Chip performance grade prediction method

Abstract

The invention provides a chip performance grade prediction method, and relates to the technical field of chip detection.The chip performance grade prediction method comprises the steps that firstly, attribute parameters of a to-be-tested chip are obtained; thirdly, performing feature extraction on the attribute parameters of the to-be-tested chip by adopting a feature extraction model to obtain feature parameters; and finally, classifying the characteristic parameters by adopting an extreme learning machine to obtain a performance grade predicted value of the chip. Wherein the feature extraction model is a neural network model, the feature extraction model performs parameter optimization by adopting a preset parameter optimization method in advance, and the extreme learning machine performs training by adopting a restart strategy in advance. According to the embodiment of the invention, the feature extraction model is adopted to perform feature extraction on the collected attribute parameters of the chip, and then the extreme learning machine is adopted to classify the extracted feature parameters, so that the performance grade of the chip can be quickly and accurately obtained.

Description

Technical Field

[0001] The present invention relates to the field of chip detection technology, and in particular to a chip performance level prediction method. Background Art

[0002] The contents of this section are merely background art related to the present invention and do not necessarily constitute prior art.

[0003] Chips are the core components of modern electronic devices. Their quality and reliability directly affect the performance and service life of the entire electronic device. Therefore, chip testing is crucial for chip quality control. Chip testing generally includes function testing, performance testing and fault detection. Among them, performance testing refers to the testing of chip performance indicators (such as power consumption, running speed, etc.). Chip testing is usually performed by professionals on the chip entity with the help of professional testing instruments. However, the performance of a chip is essentially determined by its hardware indicators and software indicators. Therefore, in addition to using instruments to perform performance testing on the chip, various hardware indicators and software indicators (attribute parameters) of the chip can also be collected to evaluate the performance of the chip. However, a performance indicator of a chip is often affected by hundreds or thousands of attribute parameters, and the relationship between each attribute parameter and the performance indicator is particularly complex. Therefore, there is currently a lack of effective methods to predict the performance of the chip through the attribute parameters of the chip. Summary of the invention

[0004] The purpose of the embodiment of the present invention is to provide a chip performance level prediction method, which can quickly and accurately predict the performance of the chip according to the attribute parameters of the chip.

[0005] To achieve the above object, an embodiment of the present invention provides a chip performance level prediction method, comprising:

[0006] Obtaining the attribute parameters of the chip to be tested;

[0007] Using a feature extraction model to extract features of the attribute parameters of the chip to be tested to obtain feature parameters;

[0008] Using an extreme learning machine to classify the characteristic parameters to obtain a performance level prediction value of the chip;

[0009] Wherein, the feature extraction model is a neural network model, and the feature extraction model is pre-optimized using a preset parameter optimization method, and the extreme learning machine is pre-trained using a restart strategy.

[0010] As an improvement of the above solution, the restart strategy includes:

[0011] Randomly initialize the weights and biases of the hidden layer of the extreme learning machine, and set the parameters of the restart strategy; wherein the parameters of the restart strategy include the restart frequency and the reset ratio;

[0012] Iteratively train the extreme learning machine using input data to obtain output weights;

[0013] Regularly evaluate whether the improvement in prediction accuracy of the extreme learning machine is greater than the preset performance improvement threshold;

[0014] When the improvement is not greater than the performance improvement threshold, partial weights and biases of the hidden layer are reset according to the reset ratio, and the extreme learning machine is iteratively trained again.

[0015] As an improvement of the above solution, after resetting part of the weights and biases of the hidden layer according to the reset ratio, the method further includes:

[0016] Calculate confidence based on the prediction accuracy and stability of the extreme learning machine;

[0017] Parameters of the restart strategy are adjusted according to the confidence level.

[0018] As an improvement of the above solution, the iterative training of the extreme learning machine using the input data specifically includes:

[0019] Input the input data into the extreme learning machine to obtain the output weight;

[0020] Performing importance weighting processing on the input data according to the feature importance score to obtain weighted input data;

[0021] Inputting the weighted input data into an extreme learning machine to obtain an adjusted output weight;

[0022] The feature importance score is updated, and the step of performing importance weighting processing on the input data is returned to iteratively train the extreme learning machine.

[0023] As an improvement of the above scheme, the feature importance score is calculated by the following formula:

[0024]

[0025] Among them, FI(x i ) represents the i-th input data x i The feature importance score of ; Performance represents the iterative performance of the extreme learning machine.

[0026] As an improvement to the above solution, the feature importance score is updated by the following formula:

[0027]

[0028] Among them, FI(x i ) new Represents the updated input data x i The feature importance score of ; λ represents the smoothing parameter; FI(x i ) represents the input data x before updating i The feature importance score of Performance new Represents the current iteration performance of the extreme learning machine.

[0029] As an improvement of the above solution, the parameter optimization method includes:

[0030] Randomly initialize the position and speed of the parameters of the feature extraction model in the parameter space; wherein the position of the parameter corresponds to the value of the parameter;

[0031] Calculate the error according to the actual label value of the sample and the predicted value of the sample depending on the parameter, and calculate the fitness value of the parameter according to the error;

[0032] Calculating the air pressure difference according to the fitness value of the parameter;

[0033] updating the speed of the parameters according to the pressure difference, the inertia weight and the learning factor;

[0034] The position of the parameter is updated according to the updated speed of the parameter, and the step of calculating the fitness value is returned to perform iterative calculations until a preset number of iterations is reached.

[0035] As an improvement of the above solution, the parameter optimization method further includes updating the speed of the parameter by adjusting the inertia weight and the learning factor, and the inertia weight and the learning factor are calculated by the following formula:

[0036]

[0037] Among them, w1(t+1) represents the inertia weight at the t+1th iteration; w 1max Indicates the maximum value of inertia weight; w 1min represents the minimum value of the inertia weight; t represents the number of iterations; T represents the maximum number of iterations; δu represents the first adaptive environmental perception adjustment factor; e represents the natural base; ESF(t) represents the environmental perception factor; c1(t+1) represents the learning factor at the t+1th iteration; c 1max represents the maximum value of the learning factor; c 1minrepresents the minimum value of the learning factor; γu represents the second adaptive environment perception adjustment factor; F(p,t) represents the fitness of the feature extraction model under parameter p at the t-th iteration; F(p,t-1) represents the fitness of the feature extraction model under parameter p at the t-1th iteration.

[0038] As an improvement of the above solution, the parameter optimization method further includes updating the position of the parameter by introducing the mutation probability of the parameter, and the position of the mutated parameter is calculated by the following formula:

[0039] p′ i =p i +δc×(p max -p min )×θ(t)×P mut

[0040]

[0041] Among them, p′ i Indicates the position of the mutated parameter; p i represents the position of the parameter before mutation; δc represents a random number generated from the interval [-1, 1]; P mut represents the probability of mutation; p max Indicates the maximum value that the parameter may take; p min represents the minimum value that the parameter can take; θ(t) represents the wind direction variable of the tth iteration; θ max represents the maximum value of the wind direction variable; t represents the current number of iterations; T represents the maximum number of iterations.

[0042] As an improvement of the above scheme, the error is calculated by the following formula:

[0043]

[0044] Where E(p) represents the error; y n Represents the actual label value of the nth sample; It indicates the predicted value of the nth sample depending on the parameter p; N1 indicates the number of samples.

[0045] Compared with the prior art, the chip performance level prediction method provided by the embodiment of the present invention first obtains the attribute parameters of the chip to be tested; then, the attribute parameters of the chip to be tested are extracted using a feature extraction model to obtain feature parameters; finally, the feature parameters are classified using an extreme learning machine to obtain a performance level prediction value of the chip; wherein the feature extraction model is a neural network model, and the feature extraction model is pre-optimized using a preset parameter optimization method, and the extreme learning machine is pre-trained using a restart strategy. The embodiment of the present invention uses a feature extraction model to extract the attributes of the collected chip, and then uses an extreme learning machine to classify the extracted feature parameters, so as to quickly and accurately obtain the performance level of the chip; and the feature extraction model of the embodiment of the present invention uses a preset parameter optimization method to optimize parameters, and can search for the optimal solution in a wider parameter space, reducing the risk of falling into a local optimal solution; the extreme learning machine of the embodiment of the present invention uses a restart strategy for training, which can effectively improve the classification accuracy and efficiency of the extreme learning machine. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a flowchart of a chip performance level prediction method provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] See also Figure 1 , is a flow chart of a chip performance level prediction method provided by an embodiment of the present invention, comprising steps S1 to S3:

[0049] S1. Obtaining the attribute parameters of the chip to be tested;

[0050] S2, using a feature extraction model to extract features of the attribute parameters of the chip to be tested to obtain feature parameters;

[0051] S3, using an extreme learning machine to classify the characteristic parameters to obtain a performance level prediction value of the chip;

[0052] Wherein, the feature extraction model is a neural network model, and the feature extraction model is pre-optimized using a preset parameter optimization method, and the extreme learning machine is pre-trained using a restart strategy.

[0053] In step S1, the attribute parameters of the chip may include the chip model, the number of cores of the processor, the operating main frequency of the chip (in GHz), the average power consumption of the chip (in W), the process technology of the chip (in nm), the L1, L2, and L3 cache sizes of the processor (in MB), the input and output interface types supported by the chip, the maximum memory capacity supported by the processor (in GB), and the supported memory type (such as DDR4), etc.

[0054] It is worth noting that the feature extraction model used in step S2 is a neural network model. For example, a 3-layer fully connected neural network can be used as the feature extraction model. Furthermore, in order to enable the feature extraction model to accurately extract useful feature parameters from the attribute parameters to improve the accuracy of subsequent chip performance prediction, the embodiment of the present invention uses a preset parameter optimization method to optimize the parameters of the feature extraction model in advance.

[0055] In step S3, the extreme learning machine is used as a classifier, and the extreme learning machine can classify the performance level of the chip according to the feature parameters extracted by the feature extraction module. Further, in order to improve the prediction accuracy and efficiency of the extreme learning machine, the embodiment of the present invention trains the extreme learning machine based on the restart strategy.

[0056] It is worth noting that the performance indicator may be energy consumption, operating time, etc. The embodiment of the present invention predicts and evaluates one of the performance indicators of the chip each time, and divides the performance indicators into different levels according to certain division rules in advance.

[0057] Compared with the prior art, the embodiment of the present invention uses a feature extraction model to extract features from the collected attribute parameters of the chip, and then uses an extreme learning machine to classify the extracted feature parameters, so as to quickly and accurately obtain the performance level of the chip.

[0058] As an optional implementation manner, the parameter optimization rule includes:

[0059] Randomly initialize the position and speed of the parameters of the feature extraction model in the parameter space; wherein the position of the parameter corresponds to the value of the parameter;

[0060] Calculate the error according to the actual label value of the sample and the predicted value of the sample depending on the parameter, and calculate the fitness value of the parameter according to the error;

[0061] Calculating the air pressure difference according to the fitness value of the parameter;

[0062] updating the speed of the parameters according to the pressure difference, the inertia weight and the learning factor;

[0063] The position of the parameter is updated according to the updated speed of the parameter, and the step of calculating the fitness value is returned to perform iterative calculations until a preset number of iterations is reached.

[0064] It is worth noting that in some neural network structures, gradient vanishing, gradient exploding or falling into local optimal solutions may occur, which in turn affects the stability of training and the performance of the model. Inspired by the formation mechanism of wind in nature and its impact on the environment, an embodiment of the present invention proposes a parameter optimization method - an optimization algorithm based on airflow dynamics.

[0065] It is understandable that wind is generated by the flow of air, its direction and strength are determined by air pressure differences, terrain and other natural factors, and wind can carry energy and transmit information, and cause the circulation and renewal of matter in the ecosystem.

[0066] The parameter optimization method proposed in the embodiment of the present invention updates the parameters by simulating the flow of wind and the pressure difference, without relying on gradient information, so it can search in a wider parameter space, avoid falling into the local optimum and accelerate convergence. In addition, the optimization algorithm based on airflow dynamics introduces the concepts of "pressure difference" and "wind speed" (speed) as the main factors for adjusting parameters to simulate the impact of wind changes in nature on the environment, thereby being able to dynamically adjust the search strategy and enhance the global search capability and adaptability.

[0067] Specifically, the parameters of the feature extraction model (neural network model) are used as particles to initialize their positions and velocities in the parameter space. For example, the parameters p can be randomly initialized by formula (1) and formula (2): i (corresponding to the particle's position) and its velocity v i :

[0068] p i (0) = p min +rand(0,1)×(p max -p min ) (1)

[0069] v i (0) = v min +rand(0,1)×(v max -v min ) (2)

[0070] Among them, p min Indicates the minimum value that the parameter can take; p max Indicates the maximum value that the parameter may take; v min Indicates the minimum value of speed; v max Indicates the maximum value of the speed; rand(0,1) indicates a random number in the interval [0,1].

[0071] Further, a fitness value is calculated for each parameter (particle), and the fitness value is determined by the performance of the feature extraction model on a specific task. Exemplarily, the fitness function F(p) is the inverse of the error, and the error is obtained based on a preset logistic regression classifier. The fitness function can be expressed as shown in formula (3):

[0072]

[0073] Wherein, F(p) represents the fitness function; E(p) is the error function, which can be cross entropy loss for example.

[0074] Furthermore, the error function E(p) can be expressed as shown in formula (4):

[0075]

[0076] Where E(p) represents the error; y n Represents the actual label value of the nth sample; It indicates the predicted value of the nth sample depending on the parameter p; N1 indicates the number of samples.

[0077] Furthermore, the pressure difference is simulated according to the fitness value. The pressure difference ΔP is determined by the fitness value. A high fitness value corresponds to a low pressure, and a low fitness value corresponds to a high pressure. Specifically, a baseline fitness F is set. base , then the pressure difference can be expressed as shown in formula (5):

[0078] ΔP(p i )=βe×(F base -F(p i )) (5)

[0079] Among them, βe represents the adjustment factor; F base Indicates the highest fitness value in the population. Exemplarily, the adjustment factor βe is set to 1.5 to increase the effect of air pressure difference on parameter update.

[0080] Furthermore, the position and speed of each parameter are updated according to the pressure difference and wind speed (i.e., the step size of parameter update) to simulate the influence of wind flow on the position of particles (parameters). The new position p of the parameter is i (t+1) and the new velocity v i (t+1) is shown in equation (6) and equation (7):

[0081] v i (t+1)=w1×v i (t)+c1×rand(0,1)×(ΔP(p i )) (6)

[0082] p i (t+1)=p i (t)+v i (t+1) (7)

[0083] Wherein, w1 represents the inertia weight, which is used to control the influence of the previous speed on the current speed; c1 represents the learning factor, which is used to control the influence of the air pressure difference on the speed. For example, w1 can be set to 0.9 and c1 can be set to 2.

[0084] As an optional implementation manner, the parameter optimization method further includes updating the speed of the parameter by adjusting the inertia weight and the learning factor, and the inertia weight and the learning factor are calculated by the following formula:

[0085]

[0086] Among them, w1(t+1) represents the inertia weight at the t+1th iteration; w 1max Indicates the maximum value of inertia weight; w 1min represents the minimum value of the inertia weight; t represents the number of iterations; T represents the maximum number of iterations; δu represents the first adaptive environmental perception adjustment factor; e represents the natural base; ESF(t) represents the environmental perception factor; c1(t+1) represents the learning factor at the t+1th iteration; c 1max represents the maximum value of the learning factor; c 1min represents the minimum value of the learning factor; γu represents the second adaptive environment perception adjustment factor; F(p,t) represents the fitness of the feature extraction model under parameter p at the t-th iteration; F(p,t-1) represents the fitness of the feature extraction model under parameter p at the t-1th iteration.

[0087] It is worth noting that the wind speed (velocity) can be dynamically adjusted according to the feedback in the search process to adapt to different regions and characteristics of the parameter space. Specifically, the speed adjustment can be achieved by dynamically adjusting the inertia weight w1 and the learning factor c1. The adjustment method of the inertia weight and the learning factor is shown in equations (8) and (9). For example, w 1max Can be set to 0.9, w 1min Can be set to 0.4, c 1max Can be set to 2.5, c 1min Can be set to 0.5.

[0088] Furthermore, the environment perception factor ESF(t) can be calculated based on the rate of change of the objective function in consecutive iterations to indicate the adjustment range required for the parameter update strategy. Specifically, let F(p,t) be the fitness of the feature extraction model under parameter p at the tth iteration, and the environment perception factor is defined as the relative rate of change of fitness in two consecutive iterations, as shown in formula (10).

[0089] Further, referring to equations (8) and (9), e -γ×ESF(t) and e -δ×ESF(t) Introducing a nonlinear factor through an exponential decay function to adjust the wind speed can ensure that the adjustment range is more significant when the environmental changes are large.

[0090] As an optional implementation, the parameter optimization method further includes updating the position of the parameter by introducing a probability of parameter mutation, and the position of the mutated parameter is calculated by the following formula:

[0091] p′ i =p i +δc×(p max -p min )×θ(t)×P mut (11)

[0092]

[0093] Among them, p′ i Indicates the position of the mutated parameter; p i represents the position of the parameter before mutation; δc represents a random number generated from the interval [-1, 1]; P mut represents the probability of mutation; p max Indicates the maximum value that the parameter may take; p min represents the minimum value that the parameter can take; θ(t) represents the wind direction variable of the tth iteration; θ max represents the maximum value of the wind direction variable; t represents the current number of iterations; T represents the maximum number of iterations.

[0094] It is worth noting that by introducing mutations to simulate the turbulence effect in natural wind, the diversity of the parameter population can be maintained and premature convergence can be avoided. Specifically, each parameter p i There is a certain probability P mut Mutation occurs, P mut It can be set to 0.1, and the mutated parameter value can be expressed as shown in formula (11).

[0095] Furthermore, the wind direction variable θ is used to simulate the change of wind direction in the real world, thereby increasing the diversity of the search process. For example, θ max Set to 30 degrees.

[0096] As an optional implementation manner, the error is calculated by the following formula:

[0097]

[0098] Where E(p) represents the error; y n Represents the actual label value of the nth sample; It indicates the predicted value of the nth sample depending on the parameter p; N1 indicates the number of samples.

[0099] Compared with the prior art, the embodiment of the present invention optimizes the parameters of the feature extraction model by adopting a preset parameter optimization method, which can search for the optimal solution in a wider parameter space, reduce the risk of falling into a local optimal solution, and improve the reliability and adaptability of the model.

[0100] It is understandable that after the feature extraction model training is completed, it can be used to extract features from the attribute parameters of the chip, and then the extracted feature parameters are input into the classifier for classification. Specifically, the embodiment of the present invention adopts an extreme learning machine based on a restart strategy as a classifier. Specifically, the embodiment of the present invention introduces a restart strategy during the training process of the extreme learning machine to avoid falling into a local optimum: when the performance improvement during the training process is not significant, some weights and biases of the hidden layer are randomly reset (ie, "restarted"). At the same time, during each restart process, a confidence evaluation mechanism is introduced to measure the prediction reliability of the current model. Models with high confidence are less affected by the restart strategy.

[0101] For ease of understanding, some parameters appearing in the embodiments of the present invention are explained below:

[0102] N represents the number of hidden layer nodes; w represents the hidden layer weight; b represents the hidden layer bias; rand(-1,1,size=N) represents the generation of random numbers ranging from -1 to 1 and the number is N; β represents the output layer weight of the extreme learning machine model; represents the Moore-Penrose pseudo-inverse of the hidden layer output matrix H; T represents the target output of the training data; H represents the output matrix of the hidden layer; g represents the activation function; X represents the input data; ΔPerformance represents the difference in performance improvement; θ represents the performance improvement threshold; ρ represents the ratio of reset weights and biases; w reset represents the weight after reset; b reset Represents the deviation after reset; C represents the confidence of the model; α represents the balance coefficient, which is used to weigh the proportion of accuracy and stability; Accuracy represents the accuracy of the model on the validation set; Stability represents the stability index of the model performance.

[0103] As an optional implementation manner, the restart strategy includes:

[0104] Randomly initialize the weights and biases of the hidden layer of the extreme learning machine, and set the parameters of the restart strategy; wherein the parameters of the restart strategy include the restart frequency and the reset ratio;

[0105] Iteratively train the extreme learning machine using input data to obtain output weights;

[0106] Regularly evaluate whether the improvement in prediction accuracy of the extreme learning machine is greater than the preset performance improvement threshold;

[0107] When the improvement is not greater than the performance improvement threshold, partial weights and biases of the hidden layer are reset according to the reset ratio, and the extreme learning machine is iteratively trained again.

[0108] Exemplarily, the initialization formulas for the hidden layer weights w and bias b are shown in equations (13) and (14):

[0109] w=rand(-1,1,size=N) (13)

[0110] b=rand(-1,1,size=N) (14)

[0111] Among them, w represents the hidden layer weight; b represents the hidden layer bias; rand(-1,1,size=N) represents generating N random numbers between -1 and 1.

[0112] As an optional implementation manner, the iterative training of the extreme learning machine using the input data specifically includes:

[0113] Input the input data into the extreme learning machine to obtain the output weight;

[0114] Performing importance weighting processing on the input data according to the feature importance score to obtain weighted input data;

[0115] Inputting the weighted input data into an extreme learning machine to obtain an adjusted output weight;

[0116] The feature importance score is updated, and the step of performing importance weighting processing on the input data is returned to iteratively train the extreme learning machine.

[0117] Exemplarily, the calculation formula of the output weight β of the extreme learning machine model is shown in formula (15):

[0118]

[0119] Among them, β represents the output weight; represents the Moore-Penrose pseudo-inverse of the hidden layer output matrix H; T represents the target output of the training data.

[0120] Furthermore, the calculation formula of the output matrix H is shown in formula (16):

[0121] H=g(w·X+b) (16)

[0122] Among them, g represents the activation function; X represents the input data.

[0123] Furthermore, it is necessary to perform importance weighting on the input data to obtain weighted input data, as shown in formula (17):

[0124] X′=X⊙FI(X) (17)

[0125] Among them, X′ represents the weighted input data; ⊙ represents element-by-element multiplication, that is, each eigenvalue (input data) is multiplied by its corresponding importance score; FI(X) represents the feature importance score of the input data.

[0126] Furthermore, the weighted input data is used to train the extreme learning machine to obtain the adjusted output weights, as shown in formula (18):

[0127]

[0128] Among them, β′ represents the adjusted output weight; H′ represents the hidden layer output matrix obtained by training with the weighted input data X′.

[0129] Furthermore, after each iterative training, the feature importance scores need to be dynamically updated according to the new model performance.

[0130] As an optional implementation, the feature importance score is calculated using the following formula:

[0131]

[0132] Among them, FI(x i ) represents the i-th input data x i The feature importance score of ; Performance represents the iterative performance of the extreme learning machine.

[0133] It is worth noting that the feature importance score is used to evaluate the importance of each input data. In formula (19), It represents the partial derivative of the model performance index with respect to the i-th input data, which is used to reflect the influence of the input change on the model performance. The model performance index (iterative performance) is specifically the prediction accuracy.

[0134] As an optional implementation, the feature importance score is updated by the following formula:

[0135]

[0136] Among them, FI(x i ) new Represents the updated input data x i The feature importance score of ; λ represents the smoothing parameter; FI(x i ) represents the input data x before updating i The feature importance score of Performance new Represents the current iteration performance of the extreme learning machine.

[0137] It is worth noting that in formula (20), the smoothing parameter λ is used to balance the influence between the old feature importance score and the new feature importance score.

[0138] Furthermore, in order to improve the accuracy of chip performance level classification, the embodiment of the present invention also regularly evaluates the performance of the extreme learning machine. If the performance improvement is not significant, some hidden layer weights and biases are reset according to the restart strategy. Specifically, the performance improvement can be calculated by formula (21):

[0139] ΔPerformance=Performance t -Performance t-1 (twenty one)

[0140] Among them, ΔPerformance represents the difference in performance improvement; Performance t Indicates the performance after the tth iteration; Performance t-1 represents the performance after the t-1th iteration.

[0141] Furthermore, the performance improvement threshold is set to θ. If ΔPerformance<θ, some hidden layer weights and biases are reset. This is the meaning of “restart” in the “restart strategy”.

[0142] Furthermore, during the “restart” process, some weights and biases are randomly selected for reset, and the ratio of the number of reset weights / biases to the number of hidden layer nodes is ρ.

[0143] Exemplarily, the hidden layer weights and biases are reset by equations (22) and (23):

[0144] w reset =rand(-1,1,size=ρ·N) (22)

[0145] breset =rand(-1,1,size=ρ·N) (23)

[0146] Among them, w reset represents the weight after reset; b reset represents the bias after reset; rand(-1,1,size=ρ·N) represents the generation of random numbers ranging from -1 to 1 and the number of which is "ρ·N"; ρ represents the ratio of reset weight and bias; N represents the number of hidden layer nodes.

[0147] As an optional implementation manner, after resetting part of the weights and biases of the hidden layer according to the reset ratio, the method further includes:

[0148] Calculate confidence based on the prediction accuracy and stability of the extreme learning machine;

[0149] Parameters of the restart strategy are adjusted according to the confidence level.

[0150] It is worth noting that, in the embodiment of the present invention, after each "restart", that is, after each reset of some hidden layer weights and biases, it is also necessary to evaluate the confidence of the extreme learning machine to adjust the parameters of the restart strategy according to the confidence, for example, to reduce the restart frequency of the extreme learning machine with high confidence.

[0151] Exemplarily, the confidence calculation formula is shown in formula (24):

[0152] C=α·Accuracy+(1-α)·Stability (24)

[0153] Among them, C represents the confidence of the model; α represents the balance coefficient, which is used to weigh the weight of accuracy and stability; Accuracy represents the accuracy of the model on the validation set; Stability represents the stability index of the model performance.

[0154] Furthermore, the trained extreme learning machine and the feature extraction model are used together for chip performance level evaluation, wherein the extreme learning machine is used to classify the feature parameters extracted by the feature extraction model and output the performance level to which the chip belongs. Compared with the prior art, the embodiment of the present invention can improve the classification accuracy and efficiency of the extreme learning machine by introducing a restart strategy and confidence level in the training process of the extreme learning machine.

[0155] Compared with the prior art, the chip performance level prediction method provided by the embodiment of the present invention first obtains the attribute parameters of the chip to be tested; then, the attribute parameters of the chip to be tested are extracted using a feature extraction model to obtain feature parameters; finally, the feature parameters are classified using an extreme learning machine to obtain a performance level prediction value of the chip; wherein the feature extraction model is a neural network model, and the feature extraction model is pre-optimized using a preset parameter optimization method, and the extreme learning machine is pre-trained using a restart strategy. The embodiment of the present invention uses a feature extraction model to extract the attributes of the collected chip, and then uses an extreme learning machine to classify the extracted feature parameters, so as to quickly and accurately obtain the performance level of the chip; and the feature extraction model of the embodiment of the present invention uses a preset parameter optimization method to optimize parameters, and can search for the optimal solution in a wider parameter space, reducing the risk of falling into a local optimal solution; the extreme learning machine of the embodiment of the present invention uses a restart strategy for training, which can effectively improve the classification accuracy and efficiency of the extreme learning machine.

[0156] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A chip performance level prediction method, characterized in that: include: Obtaining the attribute parameters of the chip to be tested; Using a feature extraction model to extract features of the attribute parameters of the chip to be tested to obtain feature parameters; Using an extreme learning machine to classify the characteristic parameters to obtain a performance level prediction value of the chip; Wherein, the feature extraction model is a neural network model, and the feature extraction model is pre-optimized using a preset parameter optimization method, and the extreme learning machine is pre-trained using a restart strategy.

2. The chip performance level prediction method according to claim 1, characterized in that: The restart strategy includes: Randomly initialize the weights and biases of the hidden layer of the extreme learning machine, and set the parameters of the restart strategy; wherein the parameters of the restart strategy include the restart frequency and the reset ratio; Iteratively train the extreme learning machine using input data to obtain output weights; Regularly evaluate whether the improvement in prediction accuracy of the extreme learning machine is greater than the preset performance improvement threshold; When the improvement is not greater than the performance improvement threshold, partial weights and biases of the hidden layer are reset according to the reset ratio, and the extreme learning machine is iteratively trained again.

3. The chip performance level prediction method according to claim 2, characterized in that: After resetting part of the weights and biases of the hidden layer according to the reset ratio, the method further includes: Calculate confidence based on the prediction accuracy and stability of the extreme learning machine; Parameters of the restart strategy are adjusted according to the confidence level.

4. The chip performance level prediction method according to claim 2, characterized in that: The iterative training of the extreme learning machine using the input data specifically includes: Input the input data into the extreme learning machine to obtain the output weight; Performing importance weighting processing on the input data according to the feature importance score to obtain weighted input data; Inputting the weighted input data into an extreme learning machine to obtain an adjusted output weight; The feature importance score is updated, and the step of performing importance weighting processing on the input data is returned to iteratively train the extreme learning machine.

5. The chip performance level prediction method according to claim 4, characterized in that: The feature importance score is calculated using the following formula: Among them, FI(x i ) represents the i-th input data x i The feature importance score of ; Performance represents the iterative performance of the extreme learning machine.

6. The chip performance level prediction method according to claim 4, characterized in that: The feature importance score is updated by the following formula: Among them, FI(x i ) new Represents the updated input data x i The feature importance score of ; λ represents the smoothing parameter; FI(x i ) represents the input data x before updating i The feature importance score of Performance new Represents the current iteration performance of the extreme learning machine.

7. The chip performance level prediction method according to claim 1, characterized in that: The parameter optimization method comprises: Randomly initialize the position and speed of the parameters of the feature extraction model in the parameter space; wherein the position of the parameter corresponds to the value of the parameter; Calculate the error according to the actual label value of the sample and the predicted value of the sample depending on the parameter, and calculate the fitness value of the parameter according to the error; Calculating the air pressure difference according to the fitness value of the parameter; updating the speed of the parameters according to the pressure difference, the inertia weight and the learning factor; The position of the parameter is updated according to the updated speed of the parameter, and the step of calculating the fitness value is returned to perform iterative calculations until a preset number of iterations is reached.

8. The chip performance level prediction method according to claim 7, characterized in that: The parameter optimization method further includes updating the speed of the parameter by adjusting the inertia weight and the learning factor, and the inertia weight and the learning factor are calculated by the following formula: Among them, w1(t+1) represents the inertia weight at the t+1th iteration; w 1max Indicates the maximum value of inertia weight; w 1min represents the minimum value of the inertia weight; t represents the number of iterations; T represents the maximum number of iterations; δu represents the first adaptive environmental perception adjustment factor; e represents the natural base; ESF(t) represents the environmental perception factor; c1(t+1) represents the learning factor at the t+1th iteration; c 1max represents the maximum value of the learning factor; c 1min represents the minimum value of the learning factor; γu represents the second adaptive environment perception adjustment factor; F(p,t) represents the fitness of the feature extraction model under parameter p at the t-th iteration; F(p,t-1) represents the fitness of the feature extraction model under parameter p at the t-1th iteration.

9. The chip performance level prediction method according to claim 7, characterized in that: The parameter optimization method further includes updating the position of the parameter by introducing a probability of parameter mutation, and the position of the mutated parameter is calculated by the following formula: p i ′ =p i +δc×(p max -p min )×θ(t)×P mut Among them, p i ′ Indicates the position of the mutated parameter; p i represents the position of the parameter before mutation; δc represents a random number generated from the interval [-1, 1]; P mut represents the probability of mutation; p max Indicates the maximum value that the parameter may take; p min represents the minimum value that the parameter can take; θ(t) represents the wind direction variable of the tth iteration; θ max represents the maximum value of the wind direction variable; t represents the current number of iterations; T represents the maximum number of iterations.

10. The chip performance level prediction method according to claim 7, characterized in that: The error is calculated by the following formula: Where E(p) represents the error; y n Represents the actual label value of the nth sample; It indicates the predicted value of the nth sample depending on the parameter p; N1 indicates the number of samples.

Images