Automatic correction method for multi-station material flow detection equipment on conveyor belt

Abstract

The invention provides an automatic correction method for multi-site material flow detection equipment on a conveyor belt, which is characterized in that on the basis of industrial material flow dynamic statistics based on fusion of detection results of the multi-site material flow equipment, a plurality of pieces of material detection equipment are deployed on the conveyor belt, and multiple sites are formed in an array form; spatial distribution characteristic data of material flowing are collected in real time, and estimation processing and clustering analysis are carried out on multi-source measurement data through data statistical analysis, specifically, firstly, material flow data are collected through multiple devices at the same time, then time offset calculation is carried out on all measurement values based on material time sequence characteristics, and the time offset is calculated; and finally, obtaining an accurate flow estimation value through k-nearest neighbor kernel estimation and a clustering processing algorithm. According to the real value fitting estimation method based on the multi-site material flow value, through multi-site material flow volume detection data fusion and algorithm optimization, high-precision dynamic estimation of the material flow can be achieved, and the actual flow value of each device is corrected.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial process monitoring, and in particular to an automatic correction method for multi-station material flow detection equipment on a conveyor belt. The method comprises a true value fitting estimation method based on the multi-station material flow values, and realizes high-precision dynamic estimation of the material flow through the fusion of multi-station material flow volume detection data and algorithm optimization, and achieves correction of the actual flow value of each device. Background Art

[0002] In the field of industrial process monitoring, accurate material flow measurement is crucial for production control, quality optimization, and energy management. Traditional material flow detection methods typically rely on a single sensor. However, due to factors such as fluctuating operating conditions, changes in material properties, and sensor drift, measurement results are prone to deviations, making it difficult to simultaneously meet real-time and accuracy requirements. Existing technologies often use fixed correction coefficients or simple averaging, which cannot adapt to the dynamic changes in complex industrial environments.

[0003] Current flow estimation methods based on single-point detection have significant limitations: high-precision sensors are expensive and complex to maintain, while standard sensors exhibit significantly increased errors when materials are unevenly distributed or in the presence of interference. Furthermore, offline statistical methods cannot correct for systematic errors introduced by changes in operating conditions in real time, resulting in reduced reliability of long-term monitoring data.

[0004] A conveyor belt production line is hundreds to thousands of meters long and is generally equipped with multiple detection devices. It is a very intuitive idea to obtain the appropriate true value of material flow through statistical analysis of the material flow detection results of multiple devices.

[0005] Single-point material flow detection technology is a method that uses local sensors to measure material flow parameters (such as flow rate, density, etc.) and then infer the overall flow rate. It is widely used in the field of industrial production process monitoring. In the mining, metallurgy, chemical and other industries, this method is often used in scenarios such as raw material transportation measurement, energy consumption statistics and process optimization. However, traditional single-point detection solutions have the inherent defect that sensors are easily affected by uneven material distribution, environmental interference and equipment drift, resulting in a significant reduction in measurement accuracy under complex working conditions, making it difficult to meet the data reliability requirements of modern intelligent manufacturing. Summary of the Invention

[0006] The present invention proposes an automatic correction method for multi-station material flow detection equipment on a conveyor belt. Based on the true value fitting estimation method of the multi-station material flow value, through the fusion of multi-station material flow volume detection data and algorithm optimization, it can achieve high-precision dynamic estimation of the material flow and correct the actual flow value of each device.

[0007] The present invention adopts the following technical solutions.

[0008] An automatic calibration method for multi-site material flow detection equipment on a conveyor belt is disclosed. The method is based on dynamic statistics of industrial material flow based on the fusion of detection results of multi-site material flow equipment. By deploying multiple material detection equipment on the conveyor belt to form multiple sites in an array, spatial distribution characteristic data of material flow is collected in real time. Data statistical analysis is then used to perform estimation processing and cluster analysis on multi-source measurement data. Specifically, the method comprises the following steps: first, material flow data is collected simultaneously by multiple devices; then, time offsets are calculated for each measurement value based on the material time series characteristics; and finally, accurate flow estimation values ​​are obtained through k-nearest neighbor kernel estimation and clustering processing algorithms.

[0009] Testing equipment installation points such as Figure 1 As shown, material detection equipment is installed at each material transfer station to obtain the real-time material flow value of a certain feed port or discharge port of the current station.

[0010] The correction methods are all explained using a certain type of material port (such as these stations are all feed ports) as an example. If another type of material port needs to be corrected, it will also be corrected according to the type.

[0011] The material detection equipment includes a high-speed camera, a line laser generator, and a terminal device. When calculating the real-time material flow value, the concave and convex characteristics reflected by the laser line of the line laser generator on the material are not affected by the material's shape, so it can support conveyor belt materials of any shape, including granular, solid and other shapes.

[0012] The automatic correction method first performs statistics and data filtering on the material flow value provided by the detection equipment to obtain the material flow value curve of each site, then calculates the time offset between each site and the upstream site, aligns the data of each site with the data of the upstream site in time, and finally performs k-nearest neighbor kernel estimation and clustering processing to obtain the final true value estimation curve. Then, a difference curve fitting calculation is performed based on the true value estimation curve and the original curve to obtain the difference curve of each site. After that, a correction calculation is performed based on the difference curve of each site to obtain the corrected curve.

[0013] The automatic correction method comprises the following steps:

[0014] Step 1: Data statistics and preprocessing;

[0015] Step 2: Calculate the time offset of each site;

[0016] Step 3: k-nearest neighbor kernel fitting estimation and clustering processing;

[0017] Step 4: Site material flow correction.

[0018] In step one, if Figure 2 , Figure 3As shown in the figure, the colored dots represent each station. Each station is equipped with material flow detection equipment to provide flow values ​​at the material inlet and outlet. When detecting material flow values ​​at each station, due to environmental and operating conditions, the flow values ​​may contain a certain amount of Gaussian noise and occasional pulse noise, resulting in temporal fluctuations in the data. To prevent occasional abnormal flow values ​​from affecting subsequent processing, data filtering is required. Let f(t), (m ≤ t ≤ n), be the data for a single station over a period of time. Assuming there are k data points within this time range, the filtered result, y(t), = median(f(t),…,f(t+k)). This data is continuously collected (and can be used for other business logic), but subsequent time offset calculations and flow value corrections are performed only within one hour after each conveyor belt line is started (different conveyor belt lines do not interfere with each other).

[0019] In step 2, due to the different distances between each station, the appearance time of a certain material at the inlet and outlet is not equal, resulting in a time offset of the material flow curve on the time axis. Here, the flow value obtained by the material detection at a certain upstream station is used as the reference timestamp t0. The time when the material arrives at the N-1 downstream stations is t1, t2, ..., t N-1 The time offset between each downstream station and the most upstream station is recorded as Δt i =t i -t0; record the material of each downstream station between t0 and t N-1 The traffic data at timestamp x in time is f i (x), and the random error caused by other factors is ε i , with the following relationship:

[0020] f i (x) = f 最上游站点 (x+Δt i )+ε i

[0021] For the calculation of the time offset Δt of a downstream station relative to the most upstream station, since t1, t2, ..., t N-1 The value of cannot be directly observed. Therefore, a reasonable preset time window size, Size, is used here to search for Δt within the upper and lower ranges of Size. When the following relationship is satisfied, the traffic in this window calculation is very small and cannot be used to calculate the time offset. This special case needs to be excluded.

[0022]

[0023] Then, this window performs a Fast Fourier Transform-based cross-correlation coefficient calculation of the flow values ​​of upstream and downstream stations Size times on the flow data, with a step size of 500 milliseconds each time. The calculation process (taking a downstream station and the most upstream station as an example, where i is the timestamp increment) is as follows:

[0024] Perform fast Fourier transform (FFT) on the material flow data series of the upstream and downstream sites respectively:

[0025]

[0026]

[0027] Where N is the sequence length. Here, the data sequence of the downstream site is flipped first, which is equivalent to taking the conjugate of F(k) in the frequency domain. k is the data index in the frequency domain, j is the imaginary unit, F(k) is the frequency domain function of the fast Fourier transform of the downstream site in a pair of upstream and downstream sites, and FFT(f(N-1-x)) is the fast Fourier transform of the data sequence flipped at a downstream site. Similarly, F 最上游站点 (k) means

[0028] Calculate the frequency domain product C(k) and calculate the inverse fast Fourier transform (IFFT): that is, convert the frequency domain product data back to the time domain, and take the real part of the result to obtain the cross-correlation coefficient R(x), and then centralize R(x) to obtain R shift (x), R shift (x) is the cross-correlation coefficient obtained by performing FFTShift(R(x)) on R(x), which moves the zero-frequency component to the center, making it easier to observe R(x). The specific formula and process are as follows:

[0029] C(x)=F(k)*F 最上游站点 (k)

[0030]

[0031] R shift The schematic diagram of the (x) function is as follows Figure 4 As shown;

[0032] Calculate the time offset:

[0033] Where g is the value that makes R shift (x) gets the coordinate of the maximum value, that is, R shift (g) = max(R shift (x)), Size is the data window size for the time offset calculation process mentioned above; through the above process, all Δt i .

[0034] The specific method of step 3 is: the time series of N stations after time axis offset correction is recorded as:

[0035] f i (x); x=1,…,n; i=1,2,…,N

[0036] Where x is the timestamp and i is the site number;

[0037] Obtain the average function corresponding to N sites at timestamp x by weighted averaging Use k x close to x i The corresponding f(x i ) is the weighted average of k-nearest neighbor kernel estimate; denoted by I x,k ={i:x i is one of the k observations closest to x},

[0038] D(x,k)=max{|x i -x|,i∈I x,k}

[0039] Use D(x,k) instead of window width h n , the obtained k-nearest neighbor kernel estimate is defined as follows:

[0040]

[0041] Take k = 5, K(u) as the normal kernel to perform k-nearest neighbor kernel estimation, and let the obtained estimation function As the initial cluster center of K-means clustering processing, denoted as g(x);

[0042] Then do the following:

[0043] Calculate the Euclidean distance d from all time functions to the center

[0044]

[0045] Calculate the distance d i The mean μ d and standard deviation σ d , through 3σ d Criteria for determining outlier curves,

[0046] If d i >μ d +3σ d , it is determined to be an outlier curve;

[0047] If the number of outlier curves is zero, g(x) is the final true value curve; otherwise, the following operations are performed until the number of outlier curves is zero.

[0048] In step 3, the cluster center g(x) is recalculated using the non-outlier curve, and the previous g(x) is recorded as g(x) old Used for multi-class judgment calculation; if the distance between these non-outlier curves and the new center g(x) is greater than 3σ d If the number of outlier curves determined by the criterion is zero, the outlier curve removed is output as an abnormal curve, and this g(x) is used as the final true value curve;

[0049] In multi-class judgment, set the inter-class difference threshold θ = 2σ d , if |g(x)–g(x) old |>θ·|g(x) old |, it is determined that there are multiple categories. At this time, the true value estimation fails, indicating that the data of N sites are not similar and the true value result cannot be obtained. It is possible that abnormal situations have occurred in the detection of multiple sites, and this situation is reported.

[0050] like Figure 5 As shown in the figure, the light-colored lines represent the filtered and time-aligned material flow curves for each site, while the dark-colored lines represent the fitted estimated material flow curves obtained through k-nearest neighbor kernel estimation and clustering. Ultimately, the true value fitted estimates of the material flow at multiple locations are obtained.

[0051] The specific method of step 4 is as follows: after obtaining the true value fitting result curve, each station has a certain deviation from the true value fitting result curve, which will form the difference points of each station. These difference points are used to fit the following quadratic polynomial function, and multiple fittings are performed to obtain the fitted difference curve:

[0052] y d =ax 2 +bx+c

[0053] Where a, b, and c are the coefficients of the quadratic term, linear term, and constant term of the difference curve to be solved respectively;

[0054] Define the loss function: x i ,y i is the coordinate of the i-th point;

[0055] Use L to measure the fitting effect. The smaller the loss value, the better the fitting effect. Finally, take the difference curve of the result with the best fitting effect as the final fitting difference curve.

[0056] The specific solution method to minimize the loss function by using the least squares method is as follows:

[0057]

[0058] Take the partial derivatives of a, b, and c and set them to zero to obtain the three simultaneous equations mentioned above, where K is the number of points. Then solve the system of equations to determine a, b, and c;

[0059] After obtaining the difference curve of each station, it is assumed that the actual flow value during the prediction process is y a , then the actual flow rate of the material flow obtained after correction is y r =y a +y d .

[0060] The terminal device adopts an embedded edge computing terminal device.

[0061] The present invention relates to the field of industrial process monitoring technology, and in particular to a true value fitting estimation method based on multi-site material flow values. By fusing multi-site material flow volume detection data and optimizing algorithms, high-precision dynamic estimation of material flow is achieved, and the actual flow value of each device is corrected. The method of the present invention effectively solves problems such as insufficient representativeness of single-point measurement and sensitivity to environmental interference, significantly improves measurement stability and accuracy in complex industrial scenarios, and can correct flow data of multiple sites through a fitting method, making the flow data reliable.

[0062] This paper proposes a true-value fitting estimation method based on multi-point material flow measurements. By deploying multiple arrays of conveyor belt flow measurement devices to collect the material flow volume on the conveyor belt, and combining this with an adaptive weighted fusion algorithm and a dynamic error compensation model, it achieves high-precision online estimation of material flow. This method effectively addresses the lack of representativeness of a single measurement point, significantly improving measurement stability and accuracy in complex industrial scenarios. Furthermore, the actual flow value of each device is corrected, ensuring reliable and stable material flow data across multiple sites. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0064] Attachment Figure 1 It is a schematic diagram of the system principle of the present invention;

[0065] Attachment Figure 2 1 is a schematic diagram of multi-site traffic statistics in step 1 of the embodiment;

[0066] Attachment Figure 3 1 is a statistical diagram of material flow at a single site in step 1 in the embodiment;

[0067] Attachment Figure 4 In the embodiment, R in step 2 shift (x) Schematic diagram of the function curve;

[0068] Attachment Figure 5 2 is a schematic diagram of a k-nearest neighbor kernel estimation curve in step 3 of the embodiment;

[0069] Attachment Figure 6 It is a schematic diagram of the algorithm flow of the present invention. DETAILED DESCRIPTION

[0070] In response to the shortcomings of the existing technology, this example proposes a dynamic statistical method for industrial material flow on a conveyor belt based on the fusion of detection results of material flow equipment at multiple sites. This method deploys an array of multiple devices on the conveyor belt to collect spatial distribution characteristic data of material flow in real time, and uses data statistical analysis to perform estimation processing and cluster analysis on multi-source measurement data. Specifically, the present invention first uses multiple devices to collect material flow data at the same time, then calculates the time offset of each measurement value based on the material time series characteristics, and finally obtains an accurate flow estimation value through k-nearest neighbor kernel estimation and clustering processing algorithm. This method effectively solves the problems of insufficient representativeness of single-point measurement and sensitivity to environmental interference, and significantly improves the measurement stability and accuracy in complex industrial scenarios. Finally, the flow data of multiple sites is corrected by a fitting method, making the flow data reliable.

[0071] As shown in the figure, an automatic correction method for multi-site material flow detection equipment on a conveyor belt is provided. The method is based on dynamic statistics of industrial material flow based on the fusion of detection results of multi-site material flow equipment. By deploying multiple material detection equipment on the conveyor belt, multiple sites are formed in the form of an array, and spatial distribution characteristic data of material flow are collected in real time. Data statistical analysis is used to estimate and cluster the multi-source measurement data. Specifically, the material flow data is first collected simultaneously by multiple devices, and then the time offset of each measurement value is calculated based on the material time series characteristics. Finally, an accurate flow estimation value is obtained through k-nearest neighbor kernel estimation and clustering processing algorithm.

[0072] Testing equipment installation points such as Figure 1 As shown, material detection equipment is installed at each material transfer station to obtain the real-time material flow value of a certain feed port or discharge port of the current station.

[0073] The correction methods are all explained using a certain type of material port (such as these stations are all feed ports) as an example. If another type of material port needs to be corrected, it will also be corrected according to the type.

[0074] The material detection equipment includes a high-speed camera, a line laser generator, and a terminal device. When calculating the real-time material flow value, the concave and convex characteristics reflected by the laser line of the line laser generator on the material are not affected by the material's shape, so it can support conveyor belt materials of any shape, including granular, solid and other shapes.

[0075] The automatic correction method first performs statistics and data filtering on the material flow value provided by the detection equipment to obtain the material flow value curve of each site, then calculates the time offset between each site and the upstream site, aligns the data of each site with the data of the upstream site in time, and finally performs k-nearest neighbor kernel estimation and clustering processing to obtain the final true value estimation curve. Then, a difference curve fitting calculation is performed based on the true value estimation curve and the original curve to obtain the difference curve of each site. After that, a correction calculation is performed based on the difference curve of each site to obtain the corrected curve.

[0076] The automatic correction method comprises the following steps:

[0077] Step 1: Data statistics and preprocessing;

[0078] Step 2: Calculate the time offset of each site;

[0079] Step 3: k-nearest neighbor kernel fitting estimation and clustering processing;

[0080] Step 4: Site material flow correction.

[0081] In step one, if Figure 2 , Figure 3 As shown in the figure, the colored dots represent each station. Each station is equipped with material flow detection equipment to provide flow values ​​at the material inlet and outlet. When detecting material flow values ​​at each station, due to environmental and operating conditions, the flow values ​​may contain a certain amount of Gaussian noise and occasional pulse noise, resulting in temporal fluctuations in the data. To prevent occasional abnormal flow values ​​from affecting subsequent processing, data filtering is required. Let f(t), (m ≤ t ≤ n), be the data for a single station over a period of time. Assuming there are k data points within this time range, the filtered result, y(t), = median(f(t),…,f(t+k)). This data is continuously collected (and can be used for other business logic), but subsequent time offset calculations and flow value corrections are performed only within one hour after each conveyor belt line is started (different conveyor belt lines do not interfere with each other).

[0082] In step 2, due to the different distances between each station, the appearance time of a certain material at the inlet and outlet is not equal, resulting in a time offset of the material flow curve on the time axis. Here, the flow value obtained by the material detection at a certain upstream station is used as the reference timestamp t0. The time when the material arrives at the N-1 downstream stations is t1, t2, ..., t N-1 The time offset between each downstream station and the most upstream station is recorded as Δt i =t i -t0; record the material of each downstream station between t0 and tN-1 The traffic data at timestamp x in time is f i (x), and the random error caused by other factors is ε i , with the following relationship:

[0083] f i (x) = f 最上游站点 (x+Δt i )+ε i

[0084] For the calculation of the time offset Δt of a downstream station relative to the most upstream station, since t1, t2, ..., t N-1 The value of cannot be directly observed. Therefore, a reasonable preset time window size, Size, is used here to search for Δt within the upper and lower ranges of Size. When the following relationship is satisfied, the traffic in this window calculation is very small and cannot be used to calculate the time offset. This special case needs to be excluded.

[0085]

[0086] Then, this window performs a Fast Fourier Transform-based cross-correlation coefficient calculation of the flow values ​​of upstream and downstream stations Size times on the flow data, with a step size of 500 milliseconds each time. The calculation process (taking a downstream station and the most upstream station as an example, where i is the timestamp increment) is as follows:

[0087] Perform fast Fourier transform (FFT) on the material flow data series of the upstream and downstream sites respectively:

[0088]

[0089] Where N is the sequence length. Here, the data sequence of the downstream site is flipped first, which is equivalent to taking the conjugate of F(k) in the frequency domain. k is the data index in the frequency domain, j is the imaginary unit, F(k) is the frequency domain function of the fast Fourier transform of the downstream site in a pair of upstream and downstream sites, and FFT(f(N-1-x)) is the fast Fourier transform of the data sequence flipped at a downstream site. Similarly, F 最上游站点 (k) means

[0090] Calculate the frequency domain product C(k) and calculate the inverse fast Fourier transform (IFFT): that is, convert the frequency domain product data back to the time domain, and take the real part of the result to obtain the cross-correlation coefficient R(x), and then centralize R(x) to obtain R shift (x), R shift(x) is the cross-correlation coefficient obtained by performing FFTShift(R(x)) on R(x), which moves the zero-frequency component to the center, making it easier to observe R(x). The specific formula and process are as follows:

[0091] C(k)=F(k)*F 最上游站点 (k)

[0092]

[0093] R shift The schematic diagram of the (x) function is as follows Figure 4 As shown;

[0094] Calculate the time offset:

[0095] Where g is the value that makes R shift (x) gets the coordinate of the maximum value, that is, R shift (g) = max(R shift (x)), Size is the data window size for the time offset calculation process mentioned above; through the above process, all Δt i .

[0096] The specific method of step 3 is: the time series of N stations after time axis offset correction is recorded as:

[0097] f i (x); x=1,…,n; i=1,2,…,N

[0098] Where x is the timestamp and i is the site number;

[0099] Obtain the average function corresponding to N sites at timestamp x by weighted averaging Use k x close to x i The corresponding f(x i ) is the weighted average of k-nearest neighbor kernel estimate; denoted by I x,k ={i:x i is one of the k observations closest to x},

[0100] D(x,k)=max{|x i -x|,i∈I x,k}

[0101] Use D(x,k) instead of window width h n , the obtained k-nearest neighbor kernel estimate is defined as follows:

[0102]

[0103] Take k = 5, K(u) as the normal kernel to perform k-nearest neighbor kernel estimation, and let the obtained estimation function As the initial cluster center of K-means clustering processing, denoted as g(x);

[0104] Then do the following:

[0105] Calculate the Euclidean distance d from all time functions to the center i

[0106]

[0107] Calculate the distance d i The mean μ d and standard deviation σ d , through 3σ d Criteria for determining outlier curves,

[0108] If d i >μ d +3σ d , it is determined to be an outlier curve;

[0109] If the number of outlier curves is zero, g(x) is the final true value curve; otherwise, the following operations are performed until the number of outlier curves is zero.

[0110] In step 3, the cluster center g(x) is recalculated using the non-outlier curve, and the previous g(x) is recorded as g(x) old Used for multi-class judgment calculation; if the distance between these non-outlier curves and the new center g(x) is greater than 3σ d If the number of outlier curves determined by the criterion is zero, the outlier curve removed is output as an abnormal curve, and this g(x) is used as the final true value curve;

[0111] In multi-class judgment, set the inter-class difference threshold θ = 2σ d , if |g(x)–g(x) old |>θ·|g(x) old |, it is determined that there are multiple categories. At this time, the true value estimation fails, indicating that the data of N sites are not similar and the true value result cannot be obtained. It is possible that abnormal situations have occurred in the detection of multiple sites, and this situation is reported.

[0112] like Figure 5 As shown in the figure, the light-colored lines represent the filtered and time-aligned material flow curves for each site, while the dark-colored lines represent the fitted estimated material flow curves obtained through k-nearest neighbor kernel estimation and clustering. Ultimately, the true value fitted estimates of the material flow at multiple locations are obtained.

[0113] The specific method of step 4 is as follows: after obtaining the true value fitting result curve, each station has a certain deviation from the true value fitting result curve, which will form the difference points of each station. These difference points are used to fit the following quadratic polynomial function, and multiple fittings are performed to obtain the fitted difference curve:

[0114] y d =ax 2 +bx+c

[0115] Where a, b, and c are the coefficients of the quadratic term, linear term, and constant term of the difference curve to be solved respectively;

[0116] Define the loss function: x i ,y i is the coordinate of the i-th point;

[0117] Use L to measure the fitting effect. The smaller the loss value, the better the fitting effect. Finally, take the difference curve of the result with the best fitting effect as the final fitting difference curve.

[0118] The specific solution method to minimize the loss function by using the least squares method is as follows:

[0119]

[0120] Take the partial derivatives of a, b, and c and set them to zero to obtain the three simultaneous equations mentioned above, where K is the number of points. Then solve the system of equations to determine a, b, and c;

[0121] After obtaining the difference curve of each station, it is assumed that the actual flow value during the prediction process is y a , then the actual flow rate of the material flow obtained after correction is y r =y a +y d .

[0122] The terminal device adopts an embedded edge computing terminal device.

Claims

1. A method for automatically calibrating a multi-station material flow rate detection device on a conveyor belt, characterized by: The method is based on the dynamic statistics of industrial material flow based on the fusion of detection results of multi-site material flow equipment. By deploying multiple material detection devices on the conveyor belt, multiple sites are formed in the form of an array, and the spatial distribution characteristic data of material flow is collected in real time. Data statistical analysis is then used to estimate and cluster the multi-source measurement data. Specifically, the method first collects material flow data simultaneously with multiple devices, then calculates the time offset of each measurement value based on the material time series characteristics, and finally obtains an accurate flow estimate value through k-nearest neighbor kernel estimation and clustering processing algorithm.

2. The automatic calibration method for a multi-station material flow detection device on a conveyor belt according to claim 1, characterized in that: The detection equipment installation point is to install material detection equipment at each material transfer station to obtain the real-time material flow value of a certain inlet or outlet of the current station; Material detection equipment includes high-speed cameras, line laser generators, and terminal equipment.

3. The automatic calibration method for a multi-station material flow detection device on a conveyor belt according to claim 2, characterized in that: The automatic correction method first performs statistics and data filtering on the material flow value provided by the detection equipment to obtain the material flow value curve of each site, then calculates the time offset between each site and the upstream site, aligns the data of each site with the data of the upstream site in time, and finally performs k-nearest neighbor kernel estimation and clustering processing to obtain the final true value estimation curve. Then, a difference curve fitting calculation is performed based on the true value estimation curve and the original curve to obtain the difference curve of each site. After that, a correction calculation is performed based on the difference curve of each site to obtain the corrected curve.

4. The automatic calibration method for a multi-station material flow detection device on a conveyor belt according to claim 3, characterized in that: The automatic correction method comprises the following steps: Step 1: Data statistics and preprocessing; Step 2: Calculate the time offset of each site; Step 3: k-nearest neighbor kernel fitting estimation and clustering processing; Step 4: Site material flow correction.

5. The automatic calibration method for a multi-station material flow rate detection device on a conveyor belt according to claim 4, characterized in that: In step 1, the data is filtered. Let the data of a single site within a period of time be f(t), (m≤t≤n). Assuming that there are k data within this time range, the filtered result y(t) = median(f(t),…,f(t+k)). These data are continuously collected, but the subsequent time offset calculation and flow value correction are only performed within 1 hour after each conveyor belt operation line is started, so that different conveyor belt operation lines do not interfere with each other.

6. The automatic calibration method for a multi-station material flow rate detection device on a conveyor belt according to claim 4, characterized in that: In step 2, the flow rate value obtained by the material detection at a certain upstream station is used as the reference time stamp t0. The time when the material arrives at the N-1 downstream stations is t1, t2, ..., t N-1 The time offset between each downstream station and the most upstream station is recorded as Δt i =t i -t0; record the material of each downstream station between t0 and t N-1 The traffic data at timestamp x in time is f i (x), and the random error caused by other factors is ε i , with the following relationship: f i (x)=f 最上游站点 (x+Δt i )+ε i To calculate the time offset Δt of a downstream station relative to the most upstream station, use a time window with a preset reasonable size (Size). Find Δt within the range above and below Size. If the following relationship is met, the traffic in this window is too small to be used for time offset calculation. This special case needs to be excluded. Then, this window performs a Fast Fourier Transform-based cross-correlation coefficient calculation on the flow data for Size upstream and downstream station flow values. The calculation process is as follows: a downstream station and the most upstream station are used for calculation, where i is the timestamp increment; Perform fast Fourier transform (FFT) on the material flow data sequences of the upstream and downstream sites respectively: Where N is the sequence length. The data sequence of the downstream site is first flipped, which is equivalent to taking the conjugate of F(k) in the frequency domain. k is the data index in the frequency domain, j is the imaginary unit, F(k) is the frequency domain function of the fast Fourier transform of the downstream site in a pair of upstream and downstream sites, and FFT(f(N-1-x)) is the fast Fourier transform of the data sequence flipped at a downstream site. Similarly, F 最上游站点 (k) the meaning of; Calculate the frequency domain product C(k) and calculate the inverse fast Fourier transform IFFT: that is, convert the frequency domain product data back to the time domain, and take the real part of the result to obtain the cross-correlation coefficient R(x), and then centralize R(x) to obtain R shift (x), R shift (x) is the cross-correlation coefficient obtained by performing FFTShift(R(x)) on R(x), which moves the zero-frequency component to the center, making it easier to observe R(x). The specific formula and process are as follows: C(k)=F(k)*F 最上游站点 (k) Calculate the time offset: Where g is the value that makes R shift (x) gets the coordinate of the maximum value, that is, R shift (g) = max(R shift (x)), Size is the data window size for the time offset calculation process mentioned above; through the above process, all Δt i .

7. The automatic calibration method for a multi-station material flow detection device on a conveyor belt according to claim 4, characterized in that: The specific method of step 3 is: the time series of N stations after time axis offset correction is recorded as: f i (x);x=1,…,n;i=1,2,…,N Where x is the timestamp and i is the site number; Obtain the average function corresponding to N sites at timestamp x by weighted averaging Use k x close to x i The corresponding f(x i ) is the weighted average of k-nearest neighbor kernel estimate of I x,k ={i:x i is one of the k observations closest to x}, D(x,k)=max{|x i -x|,i∈I x,k } Use D(x,k) instead of window width h n , the obtained k-nearest neighbor kernel estimate is defined as follows: Take k = 5, K(u) as the normal kernel to perform k-nearest neighbor kernel estimation, and let the obtained estimation function As the initial cluster center of K-means clustering processing, denoted as g(x); Then do the following: Calculate the Euclidean distance d from all time functions to the center i Calculate the distance d i The mean μ d and standard deviation σ d , through 3σ d Criteria for determining outlier curves, If d i >μ d +3σ d , it is determined to be an outlier curve; If the number of outlier curves is zero, g(x) is the final true value curve; otherwise, the following operations are performed until the number of outlier curves is zero.

8. The automatic calibration method for a multi-station material flow detection device on a conveyor belt according to claim 7, characterized in that: In step 3, the cluster center g(x) is recalculated using the non-outlier curve, and the previous g(x) is recorded as g(x) old Used for multi-class judgment calculation; if the distance between these non-outlier curves and the new center g(x) is greater than 3σ d If the number of outlier curves determined by the criterion is zero, the outlier curve removed is output as an abnormal curve, and this g(x) is used as the final true value curve; In multi-class judgment, set the inter-class difference threshold θ = 2σ d , if |g(x)–g(x) old |>θ·|g(x) old |, it is determined that there are multiple categories. At this time, the true value estimation fails, indicating that the data of N sites are not similar and the true value result cannot be obtained.

9. The automatic calibration method for a multi-station material flow detection device on a conveyor belt according to claim 4, characterized in that: The specific method of step 4 is as follows: after obtaining the true value fitting result curve, the deviation between each station and the true value fitting result curve will form the difference points of each station. These difference points are used to fit the quadratic polynomial function, and multiple fittings are performed to obtain the fitted difference curve: y d =ax 2 +nx+c Where a, b, and c are the coefficients of the quadratic term, linear term, and constant term of the difference curve to be solved respectively; Define the loss function: x i ,y i is the coordinate of the i-th point; Use L to measure the fitting effect. The smaller the loss value, the better the fitting effect. Finally, take the difference curve of the result with the best fitting effect as the final fitting difference curve. The specific solution method to minimize the loss function by using the least squares method is as follows: Take the partial derivatives of a, b, and c and set them to zero to obtain the three simultaneous equations mentioned above, where K is the number of points. Then solve the system of equations to determine a, b, and c; After obtaining the difference curve of each station, it is assumed that the actual flow value during the prediction process is y a , then the actual flow rate of the material flow obtained after correction is y r =y a +y d .

10. The automatic calibration method for multi-station material flow detection equipment on a conveyor belt according to claim 2, characterized in that: The terminal device adopts an embedded edge computing terminal device.

Images