fatigue and judgment differences in existing visual observation methods. It also overcomes the limitations of indirect judgment by fixed-point sensor detection, thus improving the accuracy of color band boundary detection. Attached Figure Description

[0023] Figure 1 is a schematic flowchart of an automatic detection method for the boundary of a separation column color band based on image recognition provided in an embodiment of the present invention.

[0024] Figure 2 is a schematic diagram of an automatic detection system for the boundary of a color band in a separation column based on image recognition, provided in an embodiment of the present invention. Detailed Implementation

[0025] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0026] As shown in Figure 1, an embodiment of the present invention proposes an automatic detection method for the boundary of a color band in a separation column based on image recognition. The method includes the following steps:

[0027] Step 1: Emit ultrasonic waves to the separation column and receive the reflected signals to obtain the raw ultrasonic echo data; and set three ultrasonic reference nuclei at the center of the sample inlet end face, the center of the sample outlet end face, and the center of the cross section of the middle section of the column to obtain the spatial coordinate vector of the ultrasonic reference nuclei.

[0028] Step 2: Construct an ultrasonic spatial measurement field based on the spatial coordinate vectors of the three ultrasonic reference cores; perform acoustic field unit mosaicking on the ultrasonic spatial measurement field to obtain the acoustic field unit topology, which contains multiple triangular acoustic field units; obtain the coordinates of each vertex of the triangular acoustic field unit, accumulate the product of the x-coordinate of each vertex and the y-coordinate of the next vertex to obtain the first accumulated value, accumulate the product of the y-coordinate of each vertex and the x-coordinate of the next vertex to obtain the second accumulated value, and take half of the absolute value of the difference between the first accumulated value and the second accumulated value as the area of ​​the triangular acoustic field unit; perform deviation feature extraction and analysis on the area of ​​all triangular acoustic field units to generate a dynamic deviation correction operator;

[0029] Step 3: Perform ultrasonic feature enhancement processing based on the original ultrasonic echo data and dynamic deviation correction operator to obtain enhanced ultrasonic feature data;

[0030] Step 4: Input the enhanced ultrasound feature data into a pre-trained deep convolutional neural network model. The pre-trained deep convolutional neural network model includes an input layer, a convolutional layer, a pooling layer, an upsampling layer, a boundary prediction convolutional layer, an activation function layer, a boundary localization layer, and a smoothing constraint unit to obtain the color band boundary feature response.

[0031] Step 5: Use the boundary localization layer to perform extreme value search on the color band boundary feature response to obtain boundary candidate points. Calculate the confidence level based on the response intensity and neighborhood distribution characteristics of the boundary candidate points. Then, obtain the color band boundary detection result through coordinate fitting, consistency correction of smooth constraint units, and connection processing.

[0032] In this embodiment of the invention, by performing sound field unit mosaicking on the ultrasonic spatial measurement field, calculating the area of ​​each triangular sound field unit and extracting deviation features, a dynamic deviation correction operator is generated to perform feature enhancement processing on the original ultrasonic echo data, effectively filtering out interference noise in the ultrasonic signal and improving the clarity and integrity of the ultrasonic feature data. At the same time, the enhanced ultrasonic feature data is input into a pre-trained deep convolutional neural network model, and the color band boundary detection result is obtained by automatic regression calculation through the boundary localization layer of the model. The entire process does not require subjective judgment by the operator, avoiding errors caused by visual fatigue, judgment differences and other factors in the existing naked eye observation method, and also overcoming the limitations of indirect judgment by fixed-point detection of sensors, thus improving the accuracy of color band boundary detection.

[0033] In a preferred embodiment of the present invention, step 1 involves transmitting ultrasonic waves to the separation column and receiving the reflected signals to obtain raw ultrasonic echo data; and setting three ultrasonic reference nuclei at the center of the sample inlet end face, the center of the sample outlet end face, and the center of the cross-section of the middle section of the column, respectively, and obtaining the spatial coordinate vector of the ultrasonic reference nuclei, which may include:

[0034] Step 101: Transmit ultrasonic pulses into the separation column and receive the reflected echo signals. Amplify, filter, and perform analog-to-digital conversion on the echo signals to obtain the original ultrasonic echo data. Specifically, for the problems of the metal separation column being opaque and having blind spots in optical detection, an ultrasonic transmitting device adapted to the metal medium is selected and fixed at a preset detection position outside the separation column to ensure that the transmitted ultrasonic pulses can effectively penetrate the metal column and propagate inside the column. After starting the ultrasonic transmitting device, the device continuously transmits ultrasonic pulses with a fixed frequency into the separation column. During the propagation of the ultrasonic pulses inside the separation column, they will encounter different medium interfaces such as the boundary of the color band inside the column, the end face of the column, and the ultrasonic reference core, thereby generating reflected echo signals. All the reflected echo signals are received in real time through a receiving device配套 with the ultrasonic transmitting device. Since the received echo signals have weak amplitudes and are mixed with environmental interference signals and clutter signals of the metal column itself and cannot be directly used for subsequent processing, a series of preprocessing of the received echo signals is required.

[0035] Use a dedicated instrumentation amplifier for ultrasonic signals as the signal amplification device to perform amplitude amplification on the echo signals. This instrumentation amplifier has the characteristics of high input impedance, low noise, and high common-mode rejection ratio, which can accurately amplify weak ultrasonic echo signals while suppressing environmental interference clutter, amplify the weak echo signals to a recognizable and processable range, and ensure that the subsequent processing can accurately capture the characteristics of the echo signals. Use an active low-pass filter to filter the amplified echo signals to filter out environmental interference clutter with frequencies higher than the ultrasonic pulse frequency and retain the effective echo signals related to the ultrasonic pulses. Complete the analog-to-digital conversion through a high-speed analog-to-digital converter. The high-speed analog-to-digital converter is an electronic device specifically used to convert analog signals into digital signals. Its core realizes the conversion through three steps: sampling, quantization, and encoding. The specific process is that the high-speed analog-to-digital converter samples the analog echo signals that have been amplified and filtered point by point according to the preset sampling frequency, captures the amplitude of the analog signal at each sampling moment, then converts the continuous amplitude signal obtained by sampling into discrete digital quantities, and converts the discrete digital quantities into binary digital signals through the encoding rule. During the conversion process, control the sampling frequency and quantization accuracy to ensure that the digital signal can completely retain the amplitude characteristics and time characteristics of the original analog echo signal, and finally obtain the original ultrasonic echo data.

[0036] Step 102: Based on the amplitude and time characteristics of echo signals at different locations in the original ultrasonic echo data, identify the reflected echoes corresponding to three ultrasonic reference nuclei, respectively, which are pre-set at the center of the sample inlet end face, the center of the sample outlet end face, and the center of the cross-section of the middle section of the column. Extract the echo signal segments corresponding to these three ultrasonic reference nuclei from the original ultrasonic echo data, and perform arrival time extraction and analysis on the three echo signal segments to obtain the arrival time information of the echo signal of each ultrasonic reference nuclei. Specifically, this includes: determining the preset positions of the three ultrasonic reference nuclei; fixing the three ultrasonic reference nuclei at the center of the sample inlet end face, the center of the sample outlet end face, and the center of the cross-section of the middle section of the column, respectively; and selecting materials with different ultrasonic reflection coefficients from the separation column material to ensure that their reflected echo signals can be distinguished from the reflected echo signals of the column itself and the color band; and performing segment-by-segment analysis on the original ultrasonic echo data to extract the arrival time information of each ultrasonic reference nuclei. The amplitude and time characteristics of the echo signal are analyzed. The time characteristic mainly refers to the propagation time of the echo signal from transmission to reception, while the amplitude characteristic refers to the peak value of the echo signal. Since the positions of the three ultrasonic reference nuclei are fixed, the propagation time of their reflected echo signals has a fixed range. Due to material differences, the amplitude of their reflected echo signals is significantly higher than the clutter signal of the column itself and the reflected echo signal of the color band. Therefore, based on the preset amplitude threshold and time range, echo signals that meet the conditions are screened from the original ultrasonic echo data, and then the reflected echoes corresponding to the three ultrasonic reference nuclei are identified. The preset amplitude threshold is 0.5V to 2.0V, and the time range is set according to the preset position of the ultrasonic reference nuclei and the ultrasonic propagation speed. The echo time range of the reference nucleus at the sample inlet is 10μs to 50μs, the echo time range of the reference nucleus at the sample outlet is 80μs to 120μs, and the echo time range of the reference nucleus in the middle section of the column is 40μs to 80μs.

[0037] After identification, based on the time range of the reflected echo signals from the three ultrasonic reference nuclei, echo signal segments corresponding to each ultrasonic reference nuclei are extracted from the original ultrasonic echo data to ensure that each segment contains only the reflected echo signal of the corresponding ultrasonic reference nuclei and is not mixed with other interference signals. Arrival time extraction and analysis are performed on each extracted echo signal segment. Starting from the moment when the ultrasonic transmitter emits a pulse, the amplitude change of the echo signal segment is monitored point by point. When the amplitude of the echo signal first reaches the preset peak threshold, the time value at that moment is recorded. This time value is the arrival time of the corresponding ultrasonic reference nuclei echo signal. The preset peak threshold is consistent with the amplitude threshold for screening ultrasonic reference nuclei echoes, which is 0.5V to 2.0V. Finally, the arrival time information of the echo signals of the three ultrasonic reference nuclei is obtained.

[0038] Step 103: Based on the arrival time information, calculate the arrival time difference between any two ultrasonic reference nuclei, and compare the arrival time difference with the fixed distance value corresponding to the pre-set geometric position relationship of the three ultrasonic reference nuclei on the separation column to calculate the actual propagation speed. Specifically, this includes: obtaining the arrival time information of the echo signals of the three ultrasonic reference nuclei obtained in step 102, and recording them as the arrival time of the first ultrasonic reference nuclei, the arrival time of the second ultrasonic reference nuclei, and the arrival time of the third ultrasonic reference nuclei, respectively. The first ultrasonic reference nuclei corresponds to the center of the sample inlet end face of the separation column, the second corresponds to the center of the sample outlet end face, and the third corresponds to the center of the cross-section of the middle section of the column; calculating the arrival time difference between any two ultrasonic reference nuclei, specifically by subtracting the arrival time of the other ultrasonic reference nuclei from the arrival time of one ultrasonic reference nuclei to obtain the arrival time difference between them. A total of three sets of arrival time differences are calculated, namely the arrival time difference between the first and second ultrasonic reference nuclei. The time difference between arrivals is calculated as follows: the time difference between arrivals between the first and third ultrasonic reference cores, and the time difference between arrivals between the second and third ultrasonic reference cores. Simultaneously, fixed distance values ​​corresponding to the geometrical positional relationships of the three ultrasonic reference cores on the separation column are pre-measured and set. These fixed distance values ​​refer to the straight-line distance between two ultrasonic reference cores, corresponding to three sets of fixed distance values, each corresponding to one of the three sets of time differences between arrivals. Since the distance the ultrasonic wave propagates in the separation column medium is equal to the propagation speed multiplied by the propagation time, and the fixed distance between two ultrasonic reference cores is equal to the propagation distance between them, with the propagation time being the time difference between arrivals, the propagation speed of the ultrasonic wave between the two sets of ultrasonic reference cores can be calculated by dividing each set of fixed distance values ​​by the corresponding time difference between arrivals. To improve the accuracy of the actual propagation speed, the three calculated propagation speeds are averaged, and the average value is taken as the actual propagation speed of the ultrasonic wave in the separation column medium.

[0039] Step 104: Based on the actual propagation speed and arrival time information, calculate the spatial distance from each ultrasonic reference core to the ultrasonic emission point. Then, combining this with the fixed spatial position of the ultrasonic emission point during the measurement process, calculate the spatial coordinate vectors of the three ultrasonic reference cores. Specifically, this includes: determining the fixed spatial position of the ultrasonic emission point during the measurement process; before the detection begins, fixing the ultrasonic emission device at a preset detection position; determining the three-dimensional spatial coordinates of the ultrasonic emission point using a spatial positioning device; and using these coordinates as a fixed reference, keeping them unchanged throughout the entire detection process; calculating the spatial distance from each ultrasonic reference core to the ultrasonic emission point based on the actual propagation speed and the arrival time information of the echo signal from each ultrasonic reference core. Specifically, the calculation method is to multiply the actual propagation speed by the arrival time of the echo signal from each ultrasonic reference core to obtain the straight-line distance from each ultrasonic reference core to the ultrasonic emission point. Since the total time for the ultrasonic wave to propagate from the emission point to the ultrasonic reference core and then reflect back to the receiving point is the arrival time, and the propagation distance is twice the distance from the emission point to the ultrasonic reference core, in actual calculation, the actual propagation speed needs to be multiplied by the arrival time and then divided by two to obtain the true spatial distance from each ultrasonic reference core to the ultrasonic emission point.

[0040] After calculating the spatial distances from the three ultrasonic reference nuclei to the ultrasonic emission point, and combining this with the fixed three-dimensional spatial coordinates of the ultrasonic emission point, a spatial geometric calculation method was used to determine the three-dimensional spatial coordinates of each ultrasonic reference nuclei. A three-dimensional rectangular coordinate system was established with the ultrasonic emission point as the origin. The X-axis was set along the axial direction of the separation column, the Y-axis along the radial direction of the separation column (horizontal), and the Z-axis along the radial direction of the separation column (vertical). Based on the preset orientations of the three ultrasonic reference nuclei, the azimuth and elevation angles of each reference nuclei in the three-dimensional coordinate system were determined. Specifically, the azimuth angle and elevation angle of the reference nuclei at the center of the sample inlet face were set to 0°, and the azimuth angle and elevation angle of the reference nuclei at the center of the sample outlet face were set to 0° and along the positive X-axis. The azimuth angle of the reference nuclei at the center of the cross-section of the middle section of the column was set to... The X-axis coordinates of each ultrasonic reference core are calculated using trigonometric functions, with the elevation angle set to 0° and the reference angle set to 0°, and the reference angle located at the midpoint of the X-axis of the reference core at the sample inlet and outlet faces. The Y-axis coordinate is calculated using the spatial distance from each ultrasonic reference core to the emission point. The X-axis coordinate is calculated as the spatial distance multiplied by the cosine of the azimuth angle and then by the cosine of the elevation angle. The Z-axis coordinate is calculated as the spatial distance multiplied by the sine of the elevation angle and then by the sine of the elevation angle. Substituting the specific values ​​of the azimuth angle, elevation angle, and spatial distance, the X-axis, Y-axis, and Z-axis coordinates of each ultrasonic reference core can be calculated. These three coordinate values ​​together constitute the spatial coordinate vector of the ultrasonic reference core. Through the above process, the spatial coordinate vectors of the three ultrasonic reference cores are calculated respectively.

[0041] This embodiment effectively solves the observation problem caused by the opacity of the metal separation column. By using ultrasound to penetrate the metal column and collect echo data, combined with the accurate positioning of the ultrasound reference core, it achieves full coverage of ultrasound data along the entire length and cross-section of the separation column. It can comprehensively capture the ultrasound characteristics of different regions inside the column, especially the characteristic information of the core region in the middle section of the column.

[0042] In a preferred embodiment of the present invention, step 2 involves constructing an ultrasonic spatial measurement field based on the spatial coordinate vectors of three ultrasonic reference cores; performing acoustic field unit mosaicking on the ultrasonic spatial measurement field to obtain an acoustic field unit topology, which includes multiple triangular acoustic field units; obtaining the coordinates of each vertex of the triangular acoustic field unit; accumulating the product of the x-coordinate of each vertex and the y-coordinate of the next vertex to obtain a first accumulated value; accumulating the product of the y-coordinate of each vertex and the x-coordinate of the next vertex to obtain a second accumulated value; taking half the absolute value of the difference between the first accumulated value and the second accumulated value as the area of ​​the triangular acoustic field unit; and performing deviation feature extraction and analysis on the areas of all triangular acoustic field units to generate a dynamic deviation correction operator, which may include:

[0043] Step 201: Based on the spatial coordinate vectors of the three ultrasonic reference cores and the original ultrasonic echo data, analyze the residual stress distribution within the separation column material to obtain residual stress distribution data. Specifically, this includes: clarifying that residual stress is generated during the manufacturing process of the metal separation column. This residual stress affects the propagation speed of ultrasonic waves within the column, leading to deviations in subsequent band boundary detection. Therefore, it is necessary to analyze the residual stress distribution. Combining the spatial coordinate vectors of the three ultrasonic reference cores obtained in Step 104, determine the specific spatial positions of the three reference cores within the separation column, using these as reference points for residual stress analysis. First, measure the standard values ​​of residual stress at the locations of the three reference cores using a stress detection instrument, serving as reference benchmarks for subsequent point-by-point analysis. Then, retrieve the original ultrasonic echo data and extract echo signal segments related to the three ultrasonic reference cores, as well as echo signal segments from different regions of the column. This is because ultrasonic waves propagate through media containing residual stress. When ultrasound propagates through a medium, its propagation speed changes, and the magnitude and direction of the residual stress affect the amount of change in propagation speed. This change is reflected in the temporal and amplitude characteristics of the echo signal. Specifically, the relationship between residual stress in metallic materials and ultrasonic propagation speed is as follows: tensile stress reduces ultrasonic propagation speed, while compressive stress increases it. The larger the absolute value of the stress, the greater the change in propagation speed. Under the same stress value, the change in ultrasonic speed along the stress direction is greater than the change in speed perpendicular to the stress direction. Based on this relationship, by analyzing the temporal and amplitude characteristics of echo signals in different regions, the actual propagation speed of ultrasound in each region is calculated based on the arrival time of the echo signal and the spatial distance from the corresponding region to the ultrasonic emission point. Then, the difference between the propagation speed of ultrasound at the location of the three ultrasonic reference cores and the propagation speed in other regions is compared, thereby analyzing the residual stress distribution state inside the separation column material point by point.

[0044] The specific process of point-by-point analysis is as follows: the internal space of the separation column is divided into several analysis points according to a preset step size. The preset step size is consistent with the node spacing of the sound field unit, ranging from 0.5mm to 2.0mm, to ensure that the spatial resolution of each analysis point meets the subsequent detection accuracy requirements. For each analysis point, its corresponding echo signal segment is extracted. Based on the arrival time of the echo signal segment and the preset spatial distance from the analysis point to the ultrasonic emission point, the actual ultrasonic wave propagation speed at the analysis point is calculated. The calculation formula is as follows: Where v is the actual propagation speed of the ultrasonic wave at the resolution point, L is the spatial distance from the resolution point to the ultrasonic emission point, and t is the arrival time of the echo signal; linear interpolation calculation is performed based on the known conditions of the three ultrasonic reference cores. Specifically, three sets of known data corresponding to the three ultrasonic reference cores are determined, and the spatial coordinates of the first ultrasonic reference core are recorded as follows: The measured standard value of residual stress is The corresponding ultrasonic propagation speed is The spatial coordinates of the second ultrasonic reference core are: The measured standard value of residual stress is The corresponding ultrasonic propagation speed is The spatial coordinates of the third ultrasonic reference core are: The measured standard value of residual stress is The corresponding ultrasonic propagation speed is Establish a correspondence between ultrasonic propagation speed as the independent variable and residual stress as the dependent variable, and clarify the velocity and stress mapping benchmarks for each reference core position.

[0045] For each resolution point, determine its spatial relative position to the three ultrasound reference nuclei, and calculate the distance from the resolution point to the first ultrasound reference nuclei. The distance is To the second ultrasound reference nucleus The distance is To the third ultrasound reference core The distance is Simultaneously calculate the pairwise distances between the three reference kernels, i.e. and The distance between them is , and The distance between them is , and The distance between them is The weight coefficients of the analytical point relative to the three benchmark kernels are calculated based on distance weights. The weight coefficients all take values ​​in the range (0,1) and satisfy the following conditions: The specific calculation formula is as follows: ; ; Based on the weighting coefficients, linear interpolation is performed to obtain the preliminary calculated value of the residual stress at the analytical point. The calculation formula is as follows: This formula correlates the standard values ​​of residual stress from the three reference cores with the spatial location of the analysis point, enabling a preliminary estimation of the residual stress. Further, the distribution direction of the residual stress is determined by considering the differences in ultrasonic velocity in different propagation directions at this analysis point. Specifically, at this analysis point, echo signal segments along the axial, radial horizontal, and radial vertical directions of the separation column are selected, and the actual ultrasonic propagation velocities in these three directions are calculated, denoted as... , , ;like and Then determine that the direction of residual stress distribution is along the axis of the separation column; if and If the residual stress distribution direction is determined to be radially horizontal; if and If the velocity values ​​in two directions are similar and both are greater than those in the third direction, then the residual stress is determined to be in a composite direction, and its distribution direction is biased towards the area between the two directions with larger velocity values.

[0046] During the analysis process, the residual stress at the locations of the three ultrasonic reference cores was used as a reference benchmark. The calculation results for each analysis point were calibrated to correct interpolation errors. Specifically, the calibration process involved comparing the preliminary calculated value of the residual stress at the analysis point obtained through linear interpolation with the standard values ​​of the residual stress at the three reference cores. Then, based on the distance from the analytical point to the corresponding benchmark kernel, the weight coefficients are adjusted for secondary interpolation calculation until the condition is met. To ensure analytical accuracy, the above point-by-point analytical process ultimately yields residual stress distribution data covering the entire separation column (i.e., the actual residual stress values ​​at each point). ).

[0047] Step 202: Based on the residual stress distribution data and the actual propagation velocity, the propagation velocity of the ultrasonic wave in the separation column medium is compensated and corrected point by point to obtain the corrected propagation velocity distribution data; based on the corrected propagation velocity distribution data and the spatial coordinate vectors of the three ultrasonic reference kernels, an ultrasonic spatial measurement field is constructed, specifically including: acquiring residual stress distribution data (points...). The actual propagation speed obtained in step 103 and step 104 clarifies the influence of residual stress on ultrasonic propagation speed. Specifically, the greater the residual stress, the greater the deviation in ultrasonic propagation speed, and the direction of influence of residual stress on propagation speed differs depending on the direction. The quantitative relationship between residual stress and ultrasonic propagation speed in metallic materials is based on the elastic mechanical properties of metallic materials and the ultrasonic propagation mechanism. Specifically, when ultrasonic waves propagate in a metallic medium, the propagation speed is positively correlated with the elastic modulus within the medium. Residual stress directly changes the elastic modulus of the metallic material; tensile stress decreases the elastic modulus, while compressive stress increases it. Changes in the elastic modulus directly lead to corresponding changes in ultrasonic propagation speed, exhibiting a linear quantitative correlation. This quantitative relationship can be quantified through specific calculation formulas. For the metallic material used in the separation column, the stress-sound velocity correlation coefficient is obtained beforehand through calibration. This coefficient is only related to the metal material and is a fixed value; for any point inside the separation column to be corrected, its residual stress (actual value) The quantitative relationship between the speed of sound and the propagation speed of ultrasound is that when there is tensile stress at that point ( When the value is positive, the deviation of the ultrasonic propagation speed It is a negative value, that is The propagation speed decreases; when there is compressive stress at that point ( When the value is negative, the deviation of the ultrasonic propagation speed It is a positive value, that is The propagation speed increases; the absolute value of the deviation is proportional to the absolute value of the actual residual stress, that is... ( (This represents the actual value of residual stress, with tensile stress being positive and compressive stress being negative). Simultaneously, this quantitative relationship needs to consider the influence of stress direction, introducing a direction influence coefficient. , The cosine of the angle between the ultrasonic wave propagation direction and the principal direction of residual stress, ranging from 0 to 1, is used to calculate the final propagation velocity deviation. = , The closer the value is to 1, the greater the influence of stress on the propagation speed. When the stress is 0, the stress has no effect on the propagation speed.

[0048] Based on the spatial location within the separation column, the ultrasonic wave propagation velocity at each spatial point is compensated and corrected point-by-point. Specifically, the correction process involves extracting the actual residual stress value at that spatial point from the residual stress distribution data. (Tensile stress is positive, compressive stress is negative), determine the angle between the ultrasonic wave propagation direction and the principal direction of the residual stress at that point, and calculate the directional influence coefficient. Substituting the values ​​into the quantitative formula above, the deviation of the ultrasonic wave propagation velocity at that point can be calculated. Finally, use the actual propagation speed at that spatial point. Add the calculated deviation value The corrected propagation speed at that spatial point is obtained. ,Right now To ensure that the propagation velocity at each spatial point can eliminate the influence of residual stress and accurately reflect the propagation state of ultrasound at that point, the corrected propagation velocities of all spatial points are integrated after point-by-point correction to obtain corrected propagation velocity distribution data covering the entire separation column. Using the spatial coordinate vectors of the three ultrasonic reference cores obtained in step 104 as the core reference points, and combining the corrected propagation velocity distribution data, an ultrasonic spatial measurement field is constructed. During the construction process, the spatial coordinate vectors of the three ultrasonic reference cores are used as the reference coordinates of the measurement field, and the corrected propagation velocity distribution data is used as the core parameters of the measurement field. The spatial position inside the separation column is correlated with the corresponding ultrasonic propagation velocity, forming a spatial measurement system that can accurately reflect the propagation characteristics of ultrasound inside the separation column.

[0049] Step 203: Using the spatial coordinate vector locations of the three ultrasonic reference cores as initial mosaicking nodes, auxiliary mosaicking nodes are generated between adjacent ultrasonic reference cores and between the ultrasonic reference cores and the separation column boundary. The initial mosaicking nodes and auxiliary mosaicking nodes are connected to obtain a grid topology covering the entire ultrasonic spatial measurement field. The grid topology contains multiple triangular acoustic field units, and the set of all triangular acoustic field units constitutes the acoustic field unit topology. Specifically, this includes: determining the spatial coordinate vector locations of the three ultrasonic reference cores and using these three locations as initial mosaicking nodes for acoustic field unit mosaicking; considering the complexity of the internal space of the separation column and the integrity of the ultrasonic spatial measurement field, auxiliary mosaicking nodes need to be generated between adjacent ultrasonic reference cores and between the ultrasonic reference cores and the separation column boundary; when generating auxiliary mosaicking nodes, the measurements of two adjacent initial mosaicking nodes are taken. and Distance between Following the principle of uniform distribution, several auxiliary tessellation nodes are set between two initial tessellation nodes to ensure that the distance between adjacent nodes meets the accuracy requirements of the sound field unit resolution. The preset threshold for the distance between adjacent nodes is 0.5mm to 2.0mm. Specifically, if two adjacent initial tessellation nodes... and The distance between them is Lmn, and the preset node spacing is... (Values ​​range from 0.5mm to 2.0mm), then the number of auxiliary tessellation nodes required is... for: The spatial coordinates of each auxiliary mosaic node are calculated using linear interpolation. For example, for the e-th auxiliary mosaic node (1≤e≤...),... ), its coordinates ( )for: ; ; Simultaneously, between each initial mosaic node and the boundary of the separation column, auxiliary mosaic nodes are uniformly set according to the geometry of the separation column to ensure that the auxiliary mosaic nodes can fit the boundary of the separation column and achieve full coverage of the column boundary area. After the auxiliary mosaic nodes are generated, the initial mosaic nodes and auxiliary mosaic nodes, as well as adjacent auxiliary mosaic nodes, are connected sequentially by straight lines to form an interconnected grid structure. This grid structure can completely cover the entire ultrasonic spatial measurement field. During the connection process, it is ensured that each connecting line segment is within the range of the ultrasonic spatial measurement field, and there is no intersection or overlap between adjacent line segments. The final grid topology structure consists of multiple triangles, each triangle being a sound field unit. The collection of all triangular sound field units constitutes the sound field unit topology structure.

[0050] Step 204: Based on the sound field unit topology, obtain the coordinates of each vertex of each triangular sound field unit. Accumulate the products of the first vertex's x-coordinate and the second vertex's y-coordinate, the second vertex's x-coordinate and the third vertex's y-coordinate, and the third vertex's x-coordinate and the first vertex's y-coordinate to obtain a first accumulated value. Then accumulate the products of the first vertex's y-coordinate and the second vertex's x-coordinate, the second vertex's y-coordinate and the third vertex's x-coordinate, and the third vertex's y-coordinate and the first vertex's x-coordinate to obtain a second accumulated value. Take half the absolute value of the difference between the first and second accumulated values ​​as the area of ​​the triangular sound field unit. Specifically, this includes: extracting the vertex information of each triangular sound field unit from the sound field unit topology obtained in step 203, clarifying the number of vertices and the spatial coordinates of each vertex; since the sound field unit is triangular, for each triangular sound field unit, record the coordinates of its three vertices sequentially, denoted as the coordinates of the first vertex. The coordinates of the second vertex and the coordinates of the third vertex To calculate the area of ​​the triangle, we choose the xy plane for projection calculation. The specific process is as follows:

[0051] Calculate the first accumulated value :

[0052] Calculate the second accumulated value :

[0053] Calculate the absolute value of the difference between the first and second accumulated values, and take half of it to obtain the projected area of ​​the triangular sound field unit on the xy plane. : ;Will The area of ​​the triangular sound field unit is used as the area of ​​the triangular sound field unit. Following the above process, the area of ​​each triangular sound field unit in the sound field unit topology is calculated one by one.

[0054] Step 205: Compare the area of ​​each triangular sound field unit with a pre-set standard unit area threshold to filter out abnormal sound field units whose area deviation exceeds the preset threshold; perform statistical analysis on the spatial distribution and deviation of the abnormal sound field units to generate a dynamic deviation correction operator, specifically including: based on the sound field unit generation principle in step 203, combined with the geometric dimensions of the separation column and the requirements of the ultrasonic spatial measurement field, pre-setting the standard unit area threshold. This threshold represents the standard area that a triangular sound field unit should have under normal circumstances. The standard unit area threshold is preset to 0.25mm² to 4.0mm²; a preset threshold for area deviation is also set. This is used to determine if there are any abnormalities in the sound field unit. The preset threshold for area deviation is ±10% to ±20% of the standard unit area threshold, i.e.: in, This is the deviation scaling factor, ranging from 0.1 to 0.2; the area of ​​each triangular sound field unit calculated in step 204 is... Each is compared with a pre-set standard unit area threshold. By comparing the areas of each sound field unit with the area threshold of the standard unit, the area deviation of that sound field unit is calculated. The calculation formula is: ; the area deviation of each sound field unit Deviation threshold from preset To make a comparison, if If an abnormal sound field unit is identified, it indicates an abnormal ultrasonic wave propagation in the corresponding separation column space region, possibly due to insufficient residual stress compensation or node embedding deviation. After identifying all abnormal sound field units, statistical analysis is performed on them, recording the spatial distribution of each unit to pinpoint the exact location of the abnormal area within the separation column and determining whether the abnormal area is concentrated in blind spots such as the middle section of the column. The deviation of each abnormal sound field unit is then calculated. Analyze the distribution pattern of deviations and classify them into levels according to their magnitude (minor deviation: 0.1 < ≤0.15; Moderate deviation: 0.15 < ≤0.2; severe deviation: >0.2), to determine the severity of the deviation; based on the statistical results of the spatial distribution location and deviation of the abnormal sound field units, combined with the ultrasonic propagation characteristics and residual stress distribution data, a dynamic deviation correction operator is generated; this operator is a structured data carrier, the core of which is to assign a unique deviation correction parameter to each sound field unit (including normal and abnormal). The specific generation logic is as follows: for normal sound field units (area deviation does not exceed the preset threshold), the corresponding deviation correction parameters are set by default as follows: amplitude compensation coefficient = 1.0 (no amplitude deviation compensation), time delay correction value = 0μs (no time delay correction), and the correction validity is marked as normal; for abnormal sound field units, the corresponding deviation correction parameters are determined based on their own deviation, spatial location, and residual stress distribution data. The specific calculation formula is: amplitude compensation coefficient = 1 + (deviation amount × Time delay correction value = deviation × ×1μs (where , This is a proportional coefficient used to adjust the correction amplitude. The value range is 0.8 to 1.2. The value ranges from 0.5 to 1.5; the larger the deviation, the more concentrated the residual stress. , The larger the value, the stronger the correction amplitude); the final generated dynamic deviation correction operator contains the core content of each sound field unit, such as the unit number, amplitude compensation coefficient, time delay correction value, correction validity mark, anomaly level (only for abnormal units), and corresponding residual stress reference value, and is stored in the form of a structured file.

[0055] This embodiment further reduces data deviation by dividing the sound field into units and correcting abnormal units, ensuring that the final enhanced ultrasonic feature data can truly and accurately reflect the internal condition of the column.

[0056] In a preferred embodiment of the present invention, step 3, performing ultrasonic feature enhancement processing based on the original ultrasonic echo data and the dynamic deviation correction operator to obtain enhanced ultrasonic feature data, may include:

[0057] Step 301: Based on the sound field unit topology, the original ultrasonic echo data is divided into various triangular sound field units according to their spatial position correspondence, obtaining the unit echo data corresponding to each triangular sound field unit. Specifically, this includes: reading the sound field unit topology output in step 203, parsing the topology information of each triangular sound field unit, including the unit number, the number of vertices, and the three-dimensional coordinates of each vertex, and constructing the spatial boundary equation of each sound field unit using the vertex coordinates: For the triangular sound field unit, a spatial plane equation aX+bY+cZ+ρ=0 is generated based on the three-point coordinates, and the coefficients a, b, and c of this equation are... The coordinates 'c' are calculated from the vertex coordinates. The plane constant 'ρ', combined with the vertex coordinate range, determines the spatial envelope range of the unit within the separation column (axial lower limit to upper limit, radial inner to outer side). This forms the spatial index information for each acoustic field unit, including the unit number, spatial boundary parameters, vertex coordinate set, and other core information. The original ultrasonic echo data is retrieved; this data is stored in binary format, and each echo data record contains a unique identifier (ID), echo amplitude sequence, emission time, arrival time, and ultrasonic transmitter probe coordinates, among other key parameters. For each echo data, its corresponding reflection spatial position is calculated. The target is the actual propagation rate of the echo data (taken from the actual propagation velocity distribution data in step 103, matched with the mean path from the transmitting probe to the reflection point); the total propagation time of the ultrasonic wave is calculated, which is the difference between the arrival time and the transmission time. Combined with the actual propagation rate, the straight-line distance from the transmitting probe to the reflection point is obtained. This distance is the total propagation time multiplied by the actual propagation rate and then divided by 2, which is the one-way propagation distance; based on the spatial coordinates of the ultrasonic transmitting probe and the ultrasonic wave propagation direction angle (horizontal azimuth angle, vertical elevation angle), the three-dimensional coordinates of the reflection point are obtained through spatial geometric calculation. The calculation logic is based on the coordinates of the transmitting probe. The system calculates the coordinates of the reflection point on the three coordinate axes by combining the one-way propagation distance and the propagation direction angle. After calculating the coordinates of the reflection point, it iterates through the spatial index information of all sound field units to determine whether the reflection point belongs to the spatial range of a certain sound field unit. For a triangular sound field unit, the coordinates of the reflection point are substituted into the unit's spatial plane equation to verify whether the point is in the plane and whether it is within the closed spatial region formed by the vertices. If the coordinates of the reflection point satisfy the spatial boundary conditions of a certain sound field unit, the information such as the ID, amplitude sequence, and arrival time of the echo data are associated with the unit number.

[0058] Spatial verification: Ensure that each reflection point belongs to only one sound field unit. If a reflection point is located at the unit boundary, it is assigned to the nearest unit according to the principle of proximity. The determination is based on the Euclidean distance from the reflection point to the unit centroid, that is, the square root of the sum of the squares of the coordinate differences between the reflection point and the unit centroid on the three coordinate axes. Integrity verification: Statistically analyze the allocation results of all echo data. If the proportion of unassigned echo data exceeds 5%, readjust the spatial boundary parameters of the sound field unit until the proportion of unassigned data is less than 1%. Integrate all echo data assigned by unit number. For each sound field unit, extract the amplitude characteristics (peak value, mean value, root mean square value) and time characteristics (mean value of arrival time, propagation time) of all echo data under it to obtain the unit echo data corresponding to each triangular sound field unit.

[0059] Step 302: Based on the dynamic deviation correction operator, obtain the deviation correction parameters corresponding to each triangular sound field unit; perform amplitude compensation and time delay correction on the unit echo data according to the deviation correction parameters to obtain the corrected unit echo data. Specifically, this includes: retrieving the generated dynamic deviation correction operator, which is a structured data carrier containing core content such as the unit number of all sound field units, the corresponding amplitude compensation coefficient, time delay correction value, correction validity mark, anomaly level, and residual stress reference value; extracting the amplitude compensation coefficient and time delay correction value corresponding to each triangular sound field unit from the dynamic deviation correction operator one by one according to the unique matching relationship of the unit number; performing correction processing on the unit echo data of each triangular sound field unit, first performing amplitude compensation operation, extracting all amplitude feature values ​​such as peak value, mean value, and root mean square value contained in the unit echo data, multiplying each amplitude feature value by the amplitude compensation coefficient corresponding to the unit to obtain the compensated value corresponding to each amplitude feature, and directly replacing the original amplitude feature value to complete the amplitude compensation.

[0060] The time delay correction operation is performed by extracting all time feature values, such as the average arrival time and propagation duration, from the unit echo data. Each time feature value is then added to the corresponding time delay correction value to obtain the corrected value for each time feature, which directly replaces the original time feature value to complete the time delay correction. For sound field units marked as normal in the dynamic deviation correction operator, the amplitude compensation coefficient is set to one and the time delay correction value to zero, without any correction processing of the unit echo data; all feature values ​​of the original unit echo data are directly retained. For sound field units marked as abnormal in the dynamic deviation correction operator, the reasonableness of the amplitude compensation coefficient and time delay correction value is verified according to the parameter adjustment rules corresponding to the marked abnormality level. After confirming no deviation, the entire correction process is completed according to the above amplitude compensation and time delay correction operation methods. After completing the amplitude compensation and time delay correction for a single triangular sound field unit, all corrected amplitude features and time features of the unit are integrated to form the unique corrected unit echo data corresponding to each triangular sound field unit.

[0061] Step 303 involves stitching and fusing the corrected echo data of all triangular acoustic field units to obtain enhanced ultrasonic feature data. Specifically, this includes: determining the spatial arrangement rules within the separation column based on its geometric structure, ensuring these rules are completely consistent with the actual spatial positions of each unit in the acoustic field unit topology; sequentially numbering and sorting all triangular acoustic field units according to the axial distribution from the inlet to the outlet and the radial distribution from the center to the column wall; extracting the corrected echo data of each triangular acoustic field unit one by one according to this sorted order, fully preserving the unit number, spatial index information, and all corresponding amplitude and temporal feature data during extraction; and performing data stitching based on the spatial boundary of the acoustic field units, seamlessly connecting the echo data of adjacent acoustic field units at the spatial boundary to ensure no numerical discontinuities or feature loss in the amplitude and temporal features of adjacent units at the boundary. For the acoustic field unit data corresponding to the core area in the middle section of the separation column and the detected blind zone, the complete feature information is retained without any simplification.

[0062] After data stitching is completed, data fusion is performed. The stitched data is then subjected to full-dimensional feature normalization to eliminate the differences in feature dimensions between different sound field units caused by differences in detection position and propagation path. The amplitude and temporal features of all sound field units are fully integrated into a continuous and uninterrupted feature dataset. Finally, a unified overall spatial coordinate reference system is added to the integrated continuous feature dataset. This coordinate reference system is completely consistent with the coordinate benchmark of the previously constructed ultrasonic spatial measurement field, so that the position of each feature point in the feature data can accurately correspond to the actual physical spatial position inside the separation column. After the coordinate system is added, the enhanced ultrasonic feature data is obtained.

[0063] In this embodiment, the original echo data is divided into various acoustic field units according to spatial location, targeted deviation correction is performed, and then the data is spliced ​​and fused to form complete enhanced ultrasonic feature data, so that the ultrasonic data can accurately correspond to the spatial location inside the separation column.

[0064] In a preferred embodiment of the present invention, step 4 involves inputting enhanced ultrasound feature data into a pre-trained deep convolutional neural network model. The pre-trained deep convolutional neural network model includes an input layer, a convolutional layer, a pooling layer, an upsampling layer, a boundary prediction convolutional layer, an activation function layer, a boundary localization layer, and a smoothing constraint unit to obtain the color band boundary feature response. This process may include:

[0065] Step 401: The enhanced ultrasound feature data is input into the input layer of the deep convolutional neural network model to obtain initial feature data. The initial feature data is then processed sequentially through multiple convolutional layers in the deep convolutional neural network model, and downsampled through pooling layers to extract depth features layer by layer, resulting in multi-scale depth feature data. Specifically, this includes: constructing the basic network architecture of the deep convolutional neural network model, which includes an input layer, multiple convolutional layers, multiple pooling layers, an upsampling layer, a boundary prediction convolutional layer, an activation function layer, and a boundary localization layer. The convolutional layers and pooling layers are arranged in an alternating series configuration. The specific number of convolutional layers is set according to the actual feature extraction requirements of the separation column color band boundary detection. Each convolutional layer is independently configured with a different size convolutional kernel and a stride. The pooling layers uniformly use mean pooling. The loss function of the model was determined to be the boundary detection loss function adapted to the boundary detection task, and the optimizer of the model was the adaptive moment estimation optimizer. After the architecture was completed, model training was carried out. Ultrasonic feature data containing different color band separation states and different column positions of the separating columns were selected to construct the model training set. All data in the training set were accurately labeled with the actual spatial position and feature information of the color band boundaries. The training set data was input into the model in batches for forward propagation calculation. Data features were extracted through the collaborative operation of convolutional layers and pooling layers. The loss value between the model prediction result and the actual labeled result was calculated in real time. The weights and bias parameters of each layer of the model were dynamically updated according to the loss value through the backpropagation algorithm. Iterative training was carried out repeatedly until the loss value of the model tended to stabilize and the prediction accuracy reached the preset detection requirements. After the model training was completed, all model parameters were saved immediately.

[0066] Enhanced ultrasound feature data is preprocessed according to the format requirements of the model input layer. The dimension, resolution, and data format of the feature data are adjusted to fully match the receiving specifications of the input layer. The preprocessed enhanced ultrasound feature data is then input into the input layer of the deep convolutional neural network model. The input layer receives the data and performs preliminary format conversion to obtain the initial feature data. The initial feature data is then input into the first convolutional layer of the model. The convolutional kernel configured in this layer performs a convolution operation with the initial feature data to extract shallow local texture features, resulting in the first convolutional layer feature data. Subsequently, the first convolutional layer feature data is input into the corresponding first pooling layer. The pooling operation downsamples the feature data, reducing the dimensionality of the feature data while fully preserving the core feature information, resulting in the first pooling layer feature data. The feature data from the first pooling layer is input into the second convolutional layer. Convolution is performed using the kernel configured in this layer to extract the mid-level texture and contour features of the feature data, resulting in the feature data for the second convolutional layer. This second convolutional layer feature data is then input into the corresponding second pooling layer for downsampling. Following the alternating convolution and pooling process described above, the feature data output from the previous pooling layer is sequentially input into subsequent convolutional and pooling layers. Each convolutional layer extracts deeper and more refined feature information based on its configured kernel size and stride. The pooling layers perform downsampling to simplify the data dimensions. Since the parameter settings of each convolutional and pooling layer differ, the feature data output by each layer corresponds to different feature scales. The feature data of different scales obtained after processing by all convolutional and pooling layers are comprehensively integrated to obtain multi-scale depth feature data.

[0067] Step 402 involves inputting multi-scale depth feature data into the upsampling layer of the deep convolutional neural network model. The upsampling operation restores the spatial resolution to the same as the initial feature data, resulting in enhanced fused feature data with uniform resolution. Specifically, this includes: adjusting the format of the multi-scale depth feature data by adjusting the dimensions and tensor format of the feature data according to the input requirements of the model's upsampling layer, ensuring that the format of the feature data fully matches the receiving standard of the upsampling layer; inputting the adjusted multi-scale depth feature data into the upsampling layer of the deep convolutional neural network model. The upsampling layer employs deconvolution as its core operation, independently configuring corresponding deconvolution kernels for different scales of depth feature data, and performing scale upsampling on the low-resolution feature data through deconvolution operations; and for each scale of depth feature data, progressively increasing its spatial resolution through upsampling operations, and interpolating and padding the pixels of the feature data during the resolution upsampling process. To ensure the continuity and integrity of feature information, the spatial resolution of all depth feature data at all scales is sequentially enlarged to be completely consistent with the spatial resolution of the initial feature data, in ascending order of feature scale. All depth feature data with unified resolution after upsampling are then subjected to feature fusion. During the fusion process, the feature values ​​of each feature data at the same spatial location are weighted and summed. The weight coefficients are set according to the actual importance of each layer's features in color band boundary detection. The weight coefficients for shallow features range from 0.6 to 0.8, and the weight coefficients for deep features range from 0.2 to 0.4. The sum of the weight coefficients for shallow and deep features is one. This weight allocation achieves complementarity between spatial location information and boundary contour feature information. The weighted summation calculation process involves multiplying the shallow feature value at the same spatial location by its corresponding weight coefficient, and then adding the deep feature value at that location multiplied by its corresponding weight coefficient to obtain the fused feature value at that location.

[0068] The weighted and summed fused feature data undergoes overall feature enhancement processing. The specific calculation process of the feature enhancement algorithm is as follows: The gradient magnitude of each feature point in the fused feature data is calculated. This gradient magnitude is the square root of the sum of the squares of the gradient values ​​in the x-direction and y-direction. The x-direction gradient value represents the gray-level change rate of the feature point in the x-axis direction, and the y-direction gradient value represents the gray-level change rate of the feature point in the y-axis direction. A boundary feature enhancement coefficient is then set, with a value ranging from 1.5 to 2.5. For feature points with gradient magnitudes greater than a preset gradient threshold, their feature values ​​are multiplied by this enhancement coefficient to achieve enhancement. The preset gradient threshold ranges from 0.4 to 0.6. Feature points with gradient magnitudes less than or equal to the preset gradient threshold are preserved. The feature values ​​remain unchanged. The specific calculation process of the noise suppression algorithm is as follows: calculate the global mean and global standard deviation of the fused feature data. The global mean is the sum of the feature values ​​of all feature points divided by the total number of feature points. The global standard deviation is the square root of the sum of the squares of the differences between the feature values ​​of all feature points and the global mean divided by the total number of feature points. Set the noise judgment threshold to the global mean plus 1.5 times the global standard deviation. For feature points with feature values ​​lower than this threshold, multiply their feature values ​​by a noise attenuation coefficient to achieve noise suppression. The noise attenuation coefficient ranges from 0.1 to 0.3. For feature points with feature values ​​higher than this threshold, keep the feature values ​​unchanged. After processing by the above feature enhancement algorithm and noise suppression algorithm, enhanced fused feature data with uniform resolution is obtained.

[0069] Step 403a involves inputting the enhanced fusion feature data into the boundary prediction convolutional layer of the deep convolutional neural network model to extract boundary response features and obtain initial boundary response feature data. Specifically, this includes configuring a dedicated convolutional kernel for color band boundary feature extraction in the boundary prediction convolutional layer of the deep convolutional neural network model. The size, number, and stride of this dedicated convolutional kernel are set according to the morphological features of the color band boundary and the actual detection accuracy requirements. All parameters of the convolutional kernel are adaptation parameters obtained after model training. The enhanced fusion feature data is then directly input into the boundary prediction convolutional layer of the deep convolutional neural network model. The specific convolution operation process is as follows: A dedicated convolutional kernel slides across the feature data matrix with a set stride. At each position, the weight parameters within the kernel are multiplied one by one with the corresponding feature value. The sum of all products is then obtained to obtain the convolutional output value at that position. During the operation, the focus is on capturing gradient change features and edge contour features related to the color band boundary in the feature data. These captured features are the response features of the color band boundary. All color band boundary response features obtained after the convolution operation are systematically integrated according to spatial dimensions. During the integration process, the spatial coordinate information and response intensity information of the feature data are fully preserved without any feature deletion. The integrated data forms the initial boundary response feature data.

[0070] Step 403b involves passing the initial boundary response feature data to the activation function layer for nonlinear mapping processing to obtain pre-activated boundary response feature data. Specifically, this includes: independently setting an activation function layer in the deep convolutional neural network model; selecting a nonlinear activation function based on the task requirements of color band boundary detection; this activation function being an adapted function determined through multiple iterations during model training, effectively enhancing the response intensity of boundary features while maximizing the suppression of background feature interference; passing the initial boundary response feature data completely and without loss to the activation function layer of the deep convolutional neural network model; substituting each feature value in the initial boundary response feature data into the activation function for nonlinear mapping calculation; amplifying the response intensity of the color band boundary features through nonlinear mapping calculation while reducing the response intensity of non-boundary background features, thereby enhancing effective boundary features and suppressing ineffective background features; and re-integrating all feature values ​​processed by the activation function according to their original spatial coordinate information, fully preserving the processed response intensity information and spatial location information during the integration process to ensure the correspondence between feature positions and actual spatial positions, resulting in pre-activated boundary response feature data.

[0071] Step 403c involves inputting the pre-activated boundary response feature data into the boundary localization layer of the deep convolutional neural network model to perform precise localization analysis of the boundary position, obtaining the color band boundary feature response. Specifically, this includes: constructing a precise boundary position localization analysis algorithm for color band boundary detection adapted to the boundary localization layer of the deep convolutional neural network model. This algorithm comprises three core components: feature selection, spatial fitting, and position calibration. All parameters of the algorithm are adaptation parameters obtained after model training. The response intensity threshold in the feature selection component ranges from 0.3 to 0.5. The pre-activated boundary response feature data is then input into the boundary localization layer of the deep convolutional neural network model. Through the feature selection component of the analysis algorithm, effective feature points with response intensity values ​​greater than or equal to the preset response intensity threshold are selected from the pre-activated boundary response feature data. These feature points belong to the color band edge. For the boundary feature points, invalid background feature points with response intensity values ​​below a threshold are directly eliminated. Through the spatial fitting step in the analysis algorithm, the selected valid feature points are spatially fitted. Based on the spatial coordinates and response intensity information of the valid feature points, combined with the geometric features of the separation column, the spatial distribution pattern of the valid feature points within the separation column is determined, initially locking the approximate spatial location of the color band boundary. Then, through the position calibration step in the analysis algorithm, combined with the actual spatial geometric features of the separation column and the coordinate reference of the previously constructed ultrasonic spatial measurement field, the initially locked color band boundary position is precisely calibrated, correcting minor positional deviations caused by feature extraction and data processing, improving the accuracy of the boundary position. Finally, the calibrated spatial position information of the color band boundary and the corresponding response intensity information are integrated to form a color band boundary feature response that accurately reflects the actual position and detailed characteristics of the color band boundary within the separation column.

[0072] In this embodiment, multi-scale feature data is restored to the initial resolution through upsampling and then weighted and fused. The clear range of weight values ​​makes feature fusion more targeted, which not only fully preserves the detailed information of the depth features, but also ensures the spatial resolution of the feature data, effectively improving the integrity and accuracy of the feature data.

[0073] In a preferred embodiment of the present invention, step 5, which involves using a boundary localization layer to perform extreme value search on the color band boundary feature response to obtain boundary candidate points, calculating the confidence level based on the response intensity and neighborhood distribution characteristics of the boundary candidate points, and then obtaining the color band boundary detection result through coordinate fitting, consistency correction of smooth constraint units, and connection processing, may include:

[0074] Step 501: Utilize the boundary localization layer of the deep convolutional neural network model to perform extreme value search processing on the color band boundary feature response, extracting the spatial locations corresponding to the local maxima of the response intensity, and using these spatial locations as boundary candidate points. Specifically, this includes: in the boundary localization layer of the deep convolutional neural network model, calling a pre-built extreme value search processing algorithm. This algorithm can perform point-by-point detection and neighborhood comparison analysis on the response intensity of the feature data; using this extreme value search processing algorithm, perform a full-range, no-dead-angle extreme value search processing on the color band boundary feature response, taking the feature point at each spatial location in the color band boundary feature response as the center, and according to a preset neighborhood range (value range is 3×3). (Pixels to 7×7 pixels) Select the surrounding feature points as neighborhood feature points, and compare the response intensity value of the central feature point with the response intensity values ​​of all neighborhood feature points one by one; if the response intensity value of the central feature point is greater than the response intensity values ​​of all neighborhood feature points, the response intensity of the central feature point is determined to be a local maximum, and the spatial coordinates corresponding to the local maximum are immediately extracted; according to the above point-by-point detection and neighborhood comparison method, all feature points in the color band boundary feature response are completely detected and compared, and the spatial coordinates corresponding to all local maxima of response intensity are extracted, and each extracted spatial coordinate is used as an independent boundary candidate point.

[0075] Step 502: Based on the response intensity values ​​at each boundary candidate point and the response distribution characteristics within its neighborhood, calculate the confidence score of each boundary candidate point as belonging to the true boundary, obtaining boundary candidate point confidence data. Specifically, this includes: calculating the confidence score for each boundary candidate point, extracting the response intensity value of the boundary candidate point itself, selecting the spatial region surrounding the boundary candidate point as its neighborhood according to a preset fixed range (value range from 3×3 pixels to 5×5 pixels), completely extracting the response intensity values ​​of all feature points within the neighborhood, statistically analyzing the extracted neighborhood response intensity values ​​to obtain the response distribution characteristics within the neighborhood, which include the mean, variance, and overall trend of the response intensity within the neighborhood; calculating the ratio of the response intensity value of the boundary candidate point to the mean of the response intensity within the neighborhood, the calculation process being that this ratio equals the response intensity value of the boundary candidate point itself divided by the mean of the response intensity within the neighborhood, this ratio directly reflects the degree of difference between the candidate point and the neighborhood features; and then combining the variance and variation of the response intensity within the neighborhood... The ratio is adjusted by a trend correction, with the correction coefficient ranging from 0.7 to 1.2. The smaller the variance of the response intensity in the neighborhood and the closer the trend conforms to the characteristics of the color band boundary, the closer the correction coefficient is to 1.2; conversely, the closer the correction coefficient is to 0.7. The corrected ratio is equal to the uncorrected ratio multiplied by the correction coefficient. The corrected ratio is then normalized. The normalization calculation is that the confidence level is equal to the corrected ratio of the candidate boundary point divided by the maximum value of the corrected ratios of all candidate boundary points. This calculation ensures that the confidence level is within the range of 0 and 1. The normalized value is the confidence level that the candidate boundary point belongs to the true boundary. The closer the value is to 1, the higher the probability that the candidate point is the true color band boundary. Following the above confidence level calculation method, the confidence level of each candidate boundary point belonging to the true boundary is calculated sequentially. The spatial coordinates of all candidate boundary points are associated and matched with their corresponding confidence levels to form a structured dataset. After integration, the confidence level data of the candidate boundary points is obtained.

[0076] Step 503: Based on the confidence data of the boundary candidate points, perform coordinate fitting processing on the boundary candidate points to obtain preliminary boundary position coordinate data; input the preliminary boundary position coordinate data into the smoothing constraint unit in the boundary localization layer of the deep convolutional neural network model for consistency correction processing to obtain corrected boundary position coordinate data; based on the corrected boundary position coordinate data, perform connection processing to obtain the color band boundary detection result, specifically including: setting a reasonable confidence threshold according to the actual results of model training and the accuracy requirements of color band boundary detection, the value of which ranges from 0.6 to 0.8, and based on this confidence threshold... The confidence data of the boundary candidate points are filtered, and false boundary candidate points with confidence values ​​below the threshold are directly removed, retaining only valid boundary candidate points with confidence values ​​higher than or equal to the threshold. Coordinate fitting is then performed on the retained valid boundary candidate points. Based on the actual spatial coordinate information of these valid candidate points, a curve fitting method adapted to the characteristics of the separating column is selected. Closely combining the geometric features of the separating column, the spatial coordinates of all retained valid boundary candidate points are fitted into a continuous curve. All spatial coordinate points of this fitted curve are extracted, and these coordinate points are integrated to obtain the preliminary boundary position coordinate data.

[0077] A smoothing constraint unit is independently set in the boundary localization layer of the deep convolutional neural network model. This unit has a built-in spatial smoothing algorithm and a consistency check algorithm. All parameters of the algorithm are the adaptation parameters determined after model training. The initial boundary position coordinate data is input into this smoothing constraint unit. The spatial smoothing algorithm processes the initial boundary position coordinate data by using a weighted moving average method. For any coordinate point in the curve, one to three adjacent coordinate points before and after the point are taken to form a smoothing window, and the smoothed coordinate value is calculated. The smoothed horizontal coordinate value is calculated by multiplying the horizontal coordinate value of each coordinate point in the window by the corresponding weight coefficient, then adding all the product results, and finally dividing the sum by the sum of all weight coefficients. The smoothed vertical coordinate value is calculated in the same way as the horizontal coordinate value. The weight coefficient is allocated according to the principle that the closer to the current coordinate point, the larger the weight value. For example, the weight allocation in the window is 0.4 for the current point, 0.3 for the adjacent point, 0.2 for the second adjacent point, and 0.1 for the farthest point. This calculation eliminates the spikes and discontinuities caused by discrete points in the fitted curve, making the transition of the curve smoother and more continuous.

[0078] The consistency verification algorithm works as follows: First, based on the geometric parameters of the separating column (such as the column radius and axial length) and the migration law of the color band boundary, a reasonable range of boundary coordinate values ​​is constructed. The radial coordinate range is 0.1 to 0.9 times the inner diameter of the separating column, and the axial coordinate range is 0 to 1 times the effective length of the separating column. Then, each smoothed coordinate point is verified one by one. If a coordinate point exceeds the reasonable range, it is corrected as follows: the corrected coordinate is equal to the smoothed coordinate plus the difference between the reasonable range center value and the smoothed coordinate, multiplied by a compensation coefficient, which ranges from 0.1 to 0.3. If a coordinate point does not exceed the range but deviates from the fitting trend of adjacent points, the deviation value between this point and the fitted line of the five adjacent points is calculated. When the deviation... When the value is greater than 0.05 times the separation column radius, it is corrected by 50% to 80% of the deviation value to ensure that the boundary position reflected by the coordinate data fully conforms to the actual situation inside the separation column. After spatial smoothing and consistency verification, the corrected boundary position coordinate data is obtained. The corrected boundary position coordinate data is then connected. Strictly following the axial and radial distribution order of the separation column, all coordinate points in the corrected boundary position coordinate data are sequentially and seamlessly connected to form a continuous, complete color band boundary contour curve that conforms to the actual situation inside the separation column. This contour curve contains all the spatial position information of the color band boundary inside the separation column, which is the final color band boundary detection result.

[0079] This embodiment forms a complete boundary coordinate processing logic from candidate point selection and fitting to smoothing correction, ensuring that the final contour curve fits the actual situation of the separation column, effectively improving the overall efficiency of boundary detection.

[0080] As shown in Figure 2, embodiments of the present invention also provide an automatic detection system for the boundary of a separated column color band based on image recognition, including:

[0081] The acquisition module is used to transmit ultrasonic waves to the separation column and receive reflected signals to obtain raw ultrasonic echo data; and three ultrasonic reference nuclei are set at the center of the sample inlet end face, the center of the sample outlet end face, and the center of the cross section of the middle section of the column to obtain the spatial coordinate vector of the ultrasonic reference nuclei.

[0082] The deviation feature extraction module is used to construct an ultrasonic spatial measurement field based on the spatial coordinate vectors of three ultrasonic reference kernels; perform acoustic field unit mosaicking on the ultrasonic spatial measurement field to obtain the acoustic field unit topology, which contains multiple triangular acoustic field units; obtain the coordinates of each vertex of the triangular acoustic field unit, accumulate the product of the x-coordinate of each vertex and the y-coordinate of the next vertex to obtain the first accumulated value, accumulate the product of the y-coordinate of each vertex and the x-coordinate of the next vertex to obtain the second accumulated value, and take half the absolute value of the difference between the first accumulated value and the second accumulated value as the area of ​​the triangular acoustic field unit; perform deviation feature extraction analysis on the area of ​​all triangular acoustic field units to generate a dynamic deviation correction operator;

[0083] The ultrasonic feature processing module is used to perform ultrasonic feature enhancement processing based on the original ultrasonic echo data and dynamic deviation correction operator to obtain enhanced ultrasonic feature data.

[0084] The feature response acquisition module is used to input enhanced ultrasound feature data into a pre-trained deep convolutional neural network model. The pre-trained deep convolutional neural network model includes an input layer, a convolutional layer, a pooling layer, an upsampling layer, a boundary prediction convolutional layer, an activation function layer, a boundary localization layer, and a smoothing constraint unit to obtain the color band boundary feature response.

[0085] The color band boundary detection module is used to perform extreme value search on the color band boundary feature response using the boundary localization layer to obtain boundary candidate points. The confidence level is calculated based on the response intensity and neighborhood distribution characteristics of the boundary candidate points. Then, through coordinate fitting, consistency correction of smooth constraint units, and connection processing, the color band boundary detection result is obtained.

[0086] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0087] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0088] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An automatic detection method for the boundary of a color band in a separation column based on image recognition, characterized in that, The method includes: Step 1, transmitting ultrasonic waves to the separation column and receiving reflected signals to obtain raw ultrasonic echo data; setting three ultrasonic reference cores at the center of the sample inlet end face, the center of the sample outlet end face, and the center of the cross-section of the middle section of the column, respectively, and obtaining the spatial coordinate vectors of the ultrasonic reference cores; Step 2, constructing an ultrasonic spatial measurement field based on the spatial coordinate vectors of the three ultrasonic reference cores; performing acoustic field unit mosaicking on the ultrasonic spatial measurement field to obtain an acoustic field unit topology, which contains multiple triangular acoustic field units; obtaining the coordinates of each vertex of the triangular acoustic field unit, accumulating the product of the x-coordinate of each vertex and the y-coordinate of the next vertex to obtain a first accumulated value, accumulating the product of the y-coordinate of each vertex and the x-coordinate of the next vertex to obtain a second accumulated value, and taking half the absolute value of the difference between the first accumulated value and the second accumulated value as the area of ​​the triangular acoustic field unit. Step 3: Extract and analyze the deviation features of the area of ​​all triangular acoustic field units to generate a dynamic deviation correction operator; Step 4: Perform ultrasonic feature enhancement processing based on the original ultrasonic echo data and the dynamic deviation correction operator to obtain enhanced ultrasonic feature data; Step 5: Input the enhanced ultrasonic feature data into a pre-trained deep convolutional neural network model, which includes an input layer, convolutional layer, pooling layer, upsampling layer, boundary prediction convolutional layer, activation function layer, boundary localization layer, and smoothing constraint unit to obtain the color band boundary feature response; Step 6: Use the boundary localization layer to perform extreme value search on the color band boundary feature response to obtain boundary candidate points, calculate the confidence level based on the response intensity and neighborhood distribution characteristics of the boundary candidate points, and then obtain the color band boundary detection result through coordinate fitting, consistency correction of the smoothing constraint unit, and connection processing.

2. The automatic detection method for the boundary of the color band of the separation column based on image recognition according to claim 1, characterized in that, Step 1 includes: transmitting ultrasonic pulses into the separation column and receiving the reflected echo signals; amplifying, filtering, and performing analog-to-digital conversion on the echo signals to obtain raw ultrasonic echo data; identifying the reflected echoes corresponding to three ultrasonic reference nuclei pre-set at the center of the sample inlet end face, the center of the sample outlet end face, and the center of the cross-section of the middle section of the column, based on the amplitude and time characteristics of the echo signals at different positions in the raw ultrasonic echo data; extracting echo signal segments corresponding to each of the three ultrasonic reference nuclei from the raw ultrasonic echo data; performing arrival time extraction and analysis on the three echo signal segments to obtain arrival time information for the echo signals of each ultrasonic reference nuclei; calculating the arrival time difference between any two ultrasonic reference nuclei based on the arrival time information; comparing the arrival time difference with the fixed distance value corresponding to the pre-set geometric position relationship of the three ultrasonic reference nuclei on the separation column to calculate the actual propagation speed; calculating the spatial distance from each ultrasonic reference nuclei to the ultrasonic emission point based on the actual propagation speed and arrival time information; and combining the fixed spatial position of the ultrasonic emission point during the measurement process to calculate the spatial coordinate vector of the three ultrasonic reference nuclei.

3. The automatic detection method for the boundary of the color band of the separation column based on image recognition according to claim 2, characterized in that, An ultrasonic spatial measurement field is constructed based on the spatial coordinate vectors of three ultrasonic reference cores; The ultrasonic spatial measurement field is processed by sound field unit mosaicking to obtain the sound field unit topology. The sound field unit topology contains multiple triangular sound field units, including: the residual stress distribution state inside the separation column material is analyzed based on the spatial coordinate vectors of three ultrasonic reference cores and the original ultrasonic echo data to obtain the residual stress distribution data. Based on the residual stress distribution data and the actual propagation speed, the propagation speed of ultrasound in the separation column medium is compensated and corrected point by point to obtain the corrected propagation speed distribution data. Based on the corrected propagation speed distribution data and the spatial coordinate vectors of the three ultrasound reference cores, an ultrasound spatial measurement field is constructed. Using the positions of the spatial coordinate vectors of the three ultrasound reference cores as initial mosaic nodes, auxiliary mosaic nodes are generated between adjacent ultrasound reference cores and between the ultrasound reference cores and the boundary of the separation column. The initial mosaic nodes and the auxiliary mosaic nodes are connected to obtain a grid topology covering the entire ultrasound spatial measurement field. The grid topology contains multiple triangular acoustic field units, and the collection of all triangular acoustic field units constitutes the acoustic field unit topology.

4. The automatic detection method for the boundary of the color band of the separation column based on image recognition according to claim 3, characterized in that, Obtain the coordinates of each vertex of the triangular sound field unit, accumulate the product of the x-coordinate of each vertex and the y-coordinate of the next vertex to obtain the first accumulated value, accumulate the product of the y-coordinate of each vertex and the x-coordinate of the next vertex to obtain the second accumulated value, and take half of the absolute value of the difference between the first accumulated value and the second accumulated value as the area of ​​the triangular sound field unit. The deviation feature extraction and analysis of the area of ​​all triangular sound field units is performed to generate a dynamic deviation correction operator. This includes: obtaining the coordinates of each vertex of each triangular sound field unit based on its topology; accumulating the products of the first vertex's x-coordinate and the second vertex's y-coordinate, the second vertex's x-coordinate and the third vertex's y-coordinate, and the third vertex's x-coordinate and the first vertex's y-coordinate to obtain a first accumulated value; accumulating the products of the first vertex's y-coordinate and the second vertex's x-coordinate, the second vertex's y-coordinate and the third vertex's y-coordinate, and the third vertex's y-coordinate and the first vertex's x-coordinate to obtain a second accumulated value; taking half the absolute value of the difference between the first and second accumulated values ​​as the area of ​​the triangular sound field unit; comparing the area of ​​each triangular sound field unit with a pre-set standard unit area threshold to filter out abnormal sound field units whose area deviation exceeds the preset threshold; and statistically analyzing the spatial distribution and deviation of the abnormal sound field units to generate a dynamic deviation correction operator.

5. The automatic detection method for the boundary of the color band of the separation column based on image recognition according to claim 4, characterized in that, Ultrasonic feature enhancement processing is performed based on the original ultrasonic echo data and dynamic deviation correction operator to obtain enhanced ultrasonic feature data. This includes: dividing the original ultrasonic echo data into various triangular acoustic field units according to their spatial correspondence based on the acoustic field unit topology, obtaining the unit echo data corresponding to each triangular acoustic field unit; obtaining the deviation correction parameters corresponding to each triangular acoustic field unit according to the dynamic deviation correction operator; performing amplitude compensation and time delay correction on the unit echo data according to the deviation correction parameters to obtain the corrected unit echo data; and stitching and fusing the corrected unit echo data corresponding to all triangular acoustic field units to obtain enhanced ultrasonic feature data.

6. The automatic detection method for the boundary of the color band of the separation column based on image recognition according to claim 5, characterized in that, Enhanced ultrasound feature data is input into a pre-trained deep convolutional neural network (DCNN) model, which includes an input layer, convolutional layers, pooling layers, upsampling layers, boundary prediction convolutional layers, activation function layers, boundary localization layers, and smoothing constraint units, to obtain color band boundary feature responses. This process includes: inputting enhanced ultrasound feature data into the input layer of the DCNN model to obtain initial feature data; sequentially processing the initial feature data through multiple convolutional layers in the DCNN model and downsampling it through pooling layers to extract depth features layer by layer, obtaining multi-scale depth feature data; inputting the multi-scale depth feature data into the upsampling layer of the DCNN model to restore it to the same spatial resolution as the initial feature data, obtaining enhanced fused feature data with uniform resolution; and inputting the enhanced fused feature data into the boundary prediction convolutional layer of the DCNN model to generate color band boundary feature responses through convolution operations and activation function processing.

7. The automatic detection method for the boundary of the color band of the separation column based on image recognition according to claim 6, characterized in that, The enhanced fusion feature data is input into the boundary prediction convolutional layer of the deep convolutional neural network model. Through convolution operations and activation function processing, the color band boundary feature response is generated. This includes: inputting the enhanced fusion feature data into the boundary prediction convolutional layer of the deep convolutional neural network model to extract boundary response features and obtain initial boundary response feature data; passing the initial boundary response feature data to the activation function layer for nonlinear mapping processing to obtain preliminary activated boundary response feature data; and inputting the preliminary activated boundary response feature data into the boundary localization layer of the deep convolutional neural network model to perform precise localization analysis of the boundary position and obtain the color band boundary feature response.

8. The automatic detection method for the boundary of the color band of the separation column based on image recognition according to claim 7, characterized in that, The process involves using a boundary localization layer to perform extreme value search on the color band boundary feature response to obtain candidate boundary points. Confidence is calculated based on the response intensity and neighborhood distribution characteristics of these candidate points. Finally, coordinate fitting, consistency correction by smoothing constraint units, and connection processing are performed to obtain the color band boundary detection result. This includes: using the boundary localization layer of a deep convolutional neural network model to perform extreme value search on the color band boundary feature response, extracting the spatial locations corresponding to local maxima of the response intensity, and using these spatial locations as candidate boundary points; calculating the confidence that each candidate boundary point belongs to the true boundary based on the response intensity value at each candidate boundary point and the response distribution characteristics within its neighborhood, obtaining confidence data for the candidate boundary points; performing coordinate fitting on the candidate boundary points based on the confidence data to obtain preliminary boundary position coordinate data; inputting the preliminary boundary position coordinate data into the smoothing constraint units of the boundary localization layer of the deep convolutional neural network model for consistency correction, obtaining corrected boundary position coordinate data; and performing connection processing based on the corrected boundary position coordinate data to obtain the color band boundary detection result.

9. An automatic detection system for the boundary of a color band in a separation column based on image recognition, wherein the system implements the method as described in any one of claims 1 to 8, characterized in that, include: The acquisition module is used to transmit ultrasonic waves to the separation column and receive the reflected signals to obtain the raw ultrasonic echo data. Three ultrasonic reference nuclei were set at the center of the sample inlet end face, the center of the sample outlet end face, and the center of the cross-section of the middle section of the column, respectively, and the spatial coordinate vectors of the ultrasonic reference nuclei were obtained. A deviation feature extraction module was used to construct an ultrasonic spatial measurement field based on the spatial coordinate vectors of the three ultrasonic reference nuclei. The ultrasonic spatial measurement field was then processed by acoustic field element mosaicking to obtain an acoustic field element topology, which contained multiple triangular acoustic field elements. The coordinates of each vertex of the triangular acoustic field element were obtained. The products of the x-coordinate of each vertex and the y-coordinate of the next vertex were accumulated to obtain a first accumulated value, and the products of the y-coordinate of each vertex and the x-coordinate of the next vertex were accumulated to obtain a second accumulated value. Half the absolute value of the difference between the first and second accumulated values ​​was taken as the area of ​​the triangular acoustic field element. Deviation feature extraction analysis was performed on the areas of all triangular acoustic field elements to generate... The system includes: a dynamic deviation correction operator; an ultrasonic feature processing module for enhancing ultrasonic features based on the original ultrasonic echo data and the dynamic deviation correction operator to obtain enhanced ultrasonic feature data; a feature response acquisition module for inputting the enhanced ultrasonic feature data into a pre-trained deep convolutional neural network model, which includes an input layer, convolutional layer, pooling layer, upsampling layer, boundary prediction convolutional layer, activation function layer, boundary localization layer, and smoothing constraint unit to obtain the color band boundary feature response; and a color band boundary detection module for using the boundary localization layer to perform extreme value search on the color band boundary feature response to obtain boundary candidate points, calculating the confidence level based on the response intensity and neighborhood distribution characteristics of the boundary candidate points, and then obtaining the color band boundary detection result through coordinate fitting, consistency correction of the smoothing constraint unit, and connection processing.

10. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 8.

Images