VCSEL chip degradation evaluation method based on multi-modal feature fusion

Abstract

The present application relates to the technical field of semiconductor device detection, and particularly relates to a VCSEL chip degradation evaluation method based on multi-modal feature fusion, which mainly solves the technical problem of insufficient sensitivity of specific degradation mechanism in the existing VCSEL chip degradation evaluation method. The method comprises the following steps: obtaining multi-modal original data of a chip to be evaluated; labeling and preprocessing image original data to obtain image processing data, and inputting the image processing data into a convolutional neural network to extract image feature information; extracting degradation sensitive parameters in photoelectric original data, inputting the degradation sensitive parameters into a multi-layer perception machine after normalization and standardization to extract parameter feature information; respectively evaluating the image feature information and the parameter feature information to generate image modal degradation indicators and photoelectric degradation indicators; splicing the two degradation indicator features to form multi-modal joint features, and inputting the multi-modal joint features into a fusion network to obtain fusion features; and inputting the fusion features into a degradation regression layer to output an evaluation result representing the degradation degree of the chip.

Description

Technical Field

[0001] This invention relates to the field of semiconductor device testing technology, and in particular to a method for evaluating the degradation of VCSEL chips based on multimodal feature fusion. Background Technology

[0002] Vertical-cavity surface-emitting laser (VCSEL) chips are widely used in optical communication, 3D sensing, lidar, and consumer electronics due to their compact structure, low threshold current, high-quality emission spot, and ease of array integration. However, in actual production and application, VCSEL chips are prone to various forms of performance degradation under the influence of manufacturing process fluctuations, packaging stress, and long-term operating stress, thus affecting the consistency, stability, and lifespan of the device. Therefore, effectively assessing the degree of degradation of VCSEL chips is of great significance for improving product reliability and reducing failure rates.

[0003] In existing technologies, the degradation assessment of VCSEL chips typically relies on the test results of electrical or optical parameters, such as threshold current, output optical power, and slope efficiency. While these methods can reflect the overall performance trend of the chip, they have certain limitations in practical applications: Firstly, in the early stages of degradation, some degradation is mainly manifested as an increase in non-radiative recombination centers or changes in local material properties, and the changes in the aforementioned electrical or optical parameters are relatively small, making it difficult to reflect the chip's degradation state in a timely manner. Secondly, when the degradation area is located around the light-emitting aperture, at the cavity edge, or in the electrode contact area, its impact on the overall electrical or optical power parameters is limited, easily leading to misjudgments or missed detections. With the development of imaging detection technology, degradation detection methods based on VCSEL emission images have gradually been applied. By acquiring near-field or far-field emission images of the chip, information such as spot morphology, spatial distribution non-uniformity, and local anomalies can be obtained, thereby identifying some structural degradation features. However, these methods also have shortcomings; in the early stages of degradation or under conditions of high imaging noise, the changes in emission image features are not obvious, making it difficult to accurately reflect the true degree of chip degradation.

[0004] It is evident that existing degradation assessment methods for VCSEL chips suffer from insufficient sensitivity to specific degradation mechanisms, making it difficult to comprehensively and stably characterize the chip's state at different degradation locations and stages. Therefore, a technical solution that can improve the accuracy and stability of degradation assessment is urgently needed. Summary of the Invention

[0005] To overcome the technical problem of insufficient sensitivity to specific degradation mechanisms in existing VCSEL chip degradation assessment methods, this invention proposes a VCSEL chip degradation assessment method based on multimodal feature fusion.

[0006] The VCSEL chip degradation evaluation method based on multimodal feature fusion provided by this invention includes the following steps:

[0007] S1. Obtain multimodal raw data of the chip to be evaluated, wherein the multimodal raw data includes image raw data and photoelectric raw data;

[0008] S2. The original image data is labeled and preprocessed to obtain image processed data; the image processed data is input into a neural network for feature extraction to obtain image feature information used to characterize the chip's light emission morphology and spatial distribution characteristics.

[0009] Degradation-sensitive parameters are extracted from the raw photoelectric data; the degradation-sensitive parameters are normalized and standardized to obtain photoelectric processed data; the photoelectric processed data is input into a multilayer perceptron for feature extraction to obtain parameter feature information used to characterize the electrical and optical performance states of the chip.

[0010] S3. Perform degradation assessments on the image feature information and the parameter feature information respectively to generate intermediate assessment parameters; the intermediate assessment parameters include image modal degradation index and photoelectric degradation index;

[0011] S4. The intermediate evaluation parameters are concatenated to form multimodal joint features; the multimodal joint features are input into the fusion network to obtain fused features;

[0012] S5. Input the fused features into the degradation regression layer and output the evaluation results used to characterize the degree of chip degradation.

[0013] Furthermore, in step S1, the raw image data includes a visible light surface image and an infrared emission image of the chip to be evaluated; the raw photoelectric data includes at least one of the current, voltage, output optical power, and emission wavelength of the chip to be evaluated.

[0014] Furthermore, in step S2, the neural network is a convolutional neural network, which includes convolutional layers, multi-scale convolutional modules, and global pooling layers. The convolutional layers are used to perform convolutional processing on the image processing data, extracting information on light spot texture, edge structure, and degradation regions to generate convolutional data. The multi-scale convolutional modules are used to perform multi-scale convolutional processing on the convolutional data, extracting local degradation texture features and large-scale brightness distribution features to generate feature maps. The global pooling layer is used to perform global pooling operations on the feature maps to obtain image feature information.

[0015] Furthermore, in step S2, the degradation-sensitive parameter includes at least one of the following: threshold current change rate, output power decrease ratio, slope efficiency change, and spectral drift.

[0016] Further, in step S2, the multilayer perceptron includes an input layer, a first fully connected layer, a second fully connected layer, and an output layer; the input layer is used to receive the photoelectric processing data; the first fully connected layer is used to perform fully connected weighted and biased processing on the photoelectric processing data, and to perform nonlinear transformation through the ReLU activation function to generate a first hidden layer feature; the second fully connected layer is used to perform fully connected weighted and biased processing on the first hidden layer feature, and to perform nonlinear transformation through the ReLU activation function to generate a second hidden layer depth feature; the output layer is used to output the second hidden layer depth feature as the parameter feature information.

[0017] Furthermore, in step S3, the process of generating the image modality degradation index is as follows: global average pooling is performed on the image feature information to compress the spatial dimension and generate an image high-level semantic vector; linear mapping is performed on the image high-level semantic vector to generate the image modality degradation index.

[0018] Furthermore, in step S3, the process of generating the photoelectric degradation index is as follows: the parameter feature information is linearly mapped to generate the photoelectric degradation index.

[0019] Furthermore, in step S4, the fusion network includes a fully connected layer, a nonlinear activation layer, and a random deactivation layer; the fully connected layer is used to perform feature mapping on the multimodal joint features to obtain fusion mapped features; the nonlinear activation layer is used to perform nonlinear transformation on the fusion mapped features to obtain fusion activated features; the random deactivation layer is used to perform random deactivation processing on the fusion activated features to achieve regularization, thereby obtaining fusion features.

[0020] Furthermore, in step S5, the degenerate regression layer is used to weight and bias the fusion features and output the evaluation results.

[0021] Furthermore, the evaluation result ranges from 0 to 1, where 0 indicates that the chip under evaluation is in a healthy state, and 1 indicates that the chip under evaluation is in a state of extreme degradation.

[0022] The VCSEL chip degradation assessment method based on multimodal feature fusion provided by this invention has the following advantages compared with existing technologies:

[0023] 1) This method acquires multimodal raw data of the chip to be evaluated and combines raw image data with raw photoelectric data. It can comprehensively characterize the degradation state of VCSEL chip from two dimensions: spatial distribution characteristics and electro-optical performance. It overcomes the problem that single detection methods are not sensitive enough to specific degradation mechanisms and realizes a comprehensive evaluation of chip state at different degradation locations and different degradation stages.

[0024] 2) Before fusion, this method first performs independent degradation assessments on image feature information and parameter feature information to generate intermediate assessment parameters, so that the degradation information of each modality can be fully explored, avoiding feature conflicts or information loss caused by large differences between modalities when directly fusing the original features, thus improving the fusion effect and assessment accuracy.

[0025] 3) This method integrates complementary information of image modalities and photoelectric modalities by performing feature splicing on intermediate evaluation parameters and inputting them into a fusion network for deep fusion. This enables reliable degradation evaluation results to be obtained through multimodal joint analysis even if the changes in single modal features are not obvious in the early stage of degradation, thereby improving the accuracy and stability of degradation evaluation. Attached Figure Description

[0026] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a flowchart illustrating the overall evaluation method in an embodiment of the present invention.

[0029] Figure 2 This diagram illustrates the structure of the convolutional neural network in an embodiment of the present invention.

[0030] Figure 3 This is a structural diagram of the multilayer perceptron in an embodiment of the present invention. Detailed Implementation

[0031] To better understand the above-mentioned objectives, features, and advantages of the present invention, the solutions of the present invention will be further described below. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other.

[0032] Many specific details are set forth in the following description in order to provide a full understanding of the invention, but the invention may also be practiced in other ways different from those described herein; obviously, the embodiments in the specification are only some embodiments of the invention, and not all embodiments.

[0033] The following is combined Figures 1 to 3 Specific embodiments of the present invention will be described in detail below.

[0034] This embodiment provides a VCSEL chip degradation evaluation method based on multimodal feature fusion, including steps S1 to S5.

[0035] S1. Obtain the multimodal raw data of the chip to be evaluated. The multimodal raw data includes image raw data and photoelectric raw data.

[0036] Specifically, the raw image data includes visible light surface images and infrared emission images of the chip to be evaluated; the raw optoelectronic data includes at least one of the following: current, voltage, output optical power, and emission wavelength of the chip to be evaluated.

[0037] S2. The original image data is labeled and preprocessed to obtain image processing data; the image processing data is input into a neural network for feature extraction to obtain image feature information used to characterize the chip's light emission morphology and spatial distribution characteristics; degradation sensitive parameters are extracted from the original optoelectronic data; the degradation sensitive parameters are normalized and standardized to obtain optoelectronic processing data; the optoelectronic processing data is input into a multilayer perceptron for feature extraction to obtain parameter feature information used to characterize the chip's electrical and optical performance states.

[0038] It is easy to understand that image feature information is used to reflect anomalies in the spatial structure or light emission uniformity of the chip being evaluated, including information such as the shape of the light emission area, the uniformity of the light spot, local dark areas, ring defects, and edge anomalies.

[0039] Specifically, the neural network is a convolutional neural network, which includes convolutional layers, multi-scale convolutional modules, and global pooling layers. The convolutional layers are used to perform convolutional processing on the image processing data, extracting information on light spot texture, edge structure, and degradation regions to generate convolutional data. The multi-scale convolutional modules are used to perform multi-scale convolutional processing on the convolutional data, extracting local degradation texture features and large-scale brightness distribution features to generate feature maps. The global pooling layers are used to perform global pooling operations on the feature maps to obtain image feature information.

[0040] Convolutional Neural Networks (CNNs) are a type of feedforward neural network that incorporates convolutional computations and has a deep structure. They are widely used in tasks such as image classification, object detection, and speech recognition. Their core idea originates from the local receptive field mechanism of the biological visual cortex, significantly reducing the number of parameters and improving generalization ability through local connections and weight sharing. The expression for a convolutional neural network is:

[0041] ;

[0042] In the formula, Represents image feature information and , Indicates the number of feature channels. and These represent the height and width of the image feature information, respectively. Represents image processing data and , and These represent the height and width of the image processing data, respectively. It represents the RGB three-channel.

[0043] Specifically, degradation-sensitive parameters include at least one of the following: threshold current change rate, output power decrease ratio, slope efficiency change, and spectral drift. Based on the collected current, voltage, and output optical power test data, current-voltage curves and optical power-current curves are plotted; wavelength drift data is determined based on the collected emission wavelength test data; the threshold current change rate is extracted from the current-voltage curve as an electrical degradation-sensitive parameter; the output power decrease ratio and slope efficiency change are extracted from the optical power-current curve as optical degradation-sensitive parameters; and spectral drift is extracted from the wavelength drift data as a spectral degradation-sensitive parameter.

[0044] Specifically, the multilayer perceptron includes an input layer, a first fully connected layer, a second fully connected layer, and an output layer. The input layer receives photoelectric processing data. The first fully connected layer performs fully connected weighting and biasing on the photoelectric processing data and performs a nonlinear transformation using the ReLU activation function to generate first hidden layer features. The second fully connected layer performs fully connected weighting and biasing on the first hidden layer features and performs a nonlinear transformation using the ReLU activation function to generate second hidden layer depth features. The output layer outputs the second hidden layer depth features as parametric feature information.

[0045] A multilayer perceptron (MLP) is a type of feedforward neural network that, by introducing the ReLU activation function, can approximate arbitrarily complex nonlinear functions, thus performing well in tasks such as classification and regression. The expression for a multilayer perceptron is:

[0046] ;

[0047] In the formula, and These represent the weight matrix and bias vector of the first fully connected layer, respectively. and These represent the weight matrix and bias vector of the second fully connected layer, respectively. Represents the ReLU activation function. This represents the first hidden layer feature. This represents the depth feature of the second hidden layer. Represents parameter feature information and , This represents the number of dimensions of the parameter feature information. Represents photoelectric processing data and , Indicates the degradation-sensitive parameter, This represents the mean of each degradation-sensitive parameter. This represents the standard deviation of each degradation-sensitive parameter.

[0048] S3. Perform degradation assessments on image feature information and parameter feature information respectively to generate intermediate assessment parameters; the intermediate assessment parameters include image modal degradation index and photoelectric degradation index.

[0049] Specifically, the image modality degradation index (EMI) generation process is as follows: Global average pooling is applied to image feature information to compress spatial dimensions and generate a high-level semantic vector; this high-level semantic vector is then linearly mapped to generate the EMI. The calculation formula is:

[0050] ;

[0051] In the formula, Represents the high-level semantic vector of the image and , This represents the number of dimensions of the high-level semantic vector of the image. Indicates global average pooling. Represents image feature information, Indicates the image modal degradation index and , and These represent the weight matrix and bias vector of the linear mapping of the image modality degradation index, respectively.

[0052] Randomly generated during network initialization and continuously updated through backpropagation and gradient descent, it is used to linearly map the high-level semantic vector of an image to the dimensional space corresponding to the image modality degradation index, and learn the mapping relationship between the high-level semantic vector of an image and the image modality degradation index. This is used to correct baseline bias caused by weight mapping, compensate for feature mean shift, and enable the model to better fit the baseline shift of the image modality degradation index.

[0053] Specifically, the generation process of the photoelectric degradation index is as follows: Linear mapping is performed on the parameter characteristic information to generate the photoelectric degradation index. The calculation formula is:

[0054] ;

[0055] In the formula, Indicates photoelectric degradation index and , and Let represent the weight matrix and bias vector, respectively, of the linear mapping of the photoelectric degradation index. This indicates parameter characteristic information.

[0056] S4. The intermediate evaluation parameters are concatenated to form multimodal joint features; the multimodal joint features are input into the fusion network to obtain fused features.

[0057] Specifically, the expression for multimodal joint features is:

[0058] ;

[0059] In the formula, This represents the joint features of multiple modalities.

[0060] Specifically, the fusion network comprises fully connected layers, nonlinear activation layers, and random deactivation layers. The fully connected layers perform feature mapping on the joint multimodal features to obtain fused mapping features. The nonlinear activation layers perform nonlinear transformations on the fused mapping features to obtain fused activation features. The random deactivation layers perform random deactivation processing on the fused activation features to achieve regularization, resulting in fused features. The fusion network can learn the correspondence between image modality degradation indices and photoelectric degradation indices and the degree of chip degradation, and establish a complementary correlation between the two degradation levels, improving the accuracy and robustness of degradation assessment.

[0061] S5. Input the fused features into the degradation regression layer and output the evaluation results used to characterize the degree of chip degradation.

[0062] Specifically, the degenerate regression layer is used to weight and bias the fused features, outputting the evaluation results. The expression for the degenerate regression layer is:

[0063] ;

[0064] In the formula, Indicates the evaluation results. and These represent the weight matrix and bias vector of the degenerate regression layer, respectively. This indicates the fusion feature.

[0065] Specifically, the evaluation result ranges from 0 to 1, where 0 indicates that the chip under evaluation is in a healthy state, and 1 indicates that the chip under evaluation is in a state of extreme degradation.

[0066] By introducing multimodal feature fusion modeling, this method achieves high-precision and robust assessment of the degradation state of VCSEL chips. This method organically combines the macroscopic physical properties of photoelectric parameters with the spatial distribution characteristics of image modalities, effectively improving the ability to distinguish between localized degraded textures and large-scale brightness distributions. Furthermore, it maintains stable assessment performance even in the presence of data noise, providing stronger technical support for the reliability analysis and lifetime prediction of VCSEL chips.

[0067] The above are merely specific embodiments of the present invention, enabling those skilled in the art to understand or implement the present invention. Although detailed descriptions have been provided with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments, and they should all be covered within the protection scope of the claims.

Claims

1. A degradation assessment method for VCSEL chips based on multimodal feature fusion, characterized in that, Includes the following steps: S1. Obtain multimodal raw data of the chip to be evaluated, wherein the multimodal raw data includes image raw data and photoelectric raw data; S2. The original image data is labeled and preprocessed to obtain image processed data; the image processed data is input into a neural network for feature extraction to obtain image feature information used to characterize the chip's light emission morphology and spatial distribution characteristics. Degradation-sensitive parameters are extracted from the raw photoelectric data; the degradation-sensitive parameters are normalized and standardized to obtain photoelectric processed data; the photoelectric processed data is input into a multilayer perceptron for feature extraction to obtain parameter feature information used to characterize the electrical and optical performance states of the chip. S3. Perform degradation assessments on the image feature information and the parameter feature information respectively to generate intermediate assessment parameters; the intermediate assessment parameters include image modal degradation index and photoelectric degradation index; S4. The intermediate evaluation parameters are concatenated to form multimodal joint features; the multimodal joint features are input into the fusion network to obtain fused features; S5. Input the fused features into the degradation regression layer and output the evaluation results used to characterize the degree of chip degradation.

2. The VCSEL chip degradation assessment method based on multimodal feature fusion according to claim 1, characterized in that, In step S1, the raw image data includes a visible light surface image and an infrared emission image of the chip to be evaluated; the raw photoelectric data includes at least one of the current, voltage, output optical power, and emission wavelength of the chip to be evaluated.

3. The VCSEL chip degradation assessment method based on multimodal feature fusion according to claim 1, characterized in that, In step S2, the neural network is a convolutional neural network, which includes convolutional layers, multi-scale convolutional modules, and global pooling layers. The convolutional layers are used to perform convolution processing on the image processing data, extracting information on light spot texture, edge structure, and degradation regions to generate convolutional data. The multi-scale convolutional modules are used to perform multi-scale convolution processing on the convolutional data, extracting local degradation texture features and large-scale brightness distribution features to generate feature maps. The global pooling layer is used to perform global pooling operations on the feature maps to obtain image feature information.

4. The VCSEL chip degradation assessment method based on multimodal feature fusion according to claim 1, characterized in that, In step S2, the degradation-sensitive parameters include at least one of the following: threshold current change rate, output power decrease ratio, slope efficiency change, and spectral drift.

5. The VCSEL chip degradation assessment method based on multimodal feature fusion according to claim 1 or 4, characterized in that, In step S2, the multilayer perceptron includes an input layer, a first fully connected layer, a second fully connected layer, and an output layer. The input layer receives the photoelectric processing data. The first fully connected layer performs fully connected weighted and biased processing on the photoelectric processing data and performs a nonlinear transformation using the ReLU activation function to generate a first hidden layer feature. The second fully connected layer performs fully connected weighted and biased processing on the first hidden layer feature and performs a nonlinear transformation using the ReLU activation function to generate a second hidden layer depth feature. The output layer outputs the second hidden layer depth feature as the parameter feature information.

6. The VCSEL chip degradation assessment method based on multimodal feature fusion according to claim 1, characterized in that, In step S3, the process of generating the image modality degradation index is as follows: global average pooling is performed on the image feature information to compress the spatial dimension and generate an image high-level semantic vector; linear mapping is performed on the image high-level semantic vector to generate the image modality degradation index.

7. The VCSEL chip degradation assessment method based on multimodal feature fusion according to claim 1, characterized in that, In step S3, the photoelectric degradation index is generated by performing linear mapping processing on the parameter feature information.

8. The VCSEL chip degradation assessment method based on multimodal feature fusion according to claim 1, characterized in that, In step S4, the fusion network includes a fully connected layer, a nonlinear activation layer, and a random deactivation layer; the fully connected layer is used to perform feature mapping on the multimodal joint features to obtain fused mapped features; The nonlinear activation layer is used to perform a nonlinear transformation on the fused mapping features to obtain fused activation features; The random deactivation layer is used to perform random deactivation processing on the fused activation features to achieve regularization and obtain fused features.

9. The VCSEL chip degradation assessment method based on multimodal feature fusion according to claim 1, characterized in that, In step S5, the degenerate regression layer is used to weight and bias the fusion features and output the evaluation results.

10. The VCSEL chip degradation evaluation method based on multimodal feature fusion according to claim 9, characterized in that, The evaluation result ranges from 0 to 1, where 0 indicates that the chip under evaluation is in a healthy state and 1 indicates that the chip under evaluation is in a state of extreme degradation.

Images