A method and system for monitoring the quality of cold-drawn tungsten wire based on big data

Abstract

The invention discloses a tungsten wire cold-drawing quality monitoring method and system based on big data, belonging to the technical field of tungsten wire cold-drawing quality monitoring. The invention collects multi-dimensional data of tension, speed, temperature and vibration, comprehensively reflects various aspects of the tungsten wire cold-drawing process, respectively extracts mean, maximum, minimum and stable values, constructs mean sequence, maximum sequence, minimum sequence and stable value sequence, divides feature sequence into early stage, middle stage and late stage, constructs fusion features, realizes evaluation of tungsten wire cold-drawing quality from three stages, and then adopts a multi-scale parallel gap compensation network to process early stage fusion features, middle stage fusion features and late stage fusion features, captures data change features of three periods, and improves the tungsten wire cold-drawing quality monitoring accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of tungsten wire cold-drawing quality monitoring, and in particular to a tungsten wire cold-drawing quality monitoring method and system based on big data. Background Art

[0002] In modern manufacturing, tungsten wire is widely used in lighting, electronics, aerospace and other fields due to its excellent high temperature performance and strength. The production of tungsten wire usually requires cold drawing to make it into filaments of different diameters to meet various application requirements. However, how to ensure and monitor product quality during the cold drawing process of tungsten wire has always been a challenge.

[0003] The cold drawing process is affected by many factors, such as drawing force, drawing speed, temperature, mechanical vibration, etc. If these factors are not effectively monitored and adjusted, the quality of tungsten wire products may be unstable, resulting in problems such as breakage and surface defects. Therefore, real-time monitoring of these parameters and analysis of their relationship are crucial to improving the quality control level of tungsten wire.

[0004] In the prior art, usually only a single parameter (such as tension or speed) is monitored to evaluate the cold drawing quality, or a simple linear combination or threshold judgment of multiple parameters is difficult to reflect the changes in various parameters in the cold drawing process. Therefore, there is a problem of low accuracy in monitoring the quality of tungsten wire cold drawing. Summary of the invention

[0005] In view of the above-mentioned deficiencies in the prior art, the present invention provides a tungsten wire cold-drawing quality monitoring method and system based on big data to solve the problem of low precision of tungsten wire cold-drawing quality monitoring in the prior art.

[0006] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is: a tungsten wire cold drawing quality monitoring method based on big data, comprising the following steps:

[0007] S1. Collect the tensile force data, drawing speed data, temperature data and vibration data during the cold drawing process of tungsten wire;

[0008] S2, extracting the mean, maximum, minimum and stable values ​​of the tension data, drawing speed data, temperature data and vibration data, respectively, and constructing the mean sequence, maximum sequence, minimum sequence and stable value sequence;

[0009] S3. Construct early fusion features, mid-term fusion features and late fusion features according to the mean value sequence, maximum value sequence, minimum value sequence and stable value sequence of different time periods;

[0010] S4, the early fusion features, the mid-term fusion features and the late fusion features constitute training samples, and based on the loss function, the multi-scale parallel gap compensation network is trained to obtain the trained multi-scale parallel gap compensation network;

[0011] S5. The trained multi-scale parallel gap compensation network is used to evaluate the cold-drawn quality of tungsten wire.

[0012] Furthermore, the mean value sequence in S2 includes: the mean value of tension, the mean value of drawing speed, the mean value of temperature and the mean value of vibration; the maximum value sequence includes: the maximum value of tension, the maximum value of drawing speed, the maximum value of temperature and the maximum value of vibration; the minimum value sequence includes: the minimum value of tension, the minimum value of drawing speed, the minimum value of temperature and the minimum value of vibration; the stable value sequence includes: the stable value of tension, the stable value of drawing speed, the stable value of temperature and the stable value of vibration.

[0013] Furthermore, the calculation formulas for the tension stability value, the drawing speed stability value, the temperature stability value and the vibration stability value are all: , where s is the tension stability value, drawing speed stability value, temperature stability value or vibration stability value, x i is the i-th value in the tension data, drawing speed data, temperature data or vibration data, N is the length of the data, || is the absolute value operation, and i is a positive integer.

[0014] Furthermore, S3 includes the following sub-steps:

[0015] S31. According to the length of time for cold drawing of tungsten wire, the cold drawing process of tungsten wire is divided into: early stage, middle stage and late stage;

[0016] S32, constructing a previous fusion feature according to the previous mean sequence, maximum sequence, minimum sequence and stable value sequence;

[0017] S33, constructing a mid-term fusion feature according to the mid-term mean sequence, maximum sequence, minimum sequence and stable value sequence;

[0018] S34. Construct a late fusion feature according to the late mean sequence, maximum sequence, minimum sequence and stable value sequence.

[0019] Furthermore, S32, S33 and S34 each include the following sub-steps:

[0020] A1. Normalize the elements in the mean value sequence, maximum value sequence, minimum value sequence and stable value sequence;

[0021] A2. For each normalized sequence, construct a matrix to obtain the mean matrix, maximum matrix, minimum matrix and stable value matrix. The formula for the sequence matrix is: A = aT a, a is a sequence, T is a transpose operation, and A is the constructed matrix;

[0022] A3. Perform weighted processing on the mean matrix, maximum matrix, minimum matrix and stable value matrix to obtain fusion features.

[0023] Further, the multi-scale parallel gap compensation network in S4 includes: an early feature extraction unit, a mid-term feature extraction unit, a late feature extraction unit, a subtractor R1, a subtractor R2, a first gap coefficient calculation unit, a second gap coefficient calculation unit, an adder A1, a fully connected layer and a compensation layer;

[0024] The early feature extraction unit is used to extract features from the early fusion features to obtain a first feature matrix;

[0025] The mid-term feature extraction unit is used to extract features from the mid-term fusion features to obtain a second feature matrix;

[0026] The late feature extraction unit is used to extract features from the late fusion features to obtain a third feature matrix;

[0027] The subtractor R1 is used to subtract the first characteristic matrix from the second characteristic matrix to obtain a first difference matrix;

[0028] The first gap coefficient calculation unit is used to calculate the first gap coefficient according to the first difference matrix;

[0029] The subtractor R2 is used to subtract the third characteristic matrix from the second characteristic matrix to obtain a second difference matrix;

[0030] The second gap coefficient calculation unit is used to calculate the second gap coefficient according to the second difference matrix;

[0031] The adder A1 is used for adding the first characteristic matrix, the second characteristic matrix and the third characteristic matrix;

[0032] The fully connected layer is used to predict the initial evaluation value according to the result of addition by adder A1;

[0033] The compensation layer is used to compensate the initial evaluation value according to the first gap coefficient and the second gap coefficient to obtain the tungsten wire cold drawing quality score.

[0034] Furthermore, the expression of the first gap coefficient calculation unit or the second gap coefficient calculation unit is: , where d is the first gap coefficient or the second gap coefficient, r jis the j-th element in the first difference matrix or the second difference matrix, j is a positive integer, || is the absolute value operation, max is the maximum value, M is the number of elements in the first difference matrix or the second difference matrix, and σ is the Sigmoid activation function.

[0035] Furthermore, the expression of the compensation layer is: , where y is the cold-drawn quality score of tungsten wire, y0 is the initial evaluation value, tanh is the hyperbolic tangent activation function, d1 is the first gap coefficient, d2 is the second gap coefficient, ω d1 is the weight of d1, b d1 is the bias of d1, ω d2 is the weight of d2, b d2 is the bias of d2.

[0036] Furthermore, the expression of the loss function in S4 is: , where L is the loss function, y is the quality score of tungsten wire cold drawing, For label.

[0037] A tungsten wire cold drawing quality monitoring system based on big data, comprising: an acquisition subsystem, a sequence construction subsystem, a feature construction subsystem, a training subsystem and an evaluation subsystem;

[0038] The acquisition subsystem is used to collect tension data, wire drawing speed data, temperature data and vibration data during the cold drawing process of tungsten wire;

[0039] The sequence construction subsystem is used to extract the mean, maximum, minimum and stable values ​​of the tension data, the drawing speed data, the temperature data and the vibration data, respectively, and construct the mean sequence, the maximum sequence, the minimum sequence and the stable value sequence;

[0040] The feature construction subsystem is used to construct early fusion features, mid-term fusion features and late fusion features according to the mean value sequence, maximum value sequence, minimum value sequence and stable value sequence of different time periods;

[0041] The training subsystem is used to form a training sample by combining early fusion features, mid-term fusion features and late fusion features, and to train the multi-scale parallel gap compensation network based on the loss function to obtain a trained multi-scale parallel gap compensation network;

[0042] The evaluation subsystem is used to evaluate the cold-drawn quality of tungsten wire using a trained multi-scale parallel gap compensation network.

[0043] The beneficial effects of the present invention are as follows: the present invention collects multi-dimensional data of tension, speed, temperature and vibration, comprehensively reflects various aspects of the tungsten wire cold drawing process, and then respectively extracts the mean, maximum, minimum and stable values, constructs a mean sequence, a maximum sequence, a minimum sequence and a stable value sequence, divides the feature sequence into early, middle and late stages, constructs fusion features, and realizes the evaluation of the tungsten wire cold drawing quality from three stages, and then uses a multi-scale parallel gap compensation network to process the early fusion features, the middle fusion features and the late fusion features, captures the data change characteristics of the three periods, and improves the tungsten wire cold drawing quality monitoring accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a flow chart of a tungsten wire cold drawing quality monitoring method based on big data;

[0045] Figure 2 Schematic diagram of the structure of the multi-scale parallel gap compensation network. DETAILED DESCRIPTION

[0046] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0047] Example 1: Figure 1 As shown, a tungsten wire cold drawing quality monitoring method based on big data includes the following steps:

[0048] S1. Collect the tensile force data, drawing speed data, temperature data and vibration data during the cold drawing process of tungsten wire;

[0049] S2, extracting the mean, maximum, minimum and stable values ​​of the tension data, drawing speed data, temperature data and vibration data, respectively, and constructing the mean sequence, maximum sequence, minimum sequence and stable value sequence;

[0050] S3. Construct early fusion features, mid-term fusion features and late fusion features according to the mean value sequence, maximum value sequence, minimum value sequence and stable value sequence of different time periods;

[0051] S4, the early fusion features, the mid-term fusion features and the late fusion features constitute training samples, and based on the loss function, the multi-scale parallel gap compensation network is trained to obtain the trained multi-scale parallel gap compensation network;

[0052] S5. The trained multi-scale parallel gap compensation network is used to evaluate the cold-drawn quality of tungsten wire.

[0053] The mean value sequence in S2 includes: the mean value of tension, the mean value of drawing speed, the mean value of temperature and the mean value of vibration; the maximum value sequence includes: the maximum value of tension, the maximum value of drawing speed, the maximum value of temperature and the maximum value of vibration; the minimum value sequence includes: the minimum value of tension, the minimum value of drawing speed, the minimum value of temperature and the minimum value of vibration; the stable value sequence includes: the stable value of tension, the stable value of drawing speed, the stable value of temperature and the stable value of vibration.

[0054] The calculation formulas for the tension stability value, the drawing speed stability value, the temperature stability value and the vibration stability value are: , where s is the tension stability value, drawing speed stability value, temperature stability value or vibration stability value, x i is the i-th value in the tension data, drawing speed data, temperature data or vibration data, N is the length of the data, || is the absolute value operation, and i is a positive integer.

[0055] When calculating the stable value, the present invention reflects the fluctuation of each value according to the difference between each value and the average value, and reflects the stability of the pulling force, drawing speed, temperature and vibration.

[0056] In this embodiment, the tension stability value, the wire drawing speed stability value, the temperature stability value and the vibration stability value can also be calculated by variance and standard deviation.

[0057] The S3 comprises the following sub-steps:

[0058] S31. According to the length of time for cold drawing of tungsten wire, the cold drawing process of tungsten wire is divided into: early stage, middle stage and late stage;

[0059] S32, constructing a previous fusion feature according to the previous mean sequence, maximum sequence, minimum sequence and stable value sequence;

[0060] S33, constructing a mid-term fusion feature according to the mid-term mean sequence, maximum sequence, minimum sequence and stable value sequence;

[0061] S34. Construct a late fusion feature according to the late mean sequence, maximum sequence, minimum sequence and stable value sequence.

[0062] The S32, S33 and S34 all include the following sub-steps:

[0063] A1. Normalize the elements in the mean value sequence, maximum value sequence, minimum value sequence and stable value sequence;

[0064] A2. For each normalized sequence, construct a matrix to obtain the mean matrix, maximum matrix, minimum matrix and stable value matrix. The formula for the sequence matrix is: A = aT a, a is a sequence, T is a transpose operation, and A is the constructed matrix;

[0065] A3. Perform weighted processing on the mean matrix, maximum matrix, minimum matrix and stable value matrix to obtain fusion features.

[0066] The present invention divides the cold drawing process into the early stage, the middle stage and the late stage, so that the network model can capture the characteristic changes in different stages. Secondly, the four sequences of each period are processed to construct the fusion feature, and the comparability of different parameters is ensured by normalization. By constructing the autocorrelation matrix A = a T a effectively captures the correlation information within the sequence and adjusts the importance of each indicator through weighted processing, thereby improving the data representation ability and enhancing the network model's ability to recognize complex patterns.

[0067] like Figure 2 As shown, the multi-scale parallel gap compensation network in S4 includes: an early feature extraction unit, a mid-term feature extraction unit, a late feature extraction unit, a subtractor R1, a subtractor R2, a first gap coefficient calculation unit, a second gap coefficient calculation unit, an adder A1, a fully connected layer and a compensation layer;

[0068] The early feature extraction unit is used to extract features from the early fusion features to obtain a first feature matrix;

[0069] The mid-term feature extraction unit is used to extract features from the mid-term fusion features to obtain a second feature matrix;

[0070] The late feature extraction unit is used to extract features from the late fusion features to obtain a third feature matrix;

[0071] The subtractor R1 is used to subtract the first characteristic matrix from the second characteristic matrix to obtain a first difference matrix;

[0072] The first gap coefficient calculation unit is used to calculate the first gap coefficient according to the first difference matrix;

[0073] The subtractor R2 is used to subtract the third characteristic matrix from the second characteristic matrix to obtain a second difference matrix;

[0074] The second gap coefficient calculation unit is used to calculate the second gap coefficient according to the second difference matrix;

[0075] The adder A1 is used for adding the first characteristic matrix, the second characteristic matrix and the third characteristic matrix;

[0076] The fully connected layer is used to predict the initial evaluation value according to the result of addition by adder A1;

[0077] The compensation layer is used to compensate the initial evaluation value according to the first gap coefficient and the second gap coefficient to obtain the tungsten wire cold drawing quality score.

[0078] In this embodiment, the early feature extraction unit, the mid-term feature extraction unit, and the late feature extraction unit use convolutional neural networks.

[0079] The present invention adopts an early feature extraction unit, a mid-term feature extraction unit, and a late feature extraction unit to process early fusion features, mid-term fusion features, and late fusion features respectively, so as to realize comprehensive feature capture of different stages of the tungsten wire cold drawing process, and adopts an adder A1 to add the features of the three stages, so that the fully connected layer can integrate the elements output by the adder A1 to perform an initial evaluation, and then subtract the first feature matrix from the second feature matrix to obtain the first gap coefficient, and subtract the third feature matrix from the second feature matrix to obtain the second gap coefficient, which effectively quantifies the differences between different stages and enhances the sensitivity to dynamic changes of the process. According to the first gap coefficient and the second gap coefficient, the initial evaluation value is compensated to improve the calculation accuracy of the tungsten wire cold drawing quality score.

[0080] The expression of the first gap coefficient calculation unit or the second gap coefficient calculation unit is: , where d is the first gap coefficient or the second gap coefficient, r j is the j-th element in the first difference matrix or the second difference matrix, j is a positive integer, || is the absolute value operation, max is the maximum value, M is the number of elements in the first difference matrix or the second difference matrix, and σ is the Sigmoid activation function.

[0081] The present invention selects the mean and maximum values ​​of the elements in the difference matrix to reflect the gap situation.

[0082] The expression of the compensation layer is: , where y is the cold-drawn quality score of tungsten wire, y0 is the initial evaluation value, tanh is the hyperbolic tangent activation function, d1 is the first gap coefficient, d2 is the second gap coefficient, ω d1 is the weight of d1, b d1 is the bias of d1, ω d2 is the weight of d2, b d2 is the bias of d2.

[0083] The present invention assigns weights and biases to the first gap coefficient and the second gap coefficient respectively, calculates the influence of each on the score, and combines the initial evaluation value to improve the prediction accuracy of the tungsten wire cold-drawn quality score.

[0084] The expression of the loss function in S4 is: , where L is the loss function, y is the quality score of tungsten wire cold drawing, For label.

[0085] In this embodiment, the weights and biases in the multi-scale parallel gap compensation network are trained by the gradient descent method.

[0086] Embodiment 2: A tungsten wire cold drawing quality monitoring system based on big data, comprising: an acquisition subsystem, a sequence construction subsystem, a feature construction subsystem, a training subsystem and an evaluation subsystem;

[0087] The acquisition subsystem is used to collect tension data, wire drawing speed data, temperature data and vibration data during the cold drawing process of tungsten wire;

[0088] The sequence construction subsystem is used to extract the mean, maximum, minimum and stable values ​​of the tension data, the drawing speed data, the temperature data and the vibration data, respectively, and construct the mean sequence, the maximum sequence, the minimum sequence and the stable value sequence;

[0089] The feature construction subsystem is used to construct early fusion features, mid-term fusion features and late fusion features according to the mean value sequence, maximum value sequence, minimum value sequence and stable value sequence of different time periods;

[0090] The training subsystem is used to form a training sample by combining early fusion features, mid-term fusion features and late fusion features, and to train the multi-scale parallel gap compensation network based on the loss function to obtain a trained multi-scale parallel gap compensation network;

[0091] The evaluation subsystem is used to evaluate the cold-drawn quality of tungsten wire using a trained multi-scale parallel gap compensation network.

[0092] The specific implementation process of Example 2 is the same as that of Example 1.

[0093] The present invention collects multi-dimensional data of tension, speed, temperature and vibration, comprehensively reflects various aspects of the tungsten wire cold drawing process, and then respectively extracts the mean, maximum, minimum and stable values, constructs the mean sequence, maximum sequence, minimum sequence and stable value sequence, divides the feature sequence into early stage, middle stage and late stage, constructs fusion features, realizes the evaluation of the tungsten wire cold drawing quality from three stages, and then adopts a multi-scale parallel gap compensation network to process the early stage fusion features, middle stage fusion features and late stage fusion features, captures the data change characteristics of the three periods, and improves the tungsten wire cold drawing quality monitoring accuracy.

[0094] The present invention collects multi-dimensional data of tension, speed, temperature and vibration, comprehensively captures the key information of the cold drawing process, and provides a rich data basis for subsequent analysis. Secondly, by extracting statistical features such as mean, maximum, minimum and stable values, a variety of sequences are constructed to effectively extract significant features in the data and better characterize the dynamic characteristics of the cold drawing process. The feature sequences are divided into early, middle and late stages and fused, which reduces the amount of data and improves the time sensitivity of the model. The multi-scale parallel gap compensation network is adopted, which can simultaneously process features of different time scales, adapt to multi-scale dynamic changes in the cold drawing process, and reduce information loss through the gap compensation mechanism. Finally, the trained network model can evaluate the cold-drawn quality of tungsten wire in real time and detect potential problems in time.

[0095] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A tungsten wire cold drawing quality monitoring method based on big data, characterized in that: The following steps are involved: S1. Collect the tensile force data, drawing speed data, temperature data and vibration data during the cold drawing process of tungsten wire; S2, extracting the mean, maximum, minimum and stable values ​​of the tension data, drawing speed data, temperature data and vibration data, respectively, and constructing the mean sequence, maximum sequence, minimum sequence and stable value sequence; S3. Construct early fusion features, mid-term fusion features and late fusion features according to the mean value sequence, maximum value sequence, minimum value sequence and stable value sequence of different time periods; S4, the early fusion features, the mid-term fusion features and the late fusion features constitute training samples, and based on the loss function, the multi-scale parallel gap compensation network is trained to obtain the trained multi-scale parallel gap compensation network; S5. Use the trained multi-scale parallel gap compensation network to evaluate the cold-drawn quality of tungsten wire; The mean value sequence in S2 includes: the mean value of tension, the mean value of wire drawing speed, the mean value of temperature and the mean value of vibration; the maximum value sequence includes: the maximum value of tension, the maximum value of wire drawing speed, the maximum value of temperature and the maximum value of vibration; the minimum value sequence includes: the minimum value of tension, the minimum value of wire drawing speed, the minimum value of temperature and the minimum value of vibration; the stable value sequence includes: the stable value of tension, the stable value of wire drawing speed, the stable value of temperature and the stable value of vibration; The calculation formulas for the tension stability value, the drawing speed stability value, the temperature stability value and the vibration stability value are: , where s is the tension stability value, drawing speed stability value, temperature stability value or vibration stability value, x i is the i-th value in the tension data, drawing speed data, temperature data or vibration data, N is the length of the data, | | is the absolute value operation, and i is a positive integer; The S3 comprises the following sub-steps: S31. According to the length of time for cold drawing of tungsten wire, the cold drawing process of tungsten wire is divided into: early stage, middle stage and late stage; S32, constructing a previous fusion feature according to the previous mean sequence, maximum sequence, minimum sequence and stable value sequence; S33, constructing a mid-term fusion feature according to the mid-term mean sequence, maximum sequence, minimum sequence and stable value sequence; S34, constructing a late fusion feature according to the late mean sequence, maximum sequence, minimum sequence and stable value sequence; The S32, S33 and S34 all include the following sub-steps: A1. Normalize the elements in the mean value sequence, maximum value sequence, minimum value sequence and stable value sequence; A2. For each normalized sequence, construct a matrix to obtain the mean matrix, maximum matrix, minimum matrix and stable value matrix. The formula for the sequence matrix is: A = a T a, a is a sequence, T is a transpose operation, and A is the constructed matrix; A3. Perform weighted processing on the mean matrix, maximum matrix, minimum matrix and stable value matrix to obtain fusion features; The multi-scale parallel gap compensation network in S4 includes: an early feature extraction unit, a mid-term feature extraction unit, a late feature extraction unit, a subtractor R1, a subtractor R2, a first gap coefficient calculation unit, a second gap coefficient calculation unit, an adder A1, a fully connected layer and a compensation layer; The early feature extraction unit is used to extract features from the early fusion features to obtain a first feature matrix; The mid-term feature extraction unit is used to extract features from the mid-term fusion features to obtain a second feature matrix; The late feature extraction unit is used to extract features from the late fusion features to obtain a third feature matrix; The subtractor R1 is used to subtract the first characteristic matrix from the second characteristic matrix to obtain a first difference matrix; The first gap coefficient calculation unit is used to calculate the first gap coefficient according to the first difference matrix; The subtractor R2 is used to subtract the third characteristic matrix from the second characteristic matrix to obtain a second difference matrix; The second gap coefficient calculation unit is used to calculate the second gap coefficient according to the second difference matrix; The adder A1 is used for adding the first characteristic matrix, the second characteristic matrix and the third characteristic matrix; The fully connected layer is used to predict the initial evaluation value according to the result of addition by adder A1; The compensation layer is used to compensate the initial evaluation value according to the first gap coefficient and the second gap coefficient to obtain the tungsten wire cold drawing quality score; The expression of the first gap coefficient calculation unit or the second gap coefficient calculation unit is: , where d is the first gap coefficient or the second gap coefficient, r j is the j-th element in the first difference matrix or the second difference matrix, j is a positive integer, || is an absolute value operation, max is the maximum value, M is the number of elements in the first difference matrix or the second difference matrix, and σ is the Sigmoid activation function; The expression of the compensation layer is: , where y is the cold-drawn quality score of tungsten wire, y0 is the initial evaluation value, tanh is the hyperbolic tangent activation function, d1 is the first gap coefficient, d2 is the second gap coefficient, ω d1 is the weight of d1, b d1 is the bias of d1, ω d2 is the weight of d2, b d2 is the bias of d2; The expression of the loss function in S4 is: , where L is the loss function, y is the quality score of tungsten wire cold drawing, For label.

2. A tungsten wire cold drawing quality monitoring system based on big data, implemented based on the tungsten wire cold drawing quality monitoring method based on big data according to claim 1, characterized in that: include: Acquisition subsystem, sequence construction subsystem, feature construction subsystem, training subsystem and evaluation subsystem; The acquisition subsystem is used to collect tension data, wire drawing speed data, temperature data and vibration data during the cold drawing process of tungsten wire; The sequence construction subsystem is used to extract the mean, maximum, minimum and stable values ​​of the tension data, the drawing speed data, the temperature data and the vibration data, respectively, and construct the mean sequence, the maximum sequence, the minimum sequence and the stable value sequence; The feature construction subsystem is used to construct early fusion features, mid-term fusion features and late fusion features according to the mean value sequence, maximum value sequence, minimum value sequence and stable value sequence of different time periods; The training subsystem is used to form a training sample by combining early fusion features, mid-term fusion features and late fusion features, and to train the multi-scale parallel gap compensation network based on the loss function to obtain a trained multi-scale parallel gap compensation network; The evaluation subsystem is used to evaluate the cold-drawn quality of tungsten wire using a trained multi-scale parallel gap compensation network.

Images