A multi-source information fusion method for intelligent manufacturing

Abstract

The present invention discloses a multi-source information fusion method for intelligent manufacturing, and belongs to the technical field of intelligent manufacturing. The present invention proposes a weight allocation method based on inter-group differences and intra-group differences, and establishes a multi-source information fusion method suitable for intelligent manufacturing. The method steps include: using multi-sensor technology to obtain multi-dimensional information of samples; calculating the inter-group differences and intra-group differences of different sensor information of samples; determining the weight of each sensor according to the ratio of the inter-group differences and intra-group differences of samples; using information fusion strategy, combined with the allocation strategy of each sensor weight, to achieve multi-source information fusion. The present invention creatively proposes a weight allocation method based on Mahalanobis distance, Euclidean distance and variance analysis. In addition, the weight allocation method of inter-group differences and intra-group differences is introduced for the first time to achieve multi-source information fusion in the intelligent manufacturing production process.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent manufacturing, and in particular relates to a multi-source information fusion method in an intelligent manufacturing production process based on a weight allocation strategy of between-group differences and within-group differences. Background Art

[0002] Intelligent manufacturing is a manufacturing process that integrates artificial intelligence and automation technology and manufacturing technology. Traditional Chinese medicine, agricultural products, food, etc. are all characterized by complex ingredients, multiple targets, and multiple pathways of action. A comprehensive and reliable evaluation system is an important part of intelligent manufacturing applications. Multi-source information fusion, coordination, and complementary multi-sensor information can improve system reliability.

[0003] The determination of sensor weights is a key issue in multi-source information fusion algorithms. The traditional method for determining weights is weighted averaging, which means that different sensors are assigned the same weight factor. However, the weighted averaging method does not take into account the properties of each sensor. In fact, even sensors of the same type may have different weights during the measurement process. A weight allocation strategy based on the properties of the sample itself needs to be established urgently.

[0004] The present invention creatively proposes a weight allocation method based on Mahalanobis distance, Euclidean distance and variance analysis. In addition, it introduces a weight allocation strategy based on inter-group differences and intra-group differences for the first time to realize multi-source information fusion in the intelligent manufacturing production process. Summary of the invention

[0005] The purpose of the present invention is to provide a weight allocation strategy based on inter-group differences and intra-group differences, and to establish a multi-source information fusion method suitable for intelligent manufacturing.

[0006] To achieve the above object, the method comprises:

[0007] Step 1: Obtain sample information through multi-sensor technology and calculate the inter-group and intra-group differences of sample information from different sensors;

[0008] Step 2: Determine the weight of each sensor based on the ratio of the difference between sample groups and the difference within the sample group;

[0009] Step 3: Adopt multi-source information fusion strategy and combine the allocation weights of each sensor to achieve multi-source information fusion.

[0010] The evaluation of the differences between and within different sensor information groups of the present invention includes the following methods:

[0011] (1) The sample original data multi-dimensional space Euclidean / Mahalanobis distance evaluation method includes the following steps:

[0012] Step 1: Test the properties of each batch of samples, repeat the test for each sample at least three times, and take the average value;

[0013] Step 2: Calculate the maximum Euclidean / Mahalanobis distance of the attributes of each batch of samples to characterize the intra-group differences of each batch of samples;

[0014] Step 3: Calculate the maximum Euclidean / Mahalanobis distance of each batch center to characterize the intergroup differences of samples from different batches;

[0015] Step 4: Establish the Euclidean / Mahalanobis distance evaluation method in multidimensional space of raw data to obtain the differences between and within groups of samples.

[0016] (2) The Euclidean / Mahalanobis distance evaluation method of the sample principal component space includes the following steps:

[0017] Step 1: Test the properties of each batch of samples, repeat the test for each sample at least three times, and take the average value;

[0018] Step 2: Use principal component analysis to reduce the dimension and obtain the characteristic variables of each sensor;

[0019] Step 3: Calculate the maximum Euclidean / Mahalanobis distance of the characteristic variables of each batch of samples to characterize the intra-group differences of each batch of samples.

[0020] Step 4: Calculate the maximum Euclidean / Mahalanobis distance of the center of each batch of characteristic variables to characterize the inter-group differences of samples from different batches;

[0021] Step 5: Establish the principal component analysis space Euclidean / Mahalanobis distance evaluation method to obtain the inter-group and intra-group differences of samples.

[0022] (3) Sample multivariate analysis of variance evaluation method, including the following steps:

[0023] Step 1: Test the properties of each batch of samples, repeat the test for each sample at least three times, and take the average value;

[0024] Step 2: Use principal component analysis to reduce the dimension and obtain the characteristic variables of each sensor;

[0025] Step 3: Calculate the within-group variance and between-group variance of each batch of samples;

[0026] Step 4: Establish a multivariate analysis of variance evaluation method to obtain the differences between and within groups of samples.

[0027] The ratio of the inter-group difference to the intra-group difference of the present invention represents the ability of the sensor to distinguish between samples between groups and samples within a group. The larger the ratio, the better the sensor's ability to distinguish, and the greater the sensor's weight. The ratio of the inter-group difference to the intra-group difference is proportional to the sensor's weight.

[0028] The multi-source information fusion of the present invention includes: data layer fusion, feature layer fusion and decision layer fusion.

[0029] (1) The method of combining data layer fusion with weight strategy includes the following steps:

[0030] Step 1: preprocess the spectrum using one or more of the following methods: derivative, smoothing, standard normal transformation (SNV), multivariate scatter correction (MSC), standardization, and normalization;

[0031] Step 2: Obtain information of different dimensions based on k different sensors after spectral preprocessing, and use the original data as the input variable of data layer fusion / the feature data obtained by principal component analysis PCA as the input variable of feature layer fusion;

[0032] Step 3: Standardize the input variables of the transformed data layer / feature layer fusion to obtain data of the same dimension from different sensors;

[0033] Step 4: Obtain the weights of k sensors based on the ratio of the difference between sensor groups and the difference within the sensor group;

[0034] Step 5: Establish a mathematical relationship model of the weight-based data layer fusion input variables.

[0035] (2) The method of combining the weight strategy at the decision layer includes the following steps:

[0036] Step 1: preprocess the spectrum using one or more of the following methods: derivative, smoothing, standard normal transformation SNV, multivariate scatter correction MSC, standardization, and normalization;

[0037] Step 2: Use the preprocessed input variables to establish mathematical relationship models for each sensor;

[0038] Step 3: Obtain the weights of k sensors based on the ratio of the difference between sensor groups and the difference within the sensor group;

[0039] Step 4: The mathematical relationship model of each sensor is combined with its weight to represent the result of decision layer fusion.

[0040] The mathematical relationship model included in the mathematical relationship model for multi-source information fusion provided by the present invention is applicable to all methods in machine learning.

[0041] In summary, the present invention provides a method for multi-source information fusion in intelligent manufacturing based on a weight allocation strategy of inter-group and intra-group differences.

[0042] The method of the present invention has the following advantages:

[0043] The present invention uses multi-sensor technology to obtain multi-dimensional information of samples, calculates the inter-group and intra-group differences of different sensor information of samples, determines the weight of each sensor according to the ratio of the inter-group and intra-group differences of samples, adopts information fusion strategy, and combines the allocation strategy of each sensor weight to realize multi-source information fusion. The detection method constructed has strong specificity, and the detection results are accurate and stable.

[0044] The present invention creatively proposes a weight allocation method based on Mahalanobis distance, Euclidean distance and variance analysis. In addition, it introduces for the first time a weight allocation method for inter-group differences and intra-group differences in feature space to achieve multi-source information fusion in the intelligent manufacturing production process. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 Flowchart of multi-source information fusion method; (a) data layer / feature layer fusion, (b) decision fusion

[0046] Figure 2 (a) Vis spectrum of bezoar powder, (b) Vis spectrum of SNV pretreatment, (c) NIR spectrum of bezoar powder, (d) NIR spectrum of SNV pretreatment, (e) chromaticity value of bezoar powder;

[0047] Figure 3 Multivariate statistical process control MSPC model based on Vis, NIR and color value methods; (a) Vis method Hotelling T 2 Statistics, (b) Vis method prediction error sum of squares SPE statistics, (c) NIR method Hotelling T 2 Statistics, (d) NIR method SPE statistics, (e) color value method Hotelling T 2 Statistics, (f) Color value method SPE statistics.

[0048] Figure 4 Multivariate statistical process control MSPC model based on NIR and MIR data layer fusion; (a) HotellingT 2 Statistics, (b) SPE statistics. DETAILED DESCRIPTION

[0049] The following is further described with reference to the accompanying drawings and implementation examples, but the present invention is not limited thereto.

[0050] Example 1: Application of an intelligent manufacturing decision-making layer fusion method in the quality evaluation of precious medicines

[0051] (1) Multi-source information collection of bezoar powder

[0052] The schematic diagram of the decision-making layer fusion method of the precious fine medicine bezoar powder is as follows Figure 1As shown. The three-dimensional information of visible spectrum Vis, near infrared spectrum NIR and chromaticity value of 30 batches of bezoar powder was collected, as follows: the visible spectrum of 90 samples was collected by visible-near infrared spectrometer, the spectral range was 400-800nm, the resolution was 0.5nm, 32 scans were made, 4 areas were collected for each sample, and the average was calculated; the NIR spectrum of 90 samples was collected by near infrared spectrometer, the spectral range was 1100-2400nm, the resolution was 0.5nm, 32 scans were made, 4 areas were collected for each sample, and the average was calculated; the chromaticity value of 90 samples was measured by electronic eye, the light source was D65, the viewing angle was 10°, 3 areas were collected for each sample, and the average was calculated.

[0053] Figure 2 .a is the original Vis spectrum of bezoar powder, Figure 2 .b is the spectrum of bezoar powder after SNV pretreatment. SNV has the function of eliminating surface scattering and optical path transformation of solid particles. Figure 2 .c is the original NIR spectrum of bezoar powder. The main characteristic bands of the sample NIR spectrum are assigned as follows: the absorption bands near 2300nm, 2100nm and 1900nm are the combined frequency absorption of CH, NH and OH stretching vibrations; the absorption bands near 1700nm, 1500nm and 1480nm are the first-frequency absorption of CH, NH and OH stretching vibrations; the absorption near 1200nm is the second-frequency absorption of CH stretching vibration. Figure 2 .d is the spectrum of the NIR spectrum of Tongren Niuhuang Qingxin Pills after SNV preprocessing. Figure 2 .e is the chromaticity value of the bezoar powder sample, where L, a and b are the coordinates of the sample in the color space, L represents black and white, a represents red and green, and b represents yellow and blue. Figure 2 .e It can be seen that the chromaticity values ​​of samples from different batches are basically the same.

[0054] (2) Establishment of the MSPC model for multivariate statistical process control based on Vis, NIR and chromaticity values

[0055] Decision fusion builds a model based on the data of each sensor, and combines the weight distribution strategy to achieve fusion in the decision center. Based on the quality evaluation of the precious medicine bezoar powder, the MSPC method is used to establish a mathematical relationship model based on the three sensors of Vis, NIR and color value data. The optimal number of principal components of the model established by bezoar powder based on Vis, NIR and electronic eye sensors is 2. The spectral interpretation rates of the first two principal components of the Vis spectrum are 92.51% and 4.93% respectively; the interpretation rates of the first two principal components of the NIR spectrum are 90.43% and 6.03% respectively; the interpretation rates of the first two principal components of the color value are 77.90% and 21.07% respectively. Figure 3is the MSPC model based on principal component analysis for Vis, NIR and chromaticity values ​​of bezoar powder, in which Hotelling T 2 and SPE are two parameter indicators of the MSPC model. 2 Samples that fall below the 95% control limit of SPE are considered abnormal samples.

[0056] The Vis spectrum of bezoar powder was preprocessed by SNV, and then the MSPC model was established. Figure 3 .a-3.b is the MSPC model Hotelling T 2 and SPE control charts, the thresholds of the control charts are 6.34 and 0.05 respectively. There are three samples outside the two confidence intervals, namely samples 88, 89 and 90. Figure 3 .c-3.d is the MSPC model of bezoar powder after SNV pretreatment, Hotelling T 2 The threshold values ​​of the and SPE control charts are 6.34 and 0.45 respectively. There are four samples that exceed the two confidence intervals, namely samples 82, 88, 89 and 90. Figure 3 .e-3f is the MSPC model of the chromaticity value of bezoar powder, T 2 The threshold values ​​of the and SPE control charts are 6.34 and 0.65 respectively. There are 8 samples outside the two confidence intervals, namely samples 2, 31, 58, 59, 60, 67, 70 and 90.

[0057] (3) Weight strategy based on decision-level fusion method

[0058] The role of different sensors in data fusion depends on the performance of their own sensors. The multivariate variance analysis evaluation method is used to calculate the weight of each sensor using the ratio of the variance between groups to the variance within groups. The number of principal components of Vis, NIR and chromaticity values ​​is selected as 2. Table 1 shows the weights of different sensors based on α. ​​Through multivariate variance analysis, the α of the Vis spectrum is 184.5177, the α of the Vis sensor is in the middle, and its corresponding weight is 0.2816; the α of the NIR spectrum is 460.1062, the α value of the NIR sensor is the largest, and the corresponding weight is also the largest at 0.7020; compared with the Vis and NIR methods, the α of the chromaticity value is the smallest at 10.7412, and the weight of the electronic eye sensor is the lowest at 0.0164. In summary, the weight of the NIR sensor is relatively large. On the one hand, it may be because the NIR sensor contains rich characteristic absorption compared with the Vis and electronic eye sensors, and on the other hand, it may be because the signal-to-noise ratio of the NIR sensor is relatively high, which can more obviously distinguish between samples between groups and samples within groups.

[0059] Table 1 Weights of different methods

[0060]

[0061] (4) Quality evaluation of bezoar powder using decision fusion method combined with weight strategy

[0062] The MSPC mathematical relationship model based on the Vis sensor for bezoar powder identified three abnormal samples, namely, samples 88, 89 and 90. The weight of the Vis sensor is 0.2816, that is, the above three samples as abnormal samples received 0.2816 votes; the MSPC model based on the NIR sensor for bezoar powder identified four abnormal samples, namely, samples 82, 88, 89 and 90. The weight of the NIR sensor is 0.7020, that is, the above four samples as abnormal samples received 0.7020 votes; the MSPC model based on the electronic eye sensor for bezoar powder identified eight abnormal samples (Table 2), however, the weight of these eight abnormal samples based on chromaticity value was only 0.0164, which did not play a key role in the sensor data fusion. For the quality evaluation of bezoar powder, the chromaticity value parameter can be ignored. According to the results of the Vis and NIR sensors, and their corresponding weights, it can be concluded that there are 4 abnormal samples based on the decision fusion strategy, namely samples 82, 88, 89 and 90.

[0063] Table 2 Results of decision fusion

[0064]

[0065] In summary, both the Vis sensor and the NIR sensor use the SNV preprocessing method, while the electronic eye sensor is not preprocessed. All three sensors use two principal components to establish the MSPC model based on the Vis, NIR sensor and electronic eye sensor. Furthermore, the weight of each sensor is determined by the ratio of the inter-group variance and the intra-group variance through multivariate variance analysis. The electronic eye sensor does not play a key role in the data fusion of the quality evaluation of bezoar powder because its weight is too low. The information advantages of different sensors complement each other. Relying on the information fusion of the Vis sensor and the NIR sensor, accurate and reliable quality evaluation results are obtained.

[0066] Example 2: Application of an intelligent manufacturing data layer fusion method in the quality evaluation of large honey pills

[0067] The schematic diagram of the big honey pill data layer fusion method is as follows Figure 1As shown. The near-infrared NIR spectrum and mid-infrared MIR spectrum of 30 batches of Niuhuang Qingxin Pills were collected in two dimensions. The measurement range of the NIR spectrum is 1100-2400nm, the resolution is 0.5nm, and the original spectrum data consists of 2601 data points; the measurement range of the MIR spectrum is 600-4000nm, the resolution is 2nm, and the original spectrum data consists of 1701 data points. Each sample was measured four times and the average value was calculated.

[0068] The inter-group and intra-group differences were calculated based on the original multidimensional space Euclidean distance evaluation method, and the quality of Niuhuang Qingxin Pills was evaluated by data layer fusion method. Both NIR and MIR spectra were preprocessed by SNV to obtain the spectral information of the two sensors after preprocessing. First, the Euclidean distance of each batch of samples based on NIR and MIR spectra was calculated respectively, and the maximum Euclidean distance within the NIR spectrum group was determined to be 2.2423, and the maximum Euclidean distance within the MIR spectrum group was 0.9477. The Euclidean distance of the center of each batch was further calculated, and the maximum Euclidean distance between NIR spectra groups was determined to be 3.7554, and the maximum Euclidean distance between MIR spectra groups was 0.9477. Finally, the ratios of the inter-group differences and intra-group differences of NIR and MIR sensors were 1.6748 and 2.1587, respectively, and the weights of the two sensors were obtained to be 0.4369 and 0.5631.

[0069] The quality of Niuhuang Qingxin Pills was evaluated by using a weighted data layer fusion method. First, the data from the two sensors were normalized to obtain data of the same dimension. The data were further combined with the weights to obtain the input variables of the weighted data layer fusion. The MSPC mathematical relationship model of 30 batches of Rentongren Niuhuang Qingxin Pills was established. The results are as follows: Figure 4 The results show that the weight-based data fusion strategy is used to find that the quality of samples 16, 18, 27, 28, 41 and 66 has large fluctuations. The multi-source information fusion strategy and the joint action of multiple sensors can realize the quality control of Chinese medicine honey pills.

Claims

1. A multi-source information fusion method for quality evaluation of precious and fine drugs, characterized in that: Different sensor information is assigned weights through the ratio of inter-group differences and intra-group differences to achieve multi-source information fusion. The precious fine medicine is bezoar powder. The specific steps are as follows: Step 1: Collect information from multiple sources about bezoar powder: The three-dimensional information of visible spectrum Vis, near-infrared spectrum NIR and chromaticity value of 30 batches of bezoar powder was collected, as follows: the visible spectrum of 90 samples was collected by visible near-infrared spectrometer, the spectral range was 400-800nm, the resolution was 0.5nm, and it was scanned 32 times. Four areas were collected for each sample and the average was calculated; the NIR spectrum of 90 samples was collected by near-infrared spectrometer, the spectral range was 1100-2400nm, the resolution was 0.5nm, and it was scanned 32 times. Four areas were collected for each sample and the average was calculated; the chromaticity values ​​of 90 samples were measured by electronic eye, the light source was D65, the viewing angle was 10°, and three areas were collected for each sample and the average was calculated; Step 2: Establishment of multivariate statistical process control MSPC model based on Vis, NIR and chromaticity values: Based on the quality evaluation of the precious fine medicine bezoar powder, the MSPC method was used to establish a mathematical relationship model based on the three sensors of Vis, NIR and color value data. The optimal number of principal components of the model established by bezoar powder based on Vis, NIR and electronic eye sensors itself was 2. The spectral interpretation rates of the first two principal components of the Vis spectrum were 92.51% and 4.93% respectively; the interpretation rates of the first two principal components of the NIR spectrum were 90.43% and 6.03% respectively; the interpretation rates of the first two principal components of the color value were 77.90% and 21.07% respectively, among which Hotelling T 2 and SPE are two parameter indicators of the MSPC model, which exceed Hotelling T 2 or the 95% control limit of SPE, are considered abnormal samples; Step 3: Weight strategy based on decision layer fusion method: The role of different sensors in data fusion depends on the performance of their own sensors. The multivariate variance analysis evaluation method is used to calculate the weight of each sensor using the ratio of the between-group variance to the within-group variance α. The number of principal components of Vis, NIR and chromaticity values ​​are all selected as 2. The following table shows the weights of different sensors based on α. Through multivariate analysis of variance, the α of the Vis spectrum is 184.5177, the α of the Vis sensor is in the middle, and its corresponding weight is 0.2816; the α of the NIR spectrum is 460.1062, the α value of the NIR sensor is the largest, and the corresponding weight is also the largest at 0.7020; compared with the Vis and NIR methods, the α of the chromaticity value is the smallest at 10.7412, and the weight of the electronic eye sensor is the lowest at 0.0164; Step 4: Quality evaluation of bezoar powder using decision fusion method combined with weight strategy: The MSPC mathematical relationship model based on the Vis sensor of bezoar powder identified three abnormal samples, namely samples 88, 89 and 90, and the weight of the Vis sensor was 0.2816, that is, the above three samples received 0.2816 votes as abnormal samples; the MSPC model based on the NIR sensor of bezoar powder identified four abnormal samples, namely samples 82, 88, 89 and 90, and the weight of the NIR sensor was 0.7020, that is, the above four samples received 0.7020 votes as abnormal samples; The MSPC model established based on the electronic eye sensor identified 8 abnormal samples, namely samples 2, 31, 58, 59, 60, 67, 70 and 90. However, the weight of these 8 abnormal samples based on the chromaticity value is only 0.0164, which does not play a key role in the sensor data fusion. For the quality evaluation of bezoar powder, the chromaticity value parameter is ignored. According to the results of the Vis and NIR sensors and their corresponding weights, it is concluded that there are 4 abnormal samples based on the decision fusion strategy, namely samples 82, 88, 89 and 90, as shown in the following table:

2. The application of the method according to claim 1 in the quality evaluation of precious fine medicines, characterized in that: The precious fine medicine is bezoar powder.

Images