Regression test simulation waveform classification method based on Mamba neural network

Abstract

The present invention discloses an automated classification method for regression test error simulation waveforms based on a Mamba neural network architecture, belonging to the field of integrated circuits. The method comprises parsing a simulation waveform file obtained from a regression test, converting a key signal waveform into a discrete sequence of integer waveforms; normalizing the numerical range of the waveform sequence to the 0-1 interval, intercepting a fixed length and inputting it into a Mamba neural network model, with the network outputting two parameters: block size and block starting position; dividing the waveform using the two characteristic parameters output by the network, calculating the thickness, degree of fluctuation, and mean characteristics of each waveform block, and calculating the overall characteristics; performing a fast Fourier transform on the waveform using integer multiples of the waveform block size as a window, and calculating frequency domain features; and merging the extracted features into a multidimensional feature vector and inputting it into a support vector machine to output a final classification result. The present invention can effectively improve the accuracy of diagnosing the cause of regression test failure, replace the manual triage process, and improve verification efficiency.

Description

Technical Field

[0001] The invention belongs to the field of integrated circuits, and in particular relates to a regression test simulation waveform classification method based on a Mamba neural network. Background Art

[0002] Regression testing is a key step in integrated circuit design and verification. Its purpose is to verify through simulation that circuits retain their original functionality after modifications. As the scale of integrated circuit designs continues to increase, the amount of data generated by regression testing is also rapidly increasing, posing a significant challenge to engineers in the analysis and classification of simulation waveforms. Traditional regression testing error analysis methods rely on manual inspection and empirical judgment, which is time-consuming and error-prone, making them unable to cope with the massive amount of waveform data in complex circuits.

[0003] Because digital circuit design often incorporates structures such as integrators, accumulators, and feedback loops, their simulation waveforms often contain many potential periodic or block features, which are crucial for fault diagnosis. However, existing classification methods often fail to effectively extract these latent waveform features, making it difficult to accurately distinguish error types. Therefore, finding a way to efficiently and accurately classify regression test error waveforms through automated techniques is a pressing issue for improving the efficiency and accuracy of regression test analysis.

[0004] As an efficient learning model, neural networks have been widely used in fields such as image processing and speech recognition. In integrated circuit testing and verification, the application of neural networks can effectively improve the speed and accuracy of data processing. In recent years, the Mamba neural network architecture, with its advantages in processing complex waveform data, has become a popular choice for waveform classification. Traditional neural network architectures often fail to fully consider the temporal and frequency domain characteristics of waveforms, resulting in unsatisfactory classification results and processing speed when faced with large amounts of test data.

[0005] Furthermore, most existing methods rely on a single classifier to classify waveforms, lacking a multi-dimensional feature fusion approach, making them incapable of handling highly complex waveforms. Classification performance can be severely limited, especially with small sample datasets, rendering it infeasible as a universal solution. Therefore, combining waveform time-domain block and frequency-domain features with efficient parallelized network architectures to accelerate the classification process has become an important research direction in error simulation waveform classification for regression testing.

[0006] To address these issues, this paper proposes a regression test simulation waveform classification method based on a Mamba neural network. By segmenting waveform data, extracting time-domain and frequency-domain features, and combining them with a support vector machine (SVM) for classification, this method achieves efficient and accurate classification results on small sample data sets, effectively improving the accuracy of regression test error analysis, replacing the tedious manual triage process, and enhancing verification efficiency. Summary of the Invention

[0007] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a regression test simulation waveform classification method based on a Mamba neural network. This method takes into account the block characteristics of the simulation waveform in the regression test, and first parses the key signal waveform in the regression test into a discrete sequence. Then, the Mamba neural network architecture is adopted to divide the waveform into blocks by taking advantage of the global receptive field and dynamic weighting. The time domain features are calculated using the block results, and the waveforms are classified using a support vector machine in combination with the frequency domain features. In order to improve the classification efficiency, a parallel network structure design is adopted to accelerate the classification process, which can achieve efficient classification of large-scale simulation waveforms under small sample data sets and improve the accuracy of diagnosis of regression test failure causes. The method proposed in the present invention can be used as a universal solution to replace the traditional manual triage process and effectively improve verification efficiency.

[0008] The present invention adopts the following technical solutions to achieve the above-mentioned purpose:

[0009] The regression test simulation waveform classification method based on Mamba neural network includes the following steps:

[0010] Step 1: parse the simulation waveform file obtained by regression testing to obtain an integer discrete sequence of key signal waveforms for subsequent processing and classification of the waveforms.

[0011] Step 2: Normalize the numerical range of the waveform sequence to the interval of 0-1, intercept the fixed length 2048 and input it into the Mamba neural network model. The network inference outputs two parameters: block size and block starting position.

[0012] Step three: Use the two characteristic parameters output by the network to divide the waveform into blocks, calculate the thickness, fluctuation degree, mean and other characteristics of each waveform, and count the overall characteristics.

[0013] Step 4: Use integer multiples of the waveform block size as a window to perform fast Fourier transform on the waveform and calculate frequency domain features such as the ratio of high-frequency and low-frequency energy and the concentration of spectral peak energy.

[0014] Step 5: Combine the features extracted from the above process to obtain a multi-dimensional feature vector, input it into the support vector machine, and output the final classification result.

[0015] Preferably, the method for processing the key signal waveform in step 1 is:

[0016] At the rising edge of the baseband clock, read the data bit values ​​of the key signal and convert the binary signal value at that moment into a decimal integer for storage. Repeat the above steps to obtain a discrete sequence of the key waveform signal.

[0017] Preferably, the method for normalizing the waveform sequence and inputting it into the Mamba neural network model in step 2 is:

[0018] Divide each value in the waveform sequence by the maximum value in the sequence to normalize the sequence values ​​to the range 0-1. Truncate the waveform sequence to a length of 2048, and padded with zeros if the length is insufficient. Input the processed 2048-length waveform sequence into the Mamba neural network model, which outputs two parameters: the block size L and the block starting position K.

[0019] Preferably, the waveform segmentation and feature extraction method in step 3 is:

[0020] Based on the model's output, the waveform is divided into blocks of equal size, starting from the block starting position K and divided into blocks of equal size, L. The waveform before the block starting position and the waveform whose length is less than L are discarded. The thickness, fluctuation degree, mean, and other characteristics of each block waveform are calculated. Based on the calculation results, the overall block thickness mean, block thickness fluctuation standard deviation, block fluctuation degree mean, and block mean standard deviation are calculated.

[0021] Preferably, the frequency domain feature calculation method in step 4 is:

[0022] Using N times the block size L (N is as large as possible and L * N does not exceed the total waveform length) output by the model as a window, perform a fast Fourier transform on the waveform sequence within the window to obtain an amplitude sequence of length L * N. Calculate the ratio of high-frequency and low-frequency energy in the amplitude sequence and the energy concentration at the spectral peak position of the integer multiple N of the amplitude sequence.

[0023] Preferably, the method for the support vector machine in step 5 to process features and obtain classification results is:

[0024] The seven eigenvalues ​​of block size L, block thickness mean, block thickness fluctuation standard deviation, block fluctuation degree mean, block mean standard deviation, high-frequency energy ratio, and spectral peak energy concentration are merged into a 7-dimensional feature vector and input into the support vector machine to finally obtain the output key signal waveform sequence classification result.

[0025] The present invention adopts the above technical solution and has the following beneficial effects:

[0026] (1) This paper proposes an automated classification method for regression test error simulation waveforms based on the Mamba neural network architecture. By segmenting the simulation waveforms used in regression testing, extracting the waveform's time domain segmentation features and frequency domain features, and combining them with a support vector machine for classification, this method achieves efficient and accurate classification of regression test waveform errors. Compared with traditional manual triage methods, this method can significantly improve classification efficiency, reduce manual intervention, and enhance the accuracy of diagnosing the causes of regression test errors.

[0027] (2) By adopting the Mamba neural network architecture, the present invention can effectively capture long-range dependencies and complex temporal patterns in waveforms. By leveraging the advantages of global receptive field and dynamic weighting, it provides better generalization, classification accuracy, and computational efficiency than traditional convolutional neural networks and transformer networks. The Mamba network can train efficient classification models on small sample data sets and is particularly suitable for the automated classification of regression test waveforms.

[0028] (3) The present invention is highly versatile and can be used as a solution to replace manual triage. By training the Mamba segmentation model using waveform training sets of different segment sizes, classification can be completed by simply retraining the support vector machine (SVM) when faced with new waveforms. Since the input feature dimension of the SVM is low and the training process is fast, this method can be quickly deployed and efficiently applied in practical applications and is suitable for different types of regression test waveform classification tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 4 is a flow chart of the automatic classification method of regression test error simulation waveform based on the Mamba neural network architecture of the present invention;

[0030] Figure 2 This is a diagram of the Mamba neural network architecture of the present invention;

[0031] Figure 3 This is a diagram of a single core layer structure in the Mamba neural network architecture of the present invention;

[0032] Figure 4 This is a schematic diagram of the thickness, fluctuation degree, and mean value characteristics of the block waveform of the present invention;

[0033] Figure 5 This is a schematic diagram of the final classification result obtained by inputting the feature vector into the support vector machine of the present invention. DETAILED DESCRIPTION

[0034] To make the objectives, technical features, and advantages of the present invention more readily apparent, the following detailed description of the specific embodiments of the present invention is provided in conjunction with the accompanying drawings. It is apparent that the embodiments described are only a portion of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative effort should fall within the scope of protection of the present invention.

[0035] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0036] Embodiment: The present invention will be further described below with reference to the accompanying drawings. Figure 1 As shown, it includes 5 steps.

[0037] 1. Read each data bit of the key signal at the rising edge of the baseband clock and convert the binary signal value at that moment into a decimal integer for storage. Repeat the above steps to obtain a discrete sequence of the key waveform signal.

[0038] Simulation waveforms are typically complex data streams composed of multiple signals that change periodically and synchronously with the baseband clock. The reason for regression test failure is often reflected in certain key signals. The present invention synchronously reads the values ​​of all data bits of key signals at the rising edge of each baseband clock.

[0039] After sampling on each rising clock edge, the acquired signal data bits are binary (i.e., each bit has only two possible states: "0" or "1"). To facilitate subsequent processing and analysis, these binary numbers are converted to decimal integers for storage. Repeating the above steps yields a discrete sequence of key waveform signals.

[0040] 2. Divide each value in the waveform sequence by the maximum value in the sequence to normalize the sequence values ​​to the range of 0-1. Truncate sequences that are too long to 2048. If the length is insufficient, add zeros to the end of the sequence to ensure that the waveform length is 2048 before entering the network.

[0041] The overall structure of the Mamba waveform block network is as follows Figure 2 As shown, the waveform is reversed and fed into the first Mamba core layer from both directions. The output of this core layer is activated using the ReLU activation function and then fed into the next core layer. After passing through multiple Mamba core layers, the output is merged into a single Mamba core layer. The final output of the core layer is passed through two linear layers to obtain two characteristic parameters: block size and block starting position.

[0042] The structure of a single Mamba core layer is as follows Figure 3 As shown in the figure, the input sequence is first divided into multiple local windows, allowing SSM to be performed in different directions while retaining the global SSM operation. The core layer implements spatial and channel attention modules before patch merging to enhance the integration of directional features and reduce redundancy. In addition, a strategy is adopted to select the most effective scanning direction for each layer, thereby optimizing computational efficiency.

[0043] Combining these technical approaches, the network architecture maximizes the model's global perception capabilities and closely simulates the manual waveform segmentation process. Furthermore, this setup enables parallel computing within a single Mamba core layer and across core layers, significantly improving the model's inference efficiency.

[0044] 3. Based on the output of the Mamba block model in step 2, the waveform is divided into waveform blocks of the same length L in sequence starting from the block starting position K and with the block size L as the step size. The waveform before the block starting position K and the waveform with a length less than L at the end are discarded.

[0045] Calculate the thickness, fluctuation degree and mean characteristics of each waveform. The specific meanings of the above characteristics are as follows Figure 4 As shown in the figure, the thickness T of the block is the difference between the second largest value and the second smallest value in the waveform block, which represents the thickness of the waveform block; the mean E of the block is the mean of the waveform block after removing the maximum and minimum values, which represents the average height of the waveform block; the fluctuation degree D of the block is the mean of the absolute value of the gradient of the waveform block sequence value, which represents the fluctuation degree of the waveform block. Figure 4 The difference between waveform blocks with larger and smaller D values ​​is shown in the figure. It can be seen that the waveform blocks with smaller D values ​​appear sparser.

[0046] Assuming that the entire waveform can be divided into m blocks, each block size is L, the block thickness mean of all waveform blocks will be calculated. , Block thickness fluctuation standard deviation , block fluctuation mean , block mean standard deviation .

[0047] Their specific calculation methods are as follows:

[0048] Block thickness average : , block thickness fluctuation standard deviation : , the mean value of block fluctuation : , block mean standard deviation : , so far all the time domain block features have been calculated.

[0049] 4. Use N times the block size L output by the Mamba block model in step 2 as a window (N is an integer, N is as large as possible and L * N does not exceed the total length of the waveform), perform fast Fourier transform on the waveform sequence in the window, and obtain an amplitude sequence of length L * N Since the amplitude sequence is centrosymmetric, the first L * N / 2 values ​​are retained.

[0050] Calculate the ratio of high-frequency and low-frequency energy in an amplitude sequence , the energy concentration at the spectral peak position of the integer multiple of the amplitude sequence N As frequency domain features. Their specific calculation methods are as follows:

[0051] Ratio of high-frequency to low-frequency energy : , , , spectral peak energy concentration : , , so far all the frequency domain features have been calculated.

[0052] 5. In the present invention, in order to achieve efficient classification of regression test error simulation waveforms, multiple eigenvalues ​​are combined into a eigenvector after feature extraction and input into the support vector machine (SVM) model for classification. Figure 5 As shown, the block size L and the block thickness mean , Block thickness fluctuation standard deviation , block fluctuation mean , block mean standard deviation , the ratio of high and low frequency energy , spectral peak energy concentration These seven eigenvalues ​​are combined into a 7-dimensional feature vector, which is then fed into the support vector machine. This 7-dimensional feature vector contains crucial information from both the time and frequency domains, fully describing every aspect of the waveform. In this way, the complex waveform information is compressed and provided as a numerical vector to the subsequent classification algorithm.

[0053] The present invention uses the RBF kernel (Radial Basis Function) as the kernel function for a support vector machine. The RBF kernel is a commonly used kernel function that effectively handles nonlinear classification problems. It improves classifier performance by mapping the input data into a high-dimensional space where nonlinear relationships between data points can be linearly segmented. Furthermore, the present invention uses GridSearchCV combined with cross-validation to automatically search for the optimal hyperparameter combination, resulting in the optimal SVM model. The final model, based on the input feature vector, generates classification results for the simulated waveform of the key signal in the regression test.

[0054] It should be noted that the above content merely illustrates the technical idea of ​​the present invention and cannot be used to limit the scope of protection of the present invention. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications all fall within the scope of protection of the claims of the present invention.

Claims

1. A regression test simulation waveform classification method based on Mamba neural network, characterized in that: The following steps are involved: Step 1: Parse the simulation waveform file obtained by regression testing and convert the key signal waveform into a discrete sequence of integer waveforms for subsequent waveform processing and classification; Step 2: Normalize the waveform sequence value range to the interval of 0-1, intercept the fixed length and input it into the Mamba neural network model. The network outputs two parameters: block size and block starting position. The method for normalizing the waveform sequence in step 2 and inputting it into the Mamba neural network model is as follows: each value in the waveform sequence is divided by the maximum value in the sequence to normalize the sequence values ​​to the range of 0-1, the waveform sequence is truncated to a length of 2048, and if the length is insufficient, the value 0 is padded at the end of the sequence. The processed waveform sequence is input into the Mamba neural network model, and the model outputs two parameters: block size L and block starting position K. Step 3: Use the two characteristic parameters output by the network to divide the waveform, calculate the thickness, fluctuation degree and mean characteristics of each waveform, and calculate the overall characteristics; Step 4: Using integer multiples of the waveform block size as a window, perform fast Fourier transform on the waveform and calculate the frequency domain features; Step 5: Combine the features extracted from the above process into a multi-dimensional feature vector and input it into the support vector machine to obtain the final classification result. The method for the support vector machine in step five to process features and obtain classification results is as follows: the seven eigenvalues ​​of block size L, block thickness mean, block thickness fluctuation standard deviation, block fluctuation degree mean, block mean standard deviation, high-frequency energy ratio, and spectral peak energy concentration are merged into a 7-dimensional feature vector, which is input into the support vector machine to finally obtain the output key signal waveform sequence classification result.

2. The regression test simulation waveform classification method based on Mamba neural network according to claim 1, wherein The method for processing the key signal waveform in step 1 is: reading each data bit value of the key signal at the rising edge of the base frequency clock, converting it into a decimal integer for storage, and looping the above steps to obtain a discrete sequence of a key waveform signal.

3. The regression test simulation waveform classification method based on Mamba neural network according to claim 1, wherein The method of waveform segmentation and feature extraction in step three is: according to the output results of the model, the waveform starts from the block starting position K, and is divided into waveform blocks of the same size according to the block size L. The waveform before the block starting position and the waveform with a length less than L at the end are discarded, and the thickness, fluctuation degree and mean characteristics of each block of waveform are calculated. According to the calculation results, the overall block thickness mean, block thickness fluctuation standard deviation, block fluctuation degree mean and block mean standard deviation are statistically calculated.

4. The regression test simulation waveform classification method based on Mamba neural network according to claim 1, wherein The frequency domain feature calculation method in step 4 is: use N times the block size L of the model output as a window, perform fast Fourier transform on the waveform sequence in the window, obtain an amplitude sequence of length L*N, calculate the ratio of high and low frequency energy of the amplitude sequence, and the energy concentration at the spectral peak position of the integer multiple of the amplitude sequence N.

Images