An injection mold for a triple-proof lampshade with an ejection mechanism

Through deep learning algorithms, the drafting speed of injection molds is monitored and analyzed, adaptive control of the three-proof lampshade products is achieved, the problem of difficulty in controlling splicing lines and drafting speed is solved, and the product quality and durability are improved.

CN116061408BActive Publication Date: 2024-11-01CIXI YUANHUI LIGHTING ELECTRIC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202211546007.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-12-05
Publication Date
2024-11-01
Estimated Expiration
2042-12-05

AI Technical Summary

Technical Problem

When existing injection molds produce three-proof lampshades, the splicing wires at the ejection position lead to poor sealing performance, poor durability, and difficult to control the drafting speed, which easily leads to product damage.

Method used

Using deep learning-based artificial intelligence algorithms, we use drafting to monitor video and speed data, extract keyframes and dynamic features, perform adaptive control, and adjust drafting speed to avoid product damage.

Benefits of technology

The production quality of three-proof lampshade products is improved, sealing performance and durability are ensured, and the risk of product damage during the drafting process is reduced.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN116061408B_ABST
    Figure CN116061408B_ABST
Patent Text Reader

Abstract

The present application relates to the field of intelligent regulation and control, and specifically discloses a three-proof lamp shade injection mold with an ejection mechanism, which monitors the real-time state of demolding by utilizing an artificial intelligence algorithm based on deep learning, and extracts and analyzes the dynamic characteristics of the demolding speed, so as to adaptively control the demolding speed at the current time point, thereby avoiding product damage caused by demolding and ensuring the production quality of the three-proof lamp shade product.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of intelligent control, and more specifically, to a triple-proof lampshade injection mold with an ejection mechanism. Background Art

[0002] With the continuous advancement of injection molding technology, improving the qualification rate of injection molding products has become the goal pursued by enterprises. Therefore, the requirements for the structural design and control system of injection molds are getting higher and higher.

[0003] At present, one of the key issues restricting the qualified rate of injection molded products is that when the injection mold with a top block structure is used for production, the traces of the splicing line will be left at the ejection position of the injection molded product, forming a fault. Especially for injection molded products such as lampshades, for example, tri-proof lampshades, due to the unreasonable design of the position of the parting line between the top block and the movable mold insert of the existing injection mold, two splicing lines are left on the side wall of the product, forming a fault. When the lampshade and the lamp shell are assembled and poured with glue, the viscous glue has poor fluidity, which is easy to cause glue scraping, resulting in uneven glue pouring, which not only affects the sealing performance of the lamp product, but also the durability of the tri-proof lamp is poor, and stress cracking is easy to occur at the splicing line position.

[0004] In response to the above technical problems, Chinese patent CN 202607984U discloses a lampshade injection mold ejection mechanism, which is composed of a group of ejection units arranged around the injection mold cavity. The ejection mechanism reduces the processing difficulty caused by the sharp corners of the movable mold insert and the product flash phenomenon caused by the collapse of the mold, and avoids the reduction of the overall strength of the ejector block caused by the excessive angle of the ejector block.

[0005] However, in the actual injection molding process, it is found that the configuration of the ejection mechanism makes it difficult to demold. Once the demolding speed is not well controlled, it is easy to cause damage to the product in the mold.

[0006] Therefore, an optimized triple-proof lampshade injection mold with an ejection mechanism is desired, which can adaptively adjust the ejection speed based on the real-time state of the ejection to avoid product damage caused by the ejection. Summary of the invention

[0007] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides a three-proof lampshade injection mold with an ejection mechanism, which uses an artificial intelligence algorithm based on deep learning to monitor the real-time state of demolding, and extracts and analyzes the dynamic characteristics of the demolding speed, so as to adaptively control the demolding speed at the current time point, thereby avoiding product damage caused by demolding and ensuring the production quality of the three-proof lampshade product.

[0008] According to one aspect of the present application, there is provided a three-proof lampshade injection mold with an ejection mechanism, comprising: a mold draft monitoring unit for acquiring the mold draft speed at a plurality of predetermined time points within a predetermined time period and a mold draft monitoring video of the predetermined time period;

[0009] A key frame extraction unit, used to extract a plurality of demolding key frames from the demolding monitoring video;

[0010] A drafting process feature extraction unit, used for obtaining a drafting tracking feature map by subjecting the plurality of drafting key frames to a convolutional neural network model using a trained temporal attention mechanism;

[0011] A drafting speed feature extraction unit, used for arranging the drafting speeds of a plurality of predetermined time points within the predetermined time period according to the time dimension as a drafting input vector and then obtaining a multi-scale drafting feature vector through a multi-scale neighborhood feature extraction module completed by training;

[0012] A dimension reduction unit, used for performing global mean pooling processing on each feature matrix along the channel dimension of the draft tracking feature map to obtain a draft tracking feature vector;

[0013] a transfer unit for calculating a transfer matrix of the multi-scale draft feature vector relative to the draft tracking feature vector as a classification feature matrix; and a control result generation unit for passing the classification feature matrix through a trained classifier to obtain a classification result, wherein the classification result is used to indicate whether the draft speed at the current time point should be increased or decreased.

[0014] In the above-mentioned triple-proof lampshade injection mold with ejection mechanism, the key frame extraction unit is further used to extract the multiple demolding key frames from the demolding monitoring video using the difference frame method.

[0015] In the above-mentioned three-proof lamp shade injection mold with ejection mechanism, the demolding process feature extraction unit includes: an adjacent frame extraction subunit, used to extract the adjacent first frame and second frame from the multiple demolding key frames; a first image local feature extraction subunit, used to pass the first frame and the second frame through the first convolution layer and the second convolution layer of the convolutional neural network model respectively to obtain a first frame feature map corresponding to the first frame and a second frame feature map corresponding to the second frame; a time attention subunit, used to perform positional point multiplication on the first frame feature map and the second frame feature map and then pass the Softmax activation function to obtain a time attention map; a second image local feature extraction subunit, used to pass the first frame through the third convolution layer of the convolutional neural network model to obtain a local feature map; a time attention application subunit, used to perform positional point multiplication on the local feature map and the time attention map to obtain the demolding tracking feature map.

[0016] In the above-mentioned three-proof lampshade injection mold with ejection mechanism, the demolding speed feature extraction unit includes: a first scale feature extraction subunit, used to input the demolding input vector into the first convolution layer of the multi-scale neighborhood feature extraction module to obtain a first neighborhood scale demolding feature vector, wherein the first convolution layer has a first one-dimensional convolution kernel of a first length; a second scale feature extraction subunit, used to input the demolding input vector into the second convolution layer of the multi-scale neighborhood feature extraction module to obtain a second neighborhood scale demolding feature vector, wherein the second convolution layer has a second one-dimensional convolution kernel of a second length, and the first length is different from the second length; and a multi-scale feature fusion subunit, used to cascade the first neighborhood scale demolding feature vector and the second neighborhood scale temperature demolding feature vector to obtain the multi-scale demolding feature vector.

[0017] In the above-mentioned three-proof lampshade injection mold with ejection mechanism, the transfer unit is further used to: calculate the transfer matrix of the multi-scale draft feature vector relative to the draft tracking feature vector as a classification feature matrix using the following formula;

[0018] Wherein, the formula is:

[0019]

[0020] Wherein V1 represents the multi-scale draft feature vector, M2 represents the transfer matrix of the draft tracking feature vector, and M represents the classification feature matrix. Represents vector multiplication.

[0021] In the above-mentioned three-proof lampshade injection mold with ejection mechanism, the control result generation unit is further used to: use the classifier to process the classification feature matrix with the following formula to obtain the classification result, and the formula is:

[0022] softmax{(W n , B n ):...:(W1,B1)|Project(F)}, where Project(F) represents the projection of the classification feature matrix into a vector, W1 to W n is the weight matrix of each fully connected layer, B1 to B n Represents the bias matrix of each fully connected layer.

[0023] In the above-mentioned triple-proof lampshade injection mold with ejection mechanism, it also includes a training module for training the convolutional neural network model using the time attention mechanism, the multi-scale neighborhood feature extraction module and the classifier.

[0024] In the above-mentioned three-proof lampshade injection mold with ejection mechanism, the training module includes: a training data acquisition unit, used to obtain training data, the training data includes training demolding speeds at multiple predetermined time points within a predetermined time period and a training demolding monitoring video of the predetermined time period; a training key frame extraction unit, used to extract multiple training demolding key frames from the training demolding monitoring video; a training demolding process feature extraction unit, used to pass the multiple training demolding key frames through the convolutional neural network model using the time attention mechanism to obtain a training demolding tracking feature map; a training demolding speed feature extraction unit, used to arrange the training demolding speeds at multiple predetermined time points within the predetermined time period according to the time dimension as a training demolding input vector, and then pass it through the multi-scale neighborhood feature extraction module to obtain a training multi-scale demolding feature vector; a training dimension reduction unit, used to extract each feature of the training demolding tracking feature map along the channel dimension. The feature matrix is ​​subjected to global mean pooling processing to obtain a training draft tracking feature vector; a training transfer unit is used to calculate the transfer matrix of the training multi-scale draft feature vector relative to the training draft tracking feature vector as a training classification feature matrix; a classification loss unit is used to pass the training classification feature matrix through a classifier to obtain a classification loss function value; a context loss measurement unit is used to calculate a local scene measurement loss function value of context statistics of the training draft tracking feature vector, wherein the local scene measurement loss function value of context statistics is related to the statistical characteristics of the feature value set of all positions of the training draft tracking feature vector; and a training unit is used to train the convolutional neural network model using the temporal attention mechanism, the multi-scale neighborhood feature extraction module and the classifier with the weighted sum of the local scene measurement loss function value of context statistics and the classification loss function value as the loss function value.

[0025] In the above-mentioned three-proof lampshade injection mold with ejection mechanism, the context loss measurement unit is further used to: calculate the local scene measurement loss function value of the context statistics of the training draft tracking feature vector according to the following formula;

[0026] Wherein, the formula is:

[0027]

[0028] where v i represents the eigenvalue of each position of the training draft tracking feature vector, μ and σ respectively represent the mean and variance of the set of all eigenvalues ​​of the training draft tracking feature vector, and L is the length of the training draft tracking feature vector, exp(·) represents the exponential operation of a number, and the exponential operation of the number represents the calculation of the value of a natural exponential function with the number as the power.

[0029] According to another aspect of the present application, there is provided a method for injection molding a tri-proof lampshade with an ejection mechanism, comprising: acquiring the demoulding speeds at a plurality of predetermined time points within a predetermined time period and a demoulding monitoring video of the predetermined time period;

[0030] Extracting a plurality of draft key frames from the draft monitoring video;

[0031] The plurality of draft key frames are passed through a trained convolutional neural network model using a temporal attention mechanism to obtain a draft tracking feature map;

[0032] Arranging the drafting speeds of a plurality of predetermined time points within the predetermined time period according to the time dimension as a drafting input vector and then performing a multi-scale neighborhood feature extraction module trained to obtain a multi-scale drafting feature vector;

[0033] Performing global mean pooling processing on each feature matrix along the channel dimension of the draft tracking feature map to obtain a draft tracking feature vector;

[0034] calculating a transfer matrix of the multi-scale draft feature vector relative to the draft tracking feature vector as a classification feature matrix; and

[0035] The classification feature matrix is ​​passed through a trained classifier to obtain a classification result, and the classification result is used to indicate whether the draft speed at the current time point should be increased or decreased.

[0036] According to another aspect of the present application, an electronic device is provided, comprising: a processor; and a memory, wherein computer program instructions are stored in the memory, and when the computer program instructions are executed by the processor, the processor executes the method for injection molding a triple-proof lamp cover with an ejection mechanism as described above.

[0037] According to another aspect of the present application, a computer-readable medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the processor executes the method for injection molding a lampshade of a triple-proof lamp with an ejection mechanism as described above.

[0038] Compared with the prior art, the present application provides a three-proof lamp shade injection mold with an ejection mechanism, which monitors the real-time state of demolding by utilizing an artificial intelligence algorithm based on deep learning, and extracts and analyzes the dynamic characteristics of the demolding speed, so as to adaptively control the demolding speed at the current time point, thereby avoiding product damage caused by demolding and ensuring the production quality of the three-proof lamp shade product. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0040] Figure 1 The diagram illustrates an application scenario of a triple-proof lampshade injection mold with an ejection mechanism according to an embodiment of the present application;

[0041] Figure 2 A block diagram of a triple-proof lampshade injection mold with an ejection mechanism according to an embodiment of the present application is illustrated;

[0042] Figure 3 A block diagram of a triple-proof lampshade injection mold with an ejection mechanism according to an embodiment of the present application is illustrated;

[0043] Figure 4 The system architecture diagram of the inference module in the injection mold of the triple-proof lampshade with the ejection mechanism according to the embodiment of the present application is illustrated;

[0044] Figure 5 The block diagram of the demolding process feature extraction unit in the injection mold of the tri-proof lamp cover with the ejection mechanism according to the embodiment of the present application is illustrated;

[0045] Figure 6 The block diagram of the extraction unit of the demolding speed in the injection mold of the tri-proof lampshade with the ejection mechanism according to the embodiment of the present application is illustrated;

[0046] Figure 7 The system architecture diagram of the training module in the injection mold of the triple-proof lampshade with the ejection mechanism according to the embodiment of the present application is illustrated;

[0047] Figure 8 A flowchart of a method for injection molding a lampshade of a tri-proof lamp with an ejection mechanism according to an embodiment of the present application is illustrated; DETAILED DESCRIPTION

[0048] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0049] Scenario Overview

[0050] Correspondingly, in the technical solution of the present application, a three-proof lamp shade injection mold with data intelligence and an ejection mechanism is proposed, which can use an artificial intelligence algorithm based on deep learning to monitor the real-time state of demolding, and extract and analyze the dynamic characteristics of the demolding speed, so as to adaptively control the demolding speed at the current time point, thereby avoiding product damage caused by demolding and ensuring the production quality of the three-proof lamp shade products.

[0051] Specifically, in the technical solution of the present application, first, the demolding monitoring data is obtained, wherein the demolding monitoring data includes the demolding speed at multiple predetermined time points within a predetermined time period and the demolding monitoring video of the predetermined time period. Next, it should be understood that considering that there are many image frames in the demolding monitoring video of the predetermined time period, this brings difficulties to the subsequent demolding real-time state feature extraction. Therefore, in order to fully extract the demolding state implicit feature information in the demolding monitoring video of the predetermined time period while simplifying the number of parameters, multiple demolding key frames are further extracted from the demolding monitoring video. That is, in a specific example of the present application, the multiple demolding key frames can be extracted from the demolding monitoring video using the difference frame method.

[0052] Then, for the multiple demolding key frames extracted from the demolding monitoring video, a convolutional neural network with excellent performance in extracting implicit features of images is used to mine the implicit features of the multiple demolding key frames. However, considering that when adaptively controlling the demolding speed, it is necessary to pay more attention to the dynamic features of the demolding state in the time dimension in order to more accurately detect the real-time state of the demolding. Therefore, in the technical solution of the present application, a convolutional neural network model with a time attention mechanism is used to process the multiple demolding key frames to extract the local implicit feature distribution information in the multiple demolding key frames, that is, the dynamic change characteristics of the demolding state, so as to obtain a demolding tracking feature map. It should be understood that the time attention mechanism allows the convolutional network model to focus its learning on the more important time dimension area of ​​the sample, thereby improving the efficiency of the network model learning samples and thus improving the accuracy of subsequent classification.

[0053] Furthermore, for the drafting speeds at multiple predetermined time points within the predetermined time period, they are arranged according to the time dimension as a drafting input vector to integrate the drafting speed data information on the time dimension within the predetermined time period. Next, considering that the drafting speed has uncertainty and volatility, a multi-scale neighborhood feature extraction module is further used to encode the drafting input vector to extract the multi-scale neighborhood correlation features of the drafting speed under different time spans, thereby obtaining a multi-scale drafting feature vector.

[0054] Next, in order to facilitate the subsequent mining of the correlation features between the dynamic features of the drafting speed and the real-time state features of the drafting to accurately perform real-time adaptive control of the drafting speed, it is necessary to construct an association model of the dynamic features of the drafting speed and the real-time state features of the drafting. That is, first, the feature matrices along the channel dimension of the drafting tracking feature map are subjected to global mean pooling processing to obtain a drafting tracking feature vector. In this way, the number of parameters can be reduced to prevent overfitting, thereby facilitating the subsequent mining of the correlation features of the dynamic features of the drafting speed and the real-time state features of the drafting. Then, the transfer matrix of the multi-scale drafting feature vector relative to the drafting tracking feature vector is calculated to represent the correlation features between the dynamic multi-scale neighborhood correlation features of the drafting speed and the real-time change features of the drafting state, and classification processing is performed based on this to obtain a classification result for indicating whether the drafting speed at the current time point should be increased or decreased. In this way, the speed of the drafting at the current time point can be adaptively and accurately controlled according to the obtained classification result.

[0055] In particular, in the technical solution of the present application, when GAP processing is performed on each feature matrix along the channel dimension of the draft tracking feature map to obtain a draft tracking feature vector, since the eigenvalue of each position of the draft tracking feature vector is obtained by global mean pooling of the feature matrix of the corresponding channel position of the draft tracking feature map, the contextual correlation between the eigenvalues ​​of the draft tracking feature vector is weaker than the contextual correlation along the channel dimension of the draft tracking feature map, thereby weakening the expression capability of the draft tracking feature vector for the contextual feature correlation of the draft tracking feature map.

[0056] Therefore, in order to improve the expression capability of the context feature association of the draft tracking feature vector to the draft tracking feature map, in the technical solution of the present application, a local scene metric loss function of the context statistics of the draft tracking feature vector is introduced, which is expressed as:

[0057]

[0058] Here, μ and σ are feature sets The mean and variance of i is the eigenvalue of each position of the draft tracking feature vector V, and L is the length of the draft tracking feature vector V.

[0059] Here, the contextual statistical local scene metric loss function is based on the global pooling of each feature matrix along the channel of the draft tracking feature map, and the feature values ​​of each position of the draft tracking feature vector V are regarded as separate channel feature descriptors as a squeezed representation of the channel scene of the feature. Therefore, the contextual statistical metric based on the class probability expression of the feature set is used as a loss function to improve the correlation of the local scenes of each channel, thereby improving the expression ability of the draft tracking feature vector for the contextual feature association of the draft tracking feature map, thereby improving the accuracy of classification. In this way, the draft speed at the current time point can be adaptively and accurately controlled, thereby avoiding damage to the product due to drafting, and ensuring the production quality of the three-proof lampshade products.

[0060] Based on this, the present application proposes a three-proof lampshade injection mold with an ejection mechanism, which includes: a demolding monitoring unit, used to obtain the demolding speeds at multiple predetermined time points within a predetermined time period and a demolding monitoring video of the predetermined time period; a key frame extraction unit, used to extract multiple demolding key frames from the demolding monitoring video; a demolding process feature extraction unit, used to obtain a demolding tracking feature map by training a convolutional neural network model using a temporal attention mechanism for the multiple demolding key frames; a demolding speed feature extraction unit, used to sort the demolding speeds at multiple predetermined time points within the predetermined time period according to the time dimension A multi-scale neighborhood feature extraction module is used as a tapping input vector and then trained to obtain a multi-scale tapping feature vector; a dimension reduction unit is used to perform global mean pooling processing on each feature matrix along the channel dimension of the tapping tracking feature map to obtain a tapping tracking feature vector; a transfer unit is used to calculate the transfer matrix of the multi-scale tapping feature vector relative to the tapping tracking feature vector as a classification feature matrix; and a control result generation unit is used to pass the classification feature matrix through a trained classifier to obtain a classification result, and the classification result is used to indicate whether the tapping speed at the current time point should be increased or decreased.

[0061] Figure 1 The figure shows an application scenario of a three-proof lampshade injection mold with an ejection mechanism according to an embodiment of the present application. Figure 1 As shown, in this application scenario, a speed sensor (e.g., Figure 1 S1) in the figure obtains the draft speed at multiple predetermined time points within a predetermined time period and the draft speed at multiple predetermined time points within a predetermined time period through a camera (for example, Figure 1 C) as shown in the figure, obtains the demoulding monitoring video of the predetermined time period. Then, the above information is input into a server (for example, Figure 1In S2), the server is capable of processing the above information using the triple-proof lampshade injection molding algorithm with an ejection mechanism to generate a classification result indicating whether the ejection speed at the current time point should be increased or decreased.

[0062] After introducing the basic principles of the present application, various non-limiting embodiments of the present application will be described in detail with reference to the accompanying drawings.

[0063] Exemplary Systems

[0064] Figure 2 The figure shows a block diagram of a three-proof lampshade injection mold with an ejection mechanism according to an embodiment of the present application. Figure 2 As shown, the triple-proof lamp cover injection mold 300 with an ejection mechanism according to an embodiment of the present application includes an inference module, wherein the inference module includes: an ejection monitoring unit 310; a key frame extraction unit 320; an ejection process feature extraction unit 330; an ejection speed feature extraction unit 340; a dimensionality reduction unit 350; a transfer unit 360; and a control result generation unit 370.

[0065] The demolding monitoring unit 310 is used to obtain the demolding speeds at multiple predetermined time points within a predetermined time period and the demolding monitoring video of the predetermined time period; the key frame extraction unit 320 is used to extract multiple demolding key frames from the demolding monitoring video; the demolding process feature extraction unit 330 is used to obtain a demolding tracking feature map by using a convolutional neural network model with a time attention mechanism that has been trained for the multiple demolding key frames; the demolding speed feature extraction unit 340 is used to arrange the demolding speeds at multiple predetermined time points within the predetermined time period into a demolding input vector according to the time dimension and then train the demolding process feature extraction unit 330 to obtain a demolding tracking feature map. The multi-scale neighborhood feature extraction module is trained to obtain a multi-scale drafting feature vector; the dimension reduction unit 350 is used to perform global mean pooling processing on each feature matrix along the channel dimension of the draft tracking feature map to obtain the draft tracking feature vector; the transfer unit 360 is used to calculate the transfer matrix of the multi-scale drafting feature vector relative to the draft tracking feature vector as a classification feature matrix; and the control result generation unit 370 is used to pass the classification feature matrix through the trained classifier to obtain a classification result, and the classification result is used to indicate whether the drafting speed at the current time point should be increased or decreased.

[0066] Figure 4 The diagram shows the system architecture of the inference module in the injection mold of the tri-proof lampshade with an ejection mechanism according to an embodiment of the present application. Figure 4As shown, in the system architecture of the three-proof lampshade injection mold 300 with an ejection mechanism, in the inference process, the demolding speeds at multiple predetermined time points within a predetermined time period and the demolding monitoring video of the predetermined time period are first obtained through the demolding monitoring unit 310; the key frame extraction unit 320 extracts multiple demolding key frames from the demolding monitoring video obtained by the demolding monitoring unit 310; then, the demolding process feature extraction unit 330 uses the multiple demolding key frames extracted by the key frame extraction unit 320 through the trained convolutional neural network model using the temporal attention mechanism to obtain a demolding tracking feature map; at the same time, the demolding speed feature extraction unit 340 uses the multiple demolding key frames within the predetermined time period obtained by the demolding monitoring unit 310 to obtain a demolding tracking feature map. The draft speeds at predetermined time points are arranged according to the time dimension as a draft input vector, and then a multi-scale neighborhood feature extraction module is trained to obtain a multi-scale draft feature vector; then, the dimension reduction unit 350 performs global mean pooling processing on each feature matrix along the channel dimension of the draft tracking feature map obtained by the draft process feature extraction unit 330 to obtain a draft tracking feature vector; the transfer unit 360 calculates the transfer matrix of the multi-scale draft feature vector relative to the draft tracking feature vector as a classification feature matrix; further, the control result generation unit 370 passes the classification feature matrix through a trained classifier to obtain a classification result, and the classification result is used to indicate whether the draft speed at the current time point should be increased or decreased.

[0067] Specifically, during the operation of the three-proof lampshade injection mold 300 with an ejection mechanism, the demolding monitoring unit 310 is used to obtain the demolding speed at multiple predetermined time points within a predetermined time period and the demolding monitoring video of the predetermined time period. In a specific example of the present application, the demolding speed at multiple predetermined time points within a predetermined time period can be obtained by a speed sensor, and the demolding monitoring video of the predetermined time period can be obtained by a camera.

[0068] Specifically, during the operation of the three-proof lampshade injection mold 300 with the ejection mechanism, the key frame extraction unit 320 is used to extract multiple drafting key frames from the drafting monitoring video. It should be understood that considering that there are many image frames in the drafting monitoring video of the predetermined time period, this brings difficulties to the subsequent real-time state feature extraction of the drafting. Therefore, in order to fully extract the implicit feature information of the drafting state in the drafting monitoring video of the predetermined time period while simplifying the number of parameters, multiple drafting key frames are further extracted from the drafting monitoring video. In a specific example of the present application, the difference frame method can be used to extract the multiple drafting key frames from the drafting monitoring video.

[0069] Specifically, during the operation of the three-proof lampshade injection mold 300 with the ejection mechanism, the demolding process feature extraction unit 330 is used to obtain a demolding tracking feature map by passing the multiple demolding key frames through the trained convolutional neural network model using the time attention mechanism. In the technical solution of the present application, for the multiple demolding key frames extracted from the demolding monitoring video, a convolutional neural network with excellent performance in the extraction of implicit features of images is used to mine the implicit features of the multiple demolding key frames. However, considering that when the demolding speed is adaptively controlled, it is necessary to pay more attention to the dynamic characteristics of the demolding state in the time dimension to more accurately detect the real-time state of the demolding. Therefore, in the technical solution of the present application, a convolutional neural network model with a time attention mechanism is used to process the multiple demolding key frames to extract the local implicit feature distribution information in the multiple demolding key frames, that is, the dynamic change characteristics of the demolding state, so as to obtain a demolding tracking feature map. It should be understandable that the temporal attention mechanism allows the convolutional network model to focus its learning on the more important time dimension area of ​​the sample, thereby improving the efficiency of the network model in learning samples and further improving the accuracy of subsequent classification.

[0070] Figure 5 The figure shows a block diagram of a demolding process feature extraction unit in a tri-proof lampshade injection mold with an ejection mechanism according to an embodiment of the present application. Figure 5 As shown, the demolding process feature extraction unit 330 includes: an adjacent frame extraction subunit 331, used to extract the adjacent first frame and second frame from the multiple demolding key frames; a first image local feature extraction subunit 332, used to pass the first frame and the second frame through the first convolution layer and the second convolution layer of the convolutional neural network model respectively to obtain a first frame feature map corresponding to the first frame and a second frame feature map corresponding to the second frame; a time attention subunit 333, used to perform positional point multiplication on the first frame feature map and the second frame feature map and then pass the Softmax activation function to obtain a time attention map; a second image local feature extraction subunit 334, used to pass the first frame through the third convolution layer of the convolutional neural network model to obtain a local feature map; a time attention application subunit 335, used to perform positional point multiplication on the local feature map and the time attention map to obtain the demolding tracking feature map.

[0071] Specifically, during the operation of the three-proof lampshade injection mold 300 with an ejection mechanism, the draft speed feature extraction unit 340 is used to arrange the draft speeds of multiple predetermined time points within the predetermined time period according to the time dimension as a draft input vector, and then obtain a multi-scale draft feature vector through a multi-scale neighborhood feature extraction module that has been trained. It should be understood that for the draft speeds of multiple predetermined time points within the predetermined time period, they are arranged according to the time dimension as a draft input vector to integrate the draft speed data information on the time dimension within the predetermined time period. Next, considering that the draft speed has uncertainty and volatility, a multi-scale neighborhood feature extraction module is further used to encode the draft input vector to extract the multi-scale neighborhood correlation features of the draft speed under different time spans, thereby obtaining a multi-scale draft feature vector.

[0072] Figure 6 The figure shows a block diagram of a demolding speed feature extraction unit in a tri-proof lampshade injection mold with an ejection mechanism according to an embodiment of the present application. Figure 6 As shown, the draft speed feature extraction unit 340 includes: a first scale feature extraction subunit 341, which is used to input the draft input vector into the first convolution layer of the multi-scale neighborhood feature extraction module to obtain a first neighborhood scale draft feature vector, wherein the first convolution layer has a first one-dimensional convolution kernel of a first length; a second scale feature extraction subunit 342, which is used to input the draft input vector into the second convolution layer of the multi-scale neighborhood feature extraction module to obtain a second neighborhood scale draft feature vector, wherein the second convolution layer has a second one-dimensional convolution kernel of a second length, and the first length is different from the second length; and a multi-scale feature fusion subunit 343, which is used to cascade the first neighborhood scale draft feature vector and the second neighborhood scale temperature draft feature vector to obtain the multi-scale draft feature vector.

[0073] Specifically, during the operation of the three-proof lampshade injection mold 300 with the ejection mechanism, the dimension reduction unit 350 is used to perform global mean pooling processing on each feature matrix along the channel dimension of the draft tracking feature map to obtain a draft tracking feature vector. It should be understood that in order to facilitate the subsequent mining of the correlation characteristics of the draft speed dynamic characteristics and the draft real-time state characteristics to accurately perform real-time adaptive control of the draft speed, it is necessary to construct an association model of the draft speed dynamic characteristics and the draft real-time state characteristics. That is, first, the global mean pooling processing is performed on each feature matrix along the channel dimension of the draft tracking feature map to obtain a draft tracking feature vector. In this way, the number of parameters can be reduced to prevent overfitting, thereby facilitating the subsequent mining of the correlation characteristics of the draft speed dynamic characteristics and the draft real-time state characteristics.

[0074] Specifically, during the operation of the tri-proof lampshade injection mold 300 with the ejection mechanism, the transfer unit 360 is used to calculate the transfer matrix of the multi-scale draft feature vector relative to the draft tracking feature vector as a classification feature matrix. In the technical solution of the present application, the transfer matrix of the multi-scale draft feature vector relative to the draft tracking feature vector is calculated to represent the correlation characteristics between the dynamic multi-scale neighborhood association characteristics of the draft speed and the real-time change characteristics of the draft state, and classification processing is performed on this basis to obtain a classification result indicating whether the draft speed at the current time point should be increased or decreased. In this way, the draft speed at the current time point can be adaptively and accurately controlled based on the obtained classification result. In a specific example of the present application, the transfer unit is further used to: calculate the transfer matrix of the multi-scale draft feature vector relative to the draft tracking feature vector as a classification feature matrix using the following formula;

[0075] Wherein, the formula is:

[0076]

[0077] Wherein V1 represents the multi-scale draft feature vector, M2 represents the transfer matrix of the draft tracking feature vector, and M represents the classification feature matrix. Represents vector multiplication.

[0078] Specifically, during the operation of the three-proof lampshade injection mold 300 with an ejection mechanism, the control result generation unit 370 is used to pass the classification feature matrix through the trained classifier to obtain a classification result, and the classification result is used to indicate whether the demolding speed at the current time point should be increased or decreased. In a specific example of the present application, the control result generation unit is further used to: use the classifier to process the classification feature matrix with the following formula to obtain a classification result, and the formula is:

[0079] softmax{(W n , B n ):...:(W1,B1)|Project(F)}, where Project(F) represents the projection of the classification feature matrix into a vector, W1 to W n is the weight matrix of each fully connected layer, B1 to D n Represents the bias matrix of each fully connected layer.

[0080] It should be understood that before using the above neural network model for inference, the convolutional neural network model using the time attention mechanism, the multi-scale neighborhood feature extraction module and the classifier need to be trained. That is to say, in the injection mold of the tri-proof lampshade with an ejection mechanism of the present application, a training module is also included for training the convolutional neural network model using the time attention mechanism, the multi-scale neighborhood feature extraction module and the classifier.

[0081] Figure 3 The figure shows a block diagram of a three-proof lampshade injection mold with an ejection mechanism according to an embodiment of the present application. Figure 3 As shown, according to the triple-proof lamp shade injection mold 300 with an ejection mechanism according to the embodiment of the present application, it also includes a training module 400, and the training module includes: a training data acquisition unit 410; a training key frame extraction unit 420; a training demolding process feature extraction unit 430; a training demolding speed feature extraction unit 440; a training dimension reduction unit 450; a training transfer unit 460; a classification loss unit 470; a context loss measurement unit 480; and a training unit 490.

[0082] Among them, the training data acquisition unit 410 is used to obtain training data, and the training data includes training demolding speeds at multiple predetermined time points within a predetermined time period and training demolding monitoring videos of the predetermined time period; the training key frame extraction unit 420 is used to extract multiple training demolding key frames from the training demolding monitoring videos; the training demolding process feature extraction unit 430 is used to pass the multiple training demolding key frames through the convolutional neural network model using the time attention mechanism to obtain a training demolding tracking feature map; the training demolding speed feature extraction unit 440 is used to arrange the training demolding speeds at multiple predetermined time points within the predetermined time period according to the time dimension as a training demolding input vector and then pass it through the multi-scale neighborhood feature extraction module to obtain a training multi-scale demolding feature vector; the training dimension reduction unit 450 is used to perform global mean pooling on each feature matrix along the channel dimension of the training demolding tracking feature map Processing to obtain a training draft tracking feature vector; the training transfer unit 460 is used to calculate the transfer matrix of the training multi-scale draft feature vector relative to the training draft tracking feature vector as a training classification feature matrix; the classification loss unit 470 is used to pass the training classification feature matrix through a classifier to obtain a classification loss function value; the context loss measurement unit 480 is used to calculate the context statistical local scene measurement loss function value of the training draft tracking feature vector, and the context statistical local scene measurement loss function value is related to the statistical characteristics of the feature value set of all positions of the training draft tracking feature vector; and the training unit 490 is used to train the convolutional neural network model using the temporal attention mechanism, the multi-scale neighborhood feature extraction module and the classifier with the weighted sum of the context statistical local scene measurement loss function value and the classification loss function value as the loss function value.

[0083] Figure 7 The diagram shows the system architecture of the training module in the injection mold of the three-proof lampshade with the ejection mechanism according to the embodiment of the present application. Figure 7As shown, in the system architecture of the three-proof lampshade injection mold 300 with an ejection mechanism, during the training process, the training data is first acquired through the training data acquisition unit 410, and the training data includes the training demolding speeds at multiple predetermined time points within a predetermined time period and the training demolding monitoring video of the predetermined time period; the training key frame extraction unit 420 extracts multiple training demolding key frames from the training demolding monitoring video acquired by the training data acquisition unit 410; the training demolding process feature extraction unit 430 passes the multiple training demolding key frames extracted by the training key frame extraction unit 420 through the convolutional neural network model using the time attention mechanism to obtain a training demolding tracking feature map; at the same time, the training demolding speed feature extraction unit 440 arranges the training demolding speeds at multiple predetermined time points within the predetermined time period acquired by the training data acquisition unit 410 according to the time dimension as a training demolding input vector, and then passes it through the multi-scale neighborhood feature extraction module to obtain a training multi-scale demolding feature vector; the training dimension reduction unit 450 extracts the training demolding key frames extracted by the training key frame extraction unit 420 through the convolutional neural network model using the time attention mechanism to obtain a training demolding tracking feature map; at the same time, the training demolding speed feature extraction unit 440 arranges the training demolding speeds at multiple predetermined time points within the predetermined time period acquired by the training data acquisition unit 410 according to the time dimension as a training demolding input vector, and then passes it through the multi-scale neighborhood feature extraction module to obtain a training multi-scale demolding feature vector; The feature matrices of the training knock tracking feature map along the channel dimension of the training knock tracking process feature extraction unit 430 are globally mean pooled to obtain the training knock tracking feature vector; then, the training transfer unit 460 calculates the transfer matrix of the training multi-scale knock feature vector relative to the training knock tracking feature vector as the training classification feature matrix; the classification loss unit 470 passes the training classification feature matrix through the classifier to obtain the classification loss function value; the context loss measurement unit 480 calculates the local scene measurement loss function value of the context statistics of the training knock tracking feature vector, and the local scene measurement loss function value of the context statistics is related to the statistical characteristics of the feature value set of all positions of the training knock tracking feature vector; further, the training unit 490 uses the weighted sum of the local scene measurement loss function value of the context statistics and the classification loss function value as the loss function value to train the convolutional neural network model using the temporal attention mechanism, the multi-scale neighborhood feature extraction module and the classifier.

[0084] In particular, in the technical solution of the present application, when GAP processing is performed on each feature matrix along the channel dimension of the draft tracking feature map to obtain a draft tracking feature vector, since the eigenvalue of each position of the draft tracking feature vector is obtained by global mean pooling of the feature matrix of the corresponding channel position of the draft tracking feature map, the contextual correlation between the eigenvalues ​​of the draft tracking feature vector is weaker than the contextual correlation along the channel dimension of the draft tracking feature map, thereby weakening the expression capability of the draft tracking feature vector for the contextual feature correlation of the draft tracking feature map.

[0085] Therefore, in order to improve the expression capability of the context feature association of the draft tracking feature vector to the draft tracking feature map, in the technical solution of the present application, a local scene metric loss function of the context statistics of the draft tracking feature vector is introduced, which is expressed as:

[0086]

[0087] Here, μ and σ are the feature sets v i The mean and variance of ∈V, v i is the eigenvalue of each position of the draft tracking feature vector V, and L is the length of the draft tracking feature vector V.

[0088] Here, the contextual statistical local scene metric loss function is based on the global pooling of each feature matrix along the channel of the draft tracking feature map, and the feature values ​​of each position of the draft tracking feature vector V are regarded as separate channel feature descriptors as a squeezed representation of the channel scene of the feature. Therefore, the contextual statistical metric based on the class probability expression of the feature set is used as a loss function to improve the correlation of the local scenes of each channel, thereby improving the expression ability of the draft tracking feature vector for the contextual feature association of the draft tracking feature map, thereby improving the accuracy of classification. In this way, the draft speed at the current time point can be adaptively and accurately controlled, thereby avoiding damage to the product due to drafting, and ensuring the production quality of the three-proof lampshade products.

[0089] In summary, the three-proof lamp shade injection mold 300 with a baseband ejection mechanism according to the embodiment of the present application is explained, which monitors the real-time state of demolding by utilizing an artificial intelligence algorithm based on deep learning, and extracts and analyzes the dynamic characteristics of the demolding speed, so as to adaptively control the demolding speed at the current time point, thereby avoiding damage to the product due to demolding and ensuring the production quality of the three-proof lamp shade product.

[0090] As described above, the three-proof lamp shade with an ejection mechanism according to the embodiment of the present application can be implemented in various terminal devices. In one example, the three-proof lamp shade injection mold 300 with an ejection mechanism according to the embodiment of the present application can be integrated into the terminal device as a software module and / or a hardware module. For example, the three-proof lamp shade injection mold 300 with an ejection mechanism can be a software module in the operating system of the terminal device, or can be an application developed for the terminal device; of course, the three-proof lamp shade injection mold 300 with an ejection mechanism can also be one of the many hardware modules of the terminal device.

[0091] Alternatively, in another example, the tri-proof lamp shade injection mold 300 with ejection mechanism and the terminal device may also be separate devices, and the tri-proof lamp shade injection mold 300 with ejection mechanism may be connected to the terminal device via a wired and / or wireless network and transmit interactive information in accordance with an agreed data format.

[0092] Exemplary Methods

[0093] Figure 8 The figure shows a flow chart of a method for injection molding a lampshade of a tri-proof lamp with an ejection mechanism according to an embodiment of the present application. Figure 8 As shown, according to the injection molding method of a three-proof lamp shade with an ejection mechanism in an embodiment of the present application, the method comprises the following steps: S110, obtaining the demolding speeds at multiple predetermined time points within a predetermined time period and the demolding monitoring video of the predetermined time period; S120, extracting multiple demolding key frames from the demolding monitoring video; S130, passing the multiple demolding key frames through a trained convolutional neural network model using a temporal attention mechanism to obtain a demolding tracking feature map; S140, arranging the demolding speeds at multiple predetermined time points within the predetermined time period according to the time dimension as a demolding input vector. After measurement, a multi-scale neighborhood feature extraction module is trained to obtain a multi-scale drafting feature vector; S150, global mean pooling is performed on each feature matrix along the channel dimension of the draft tracking feature map to obtain a draft tracking feature vector; S160, a transfer matrix of the multi-scale drafting feature vector relative to the draft tracking feature vector is calculated as a classification feature matrix; and, S170, the classification feature matrix is ​​passed through a trained classifier to obtain a classification result, and the classification result is used to indicate whether the drafting speed at the current time point should be increased or decreased.

[0094] In one example, in the above-mentioned method for injection molding a tri-proof lamp cover with an ejection mechanism, the step S120 includes: extracting the multiple demolding key frames from the demolding monitoring video using a difference frame method.

[0095] In one example, in the above-mentioned method for injection molding a three-proof lamp cover with an ejection mechanism, the step S130 includes: extracting an adjacent first frame and a second frame from the multiple demolding key frames; passing the first frame and the second frame through the first convolution layer and the second convolution layer of the convolutional neural network model respectively to obtain a first frame feature map corresponding to the first frame and a second frame feature map corresponding to the second frame; performing positional point multiplication on the first frame feature map and the second frame feature map and then passing through a Softmax activation function to obtain a temporal attention map; passing the first frame through the third convolution layer of the convolutional neural network model to obtain a local feature map; performing positional point multiplication on the local feature map and the temporal attention map to obtain the demolding tracking feature map.

[0096] In one example, in the above-mentioned method for injection molding a tri-proof lamp cover with an ejection mechanism, the step S140 includes: inputting the draft input vector into the first convolution layer of the multi-scale neighborhood feature extraction module to obtain a first neighborhood scale draft feature vector, wherein the first convolution layer has a first one-dimensional convolution kernel of a first length; inputting the draft input vector into the second convolution layer of the multi-scale neighborhood feature extraction module to obtain a second neighborhood scale draft feature vector, wherein the second convolution layer has a second one-dimensional convolution kernel of a second length, and the first length is different from the second length; and cascading the first neighborhood scale draft feature vector and the second neighborhood scale temperature draft feature vector to obtain the multi-scale draft feature vector.

[0097] In one example, in the above-mentioned method for injection molding a tri-proof lampshade with an ejection mechanism, the step S160 includes: calculating a transfer matrix of the multi-scale draft feature vector relative to the draft tracking feature vector as a classification feature matrix using the following formula;

[0098] Wherein, the formula is:

[0099]

[0100] Wherein V1 represents the multi-scale draft feature vector, M2 represents the transfer matrix of the draft tracking feature vector, and M represents the classification feature matrix. Represents vector multiplication.

[0101] In one example, in the above-mentioned method for injection molding a lampshade of a tri-proof lamp with an ejection mechanism, the step S170 includes: using the classifier to process the classification feature matrix using the following formula to obtain a classification result, and the formula is:

[0102] softmax{(W n , B n ):...:(W1,B1)|Project(F)}, where Project(F) represents the projection of the classification feature matrix into a vector, W1 to W n is the weight matrix of each fully connected layer, B1 to B n Represents the bias matrix of each fully connected layer.

[0103] In summary, according to the embodiment of the present application, the injection molding method of the triple-proof lamp shade with ejection mechanism is explained, which monitors the real-time state of demolding by utilizing an artificial intelligence algorithm based on deep learning, and extracts and analyzes the dynamic characteristics of the demolding speed, so as to adaptively control the demolding speed at the current time point, thereby avoiding damage to the product due to demolding and ensuring the production quality of the triple-proof lamp shade product.

Claims

1. An injection mold for a triple-proof lampshade with an ejection mechanism, characterized in that: include: A demoulding monitoring unit, used to obtain the demoulding speeds at a plurality of predetermined time points within a predetermined time period and the demoulding monitoring video of the predetermined time period; A key frame extraction unit, used to extract a plurality of demolding key frames from the demolding monitoring video; A drafting process feature extraction unit, used for obtaining a drafting tracking feature map by subjecting the plurality of drafting key frames to a convolutional neural network model using a trained temporal attention mechanism; A drafting speed feature extraction unit, used for arranging the drafting speeds of a plurality of predetermined time points within the predetermined time period according to the time dimension as a drafting input vector and then obtaining a multi-scale drafting feature vector through a multi-scale neighborhood feature extraction module completed by training; A dimension reduction unit, used for performing global mean pooling processing on each feature matrix along the channel dimension of the draft tracking feature map to obtain a draft tracking feature vector; a transfer unit, used to calculate a transfer matrix of the multi-scale draft feature vector relative to the draft tracking feature vector as a classification feature matrix; and a control result generation unit, used to pass the classification feature matrix through a trained classifier to obtain a classification result, wherein the classification result is used to indicate whether the draft speed at the current time point should be increased or decreased; Wherein, the drafting process feature extraction unit comprises: an adjacent frame extraction subunit, used for extracting adjacent first and second frames from the plurality of drafting key frames; A first image local feature extraction subunit is used to pass the first frame and the second frame through the first convolution layer and the second convolution layer of the convolutional neural network model respectively to obtain a first frame feature map corresponding to the first frame and a second frame feature map corresponding to the second frame; A temporal attention subunit, configured to perform a positional dot multiplication on the first frame feature map and the second frame feature map and then pass the result through a Softmax activation function to obtain a temporal attention map; A second image local feature extraction subunit, configured to pass the first frame through a third convolutional layer of the convolutional neural network model to obtain a local feature map; A temporal attention applying subunit, used for performing position point multiplication of the local feature map and the temporal attention map to obtain the draft tracking feature map; Wherein, the draft speed feature extraction unit comprises: a first scale feature extraction subunit, configured to input the draft input vector into a first convolutional layer of the multi-scale neighborhood feature extraction module to obtain a first neighborhood scale draft feature vector, wherein the first convolutional layer has a first one-dimensional convolution kernel of a first length; a second scale feature extraction subunit, configured to input the draft input vector into a second convolution layer of the multi-scale neighborhood feature extraction module to obtain a second neighborhood scale draft feature vector, wherein the second convolution layer has a second one-dimensional convolution kernel of a second length, and the first length is different from the second length; and a multi-scale feature fusion subunit, configured to cascade the first neighborhood scale draft feature vector and the second neighborhood scale temperature draft feature vector to obtain the multi-scale draft feature vector; Wherein, the transfer unit is further used to: calculate the transfer matrix of the multi-scale draft feature vector relative to the draft tracking feature vector as a classification feature matrix using the following formula; Wherein, the formula is: Wherein V1 represents the multi-scale draft feature vector, M2 represents the transfer matrix of the draft tracking feature vector, and M represents the classification feature matrix. Represents vector multiplication; Wherein, the control result generating unit is further used to: use the classifier to process the classification feature matrix according to the following formula to obtain a classification result, and the formula is: softmax{(W n , B n ):...:(W1,B1)|Poject(F)}, where Project(F) represents projecting the classification feature matrix into a vector, W1 to W n is the weight matrix of each fully connected layer, B1 to B n Represents the bias matrix of each fully connected layer.

2. The triple-proof lampshade injection mold with ejection mechanism according to claim 1, characterized in that: The key frame extraction unit is further used to extract the multiple demolding key frames from the demolding monitoring video using a difference frame method.

3. The triple-proof lampshade injection mold with ejection mechanism according to claim 2 is characterized in that: It also includes a training module for training the convolutional neural network model using the temporal attention mechanism, the multi-scale neighborhood feature extraction module and the classifier.

4. The triple-proof lampshade injection mold with ejection mechanism according to claim 3 is characterized in that: The training module includes: a training data acquisition unit for acquiring training data, wherein the training data includes training drafting speeds at multiple predetermined time points within a predetermined time period and a training drafting monitoring video of the predetermined time period; A training key frame extraction unit, used to extract a plurality of training drafting key frames from the training drafting monitoring video; A training drafting process feature extraction unit, used for passing the plurality of training drafting key frames through the convolutional neural network model using the temporal attention mechanism to obtain a training drafting tracking feature map; A training draft speed feature extraction unit, used for arranging the training draft speeds at a plurality of predetermined time points within the predetermined time period according to the time dimension as a training draft input vector and then passing it through the multi-scale neighborhood feature extraction module to obtain a training multi-scale draft feature vector; A training dimension reduction unit, used for performing global mean pooling processing on each feature matrix along the channel dimension of the training draft tracking feature map to obtain a training draft tracking feature vector; A training transfer unit, used for calculating a transfer matrix of the training multi-scale draft feature vector relative to the training draft tracking feature vector as a training classification feature matrix; A classification loss unit, used for passing the training classification feature matrix through a classifier to obtain a classification loss function value; A context loss measurement unit, used to calculate the local scene measurement loss function value of the context statistics of the training draft tracking feature vector, wherein the local scene measurement loss function value of the context statistics is related to the statistical characteristics of the feature value set of all positions of the training draft tracking feature vector; and a training unit, used to train the convolutional neural network model using the temporal attention mechanism, the multi-scale neighborhood feature extraction module and the classifier using the weighted sum of the local scene measurement loss function value of the context statistics and the classification loss function value as the loss function value.

5. The triple-proof lampshade injection mold with ejection mechanism according to claim 4, characterized in that: The context loss measurement unit is further used to: calculate the local scene measurement loss function value of the context statistics of the training draft tracking feature vector according to the following formula; Wherein, the formula is: where v i represents the eigenvalue of each position of the training draft tracking feature vector, μ and σ respectively represent the mean and variance of the set of all eigenvalues ​​of the training draft tracking feature vector, and L is the length of the training draft tracking feature vector, exp(·) represents the exponential operation of a number, and the exponential operation of the number represents the calculation of the value of a natural exponential function with the number as the power.