An integrated circuit control system based on artificial intelligence

Abstract

The application discloses an integrated circuit control system based on artificial intelligence, which comprises a data acquisition module, a feature extraction module, a control decision mapping module, a loss design module, an integrated circuit control model establishment module and an integrated circuit control module. The application belongs to the field of circuit control and specifically relates to an integrated circuit control system based on artificial intelligence. The scheme adopts a double-branch parallel architecture, locks burst, peak and slow change trends, can quickly respond to large control instruction changes through control decision mapping design, provides smooth nonlinear feedback for small errors and stable states, avoids excessive adjustment and unnecessary fluctuations, adopts a logarithmic function to inhibit the effect when accidental sensor noise and burst transient states occur, thereby avoiding control jitter, can provide appropriate feedback signals for the control system in a small error area, and further improves the integrated circuit control effect.

Description

TECHNICAL FIELD

[0001] The application relates to the field of circuit control, and particularly relates to an integrated circuit control system based on artificial intelligence. BACKGROUND

[0002] The integrated circuit control system is a system based on integrated circuit technology, which monitors and analyzes the running state of a circuit, and adjusts and controls the circuit parameters according to the analysis result, so as to realize specific functions and performance requirements. However, the general integrated circuit control system has the problems of insufficient capturing ability for local drastic changes of the integrated circuit, insufficient description of long-term trends and slow change characteristics, lack of stability strategy for large-scale control, and thus poor integrated circuit control effect; the general integrated circuit control system is sensitive to occasional sensor noise and sudden transient anomalies in the sensor signal, and has poor feedback in the small error area, and thus the integrated circuit control is jittered. SUMMARY

[0003] In view of the above problems, the application provides an integrated circuit control system based on artificial intelligence, which can overcome the defects of the prior art. The general integrated circuit control system has the problems of insufficient capturing ability for local drastic changes of the integrated circuit, insufficient description of long-term trends and slow change characteristics, lack of stability strategy for large-scale control, and thus poor integrated circuit control effect. The present application adopts a double-branch parallel architecture, which can quickly lock sudden peaks and spikes through a small core + attention, and can lock slow trends through a large core + channel gating. Based on multi-level global-local feature enhancement, irrelevant redundancy is removed, the fast response capability is retained, and the slow trend and periodic change of the integrated circuit are captured. Through control decision mapping design, the control system can quickly respond to large-scale control instruction changes and enhance the response capability of the system in the case of device anomaly, fault repair or high temperature alarm. For small errors and stable states, the control system provides smooth nonlinear feedback, avoids excessive adjustment and unnecessary fluctuations, and improves the stability of the integrated circuit control system. The general integrated circuit control system is sensitive to occasional sensor noise and sudden transient anomalies in the sensor signal, and has poor feedback in the small error area, and thus the integrated circuit control is jittered. The present application constructs an adjustable smooth loss function. When occasional sensor noise and sudden transients occur and the error fluctuates greatly, a logarithmic function is used for suppression, so that the control system is not sensitive to these anomalies and the control jitter is avoided. In the small error area, the control system can provide appropriate feedback signals, and thus the integrated circuit control effect is improved.

[0004] The technical scheme adopted by the present application is as follows: the present application provides an integrated circuit control system based on artificial intelligence, comprising a data acquisition module, a feature extraction module, a control decision mapping module, a loss design module, an integrated circuit control model establishment module and an integrated circuit control module;

[0005] The data acquisition module acquires internal historical signals of the integrated circuit and constructs a historical circuit sample set;

[0006] The feature extraction module captures local mutations and global long-range trends based on a double-branch parallel architecture and adaptively fuses the features of the two;

[0007] The control decision mapping module flattens the fused features into a vector, inputs the vector into a lightweight control network, and maps the network output to actual circuit control instructions through an activation function;

[0008] The loss design module designs a loss function based on the actual response and the expected response of the circuit;

[0009] The integrated circuit control model establishment module establishes an integrated circuit control model based on the historical circuit sample set, in combination with the feature extraction module, the control decision mapping module and the loss design module;

[0010] The integrated circuit control module controls real-time integrated circuits based on the established integrated circuit control model.

[0011] Further, the data acquisition module acquires internal historical signals of the integrated circuit and determines an expected response, taking the expected response as a data label; the historical circuit sample set is constructed by dividing a sliding window with a length of T.

[0012] Further, the feature extraction module extracts features from the input X of the historical circuit sample set, including two parallel branches of a local fine branch and a global receptive branch; the local fine branch adopts a small convolution kernel with a size of k1xk1, followed by channel attention and spatial reconstruction, represented as: Wherein, F1 is the output of the local branch; CR(·) is a Channel-attention+Spatial-reconstruction module; is a convolution operation with a convolution kernel of k1xk1; the global receptive branch adopts a large convolution kernel with a size of k2xk2 and combines a learnable channel weight, represented as: Wherein, F2 is the output of the global branch; W gate is a learnable channel weight vector; (·) is a convolution operation with a convolution kernel of k2xk2; prior feature fusion is represented as: F comb =F1+F2; wherein, Fcomb is the prior fusion output; the abnormal response channel is reinforced by the global branch, denoted as: X G =CIBlock k (GMP(F comb ));wherein X G is the feature after the global branch reinforcement; GMP(·) is the global maximum pooling; CIBlock k (·) is the cascaded k times CIBlock; CIBlock consists of: pointwise convolution→PReLU→Dropout→BatchNorm; the local branch promotion channel interaction is denoted as: X L =CIBlock k (F comb );wherein X L is the fusion feature after the local branch processing; the fusion weight is outputted, and the formula used is: X′ G =Repeat(X G ,size(X L ));X concat =Concat(X G ,X′ G );X fused =PWConv(X concat );W=σ s (X fused ); wherein size(·) is the shape information returned; X′ G is the broadcasted global feature; X out is the fusion feature tensor; Repeat(X G ,size(X L )) is to copy the global feature X G to the same shape as X L ; X concat is the splicing result; PWConv(·) is the 1×1 convolution; X fused is to map the spliced feature back to the original channel dimension to obtain the fusion feature representation; Concat(X G ,X′ G ) is to splice the original and broadcasted global features in the channel dimension; W=σ(X fused ) is the adaptive fusion coefficient tensor; σ s (·) is the Sigmoid activation function.

[0013] Further, the control decision mapping module flattens X out into a vector x, inputs a lightweight control network, denoted as: x=Flatten(X out); u = CtrlNet(x); where, Flatten(·) is a flattening operation; CtrlNet(·) is a function mapping of the lightweight control network; u is a multi-channel control instruction, each component u j corresponds to a physical control quantity; and u j is mapped to an actual control instruction by an activation function mapping, denoted as: wherein, is an actual control instruction, mapped to an actual control command range; sigma(·) is an activation function mapping; lambda, a and epsilon control the shape and characteristics of the activation function; U max and U min are the maximum and minimum values of the control command range, respectively.

[0014] Further, the loss design module takes the sensor signal waveform actually exhibited by the circuit after the actual control command is applied as an actual response for the ith sample, and sets the error between the actual response and the expected response as r i , and constructs an adjustable smooth loss L, using the formula: The overall gradient is denoted as: wherein, N is the total number of samples; k is a sensitivity parameter; and r is the overall error.

[0015] Further, the integrated circuit control model establishment module divides the historical circuit sample set into a test set and a training set, and constructs an integrated circuit control model based on the feature extraction module, the control decision mapping module and the loss design module. An error threshold is set, and when the loss of the integrated circuit control model on the training set converges or reaches the maximum training number, the integrated circuit control model training is completed. The error index of the actual response and the expected response of the test set of the integrated circuit control model trained is calculated, and if the error index is less than the corresponding threshold, the integrated circuit control model establishment is completed, otherwise the model parameters are adjusted for retraining.

[0016] Further, the integrated circuit control module controls the integrated circuit based on the actual control command obtained by the integrated circuit control model based on the established integrated circuit control model, and inputs the internal signals of the integrated circuit to the integrated circuit control model in real time.

[0017] The above scheme has the following beneficial effects:

[0018] (1) For the general integrated circuit control system, the capture ability of the local sharp change of the integrated circuit is insufficient, the long-term trend and slow change characteristics are insufficient, and the stability strategy for large amplitude control is lacking, thereby leading to the problem of poor integrated circuit control effect, the scheme adopts a double-branch parallel architecture, and the small core + attention is used to quickly lock the burst and peak, and the large core + channel gate is used to lock the slow trend; Based on multi-level global-local feature enhancement, irrelevant redundancy is removed, fast response capability is retained, and global slow trend and periodic change of integrated circuit are captured; Through control decision mapping design, for large amplitude control instruction change, it can quickly respond and enhance the response ability of the system in the case of device exception, fault repair or high temperature alarm; For small errors and stable state, provide smooth nonlinear feedback, avoid excessive adjustment and unnecessary fluctuations, and improve the stability of the integrated circuit control system.

[0019] (2) For the general integrated circuit control system, the occurrence of sensor signal noise and burst transient anomaly is sensitive, the feedback in the small error area is poor, and the integrated circuit control is dithered, the scheme constructs an adjustable smooth loss function, when the occurrence of sensor noise and burst transient anomaly, the error produces large fluctuation, the logarithmic function is used for suppression, so that the control system is not sensitive to these anomalies, and the control dithering is avoided; In the small error area, the control system can provide appropriate feedback signal; And improve the integrated circuit control effect. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 A flowchart of an integrated circuit control system based on artificial intelligence is provided.

[0021] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation on the present application. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0023] In the description of the present application, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0024] Embodiment one, refer to Figure 1 The present application provides an integrated circuit control system based on artificial intelligence, comprising a data acquisition module, a feature extraction module, a control decision mapping module, a loss design module, an integrated circuit control model establishment module and an integrated circuit control module.

[0025] The data acquisition module acquires internal historical signals of the integrated circuit, constructs a historical circuit sample set, and sends the data to the feature extraction module.

[0026] The feature extraction module captures local mutations and global long-range trends based on a double-branch parallel architecture, and adaptively fuses the features of the two; and sends the data to the control decision mapping module.

[0027] The control decision mapping module flattens the fused features into a vector, inputs the lightweight control network, and maps the network output to the actual control instruction of the circuit through the activation function; and sends the data to the loss design module.

[0028] The loss design module designs a loss function based on the actual response and the expected response of the circuit; and sends the data to the integrated circuit control model establishment module.

[0029] The integrated circuit control model establishment module establishes an integrated circuit control model based on the historical circuit sample set, in combination with the feature extraction module, the control decision mapping module and the loss design module; and sends the data to the integrated circuit control module.

[0030] The integrated circuit control module controls the real-time integrated circuit based on the established integrated circuit control model.

[0031] Embodiment two, refer to Figure 1 Based on the above embodiment, the data acquisition module acquires the internal historical signals of the integrated circuit and determines the expected response, taking the expected response as the data label, and the historical signals include voltage waveform, current, temperature and frequency; a sliding window with a length T is used for segmentation, and a historical circuit sample set is constructed; wherein each input matrix of the historical circuit sample set is regarded as a two-dimensional matrix of sensor channel x sampling time, taking into account the interaction of different sensor channels and time modes; the expected response is the target waveform of each channel of the integrated circuit under ideal control.

[0032] Embodiment three, see Figure 1 , based on the above embodiment, the feature extraction module captures the sudden event and long-range trend through a double-branch parallel architecture; the feature extraction is performed on the input X of the historical circuit sample set, including two parallel branches of a local fine branch and a global receptive branch; the local fine branch adopts a small convolution kernel with a size of 3x3, followed by channel attention and spatial reconstruction, which focuses on capturing local mutations of the sudden current peak and temperature surge of the sensing signal, and is expressed as: F1 = CR(Conv3(X)); wherein F1 is the output of the local branch, which mainly represents short-time peaks and surges; CR(·) is a Channel-attention+Spatial-reconstruction module; Conv3(·) is a convolution operation with a 3x3 convolution kernel; the global receptive branch adopts a large convolution kernel with a size of 7x7, and combines with a learnable channel weight, which focuses on the global periodicity and long-range dependence of slow-changing hot spots and frequency drift trends, and is expressed as: F2 = W gate ⊙Conv7(X); wherein F2 is the output of the global branch; W gate is a learnable channel weight vector, which has the same shape as the channel number; Conv7(·) is a convolution operation with a 7x7 convolution kernel; prior feature fusion is expressed as: F comb = F1 + F2; wherein F comb is the prior fusion output; the most intense abnormal response channel is strengthened through the global branch, which is expressed as: X G = CIBlock k (GMP(F comb )); wherein X G is the feature after the global branch strengthening; GMP(·) is a global maximum pooling; CIBlock k (·) is a k-time CIBlock in cascade; CIBlock is composed of point-by-point convolution->PReLU->Dropout->BatchNorm; the channel interaction of the local branch is promoted, which is expressed as: X L = CIBlock k (F comb ); wherein X L is the fusion feature after the local branch processing; the fusion weight is outputted, and the formula used is: X′ G = Repeat(X G , size(X L )); X concat = Concat(X G , X′ G ); X fused = PWConv(X concat ); W = σ s (X fused); where size(·) returns shape information; X G is the broadcasted global feature; X out is the fused feature tensor, capturing both local and global information of the integrated circuit; Repeat(X G , size(X L )) is to replicate the global feature X G to the same shape as X L ; X concat is the concatenation result; PWConv(·) is a 1x1 convolution; X fused is to map the concatenated feature back to the original channel dimension, resulting in the fused feature representation; Concat(X G , X G ) is to concatenate the original and broadcasted global features in the channel dimension; W = σ(X fused ) is the adaptive fusion coefficient tensor; σ s (·) is the Sigmoid activation function; By removing irrelevant redundancy through low-level convolution, saving edge computing power, high-level large receptive field capturing slow trend, early warning hot spot, adaptive fusion according to the current dynamic distribution of attention, taking into account fast response and global stability.

[0033] Embodiment four, see Figure 1 , this embodiment is based on the above embodiment, the control decision mapping module to X out is flattened into a vector x, input into the lightweight control network, denoted as: x = Flatten(X out ); u = CtrlNet(x); where Flatten(·) is the flattening operation; CtrlNet(·) is the function mapping of the lightweight control network; u is the multi-channel control instruction, each component u j corresponds to a physical control quantity, including voltage regulation, current size, temperature threshold, switch signal; the activation function mapping is performed on u j to the actual control instruction, denoted as: where, is the actual control instruction, mapped to the actual control command range; σ(·) is the activation function mapping; λ, a and ε control the shape and characteristics of the activation function; U max and U min are the maximum and minimum values of the control command range, respectively; to ensure the differentiability of the three segments, take In many integrated circuit control systems, there are large control instructions, including rapid adjustment of voltage and current, especially in the case of device anomalies, fault repair or high temperature alarm, the response to these large input changes is slow, and the scheme enhances the response ability to such sudden signals through the linear part at both ends; and provides smooth nonlinear feedback for small errors or steady state, which is crucial for the stability of the control system; In practical applications, integrated circuit control systems need to maintain small errors and quickly converge to the target state when running smoothly, which can avoid excessive adjustment and unnecessary fluctuations.

[0034] By performing the above operation, for the general integrated circuit control system, there is a lack of capture ability for local sharp changes of integrated circuits, a lack of description of long-term trends and slow-changing characteristics, a lack of stability strategy for large-scale control, which leads to poor integrated circuit control effect, the scheme adopts a dual-branch parallel architecture, and through small core + attention quickly locks sudden, peak, large core + channel gate lock slow trend; Based on multi-level global-local feature enhancement, irrelevant redundancy is removed, both fast response ability is retained, and slow-changing trends and periodic changes of integrated circuits are captured; Through control decision mapping design, for large-scale control instruction changes, it can quickly respond and enhance the system's ability to cope with device anomalies, fault repair or high temperature alarm; and provides smooth nonlinear feedback for small errors and steady state, avoiding excessive adjustment and unnecessary fluctuations, and improving the stability of the integrated circuit control system.

[0035] Embodiment five, refer to Figure 1 Based on the above embodiment, the loss design module for the i-th sample applies the sensor signal waveform actually exhibited by the circuit after the actual control command is applied as the actual response, and sets the error between the actual response and the expected response as r i And build adjustable smooth loss L, the formula is: The overall gradient is represented as: Where N is the total number of samples; k is the sensitivity parameter; r is the overall error; the gradient is saturated to ±k for occasional sensor noise and sudden transient, avoiding control jitter caused by abnormal readings; provide smooth nonlinear feedback in the small error area, accelerate convergence and ensure control accuracy.

[0036] By performing the above operation, for the general integrated circuit control system, there is a problem of being sensitive to incidental sensor noise and sudden transient anomalies in the sensor signal, poor feedback to small error area, and further leading to integrated circuit control jitter. The scheme constructs a adjustable smoothing loss function. When incidental sensor noise and sudden transient anomalies occur, and the error produces large fluctuations, a logarithmic function is used for suppression, so that the control system is not sensitive to these anomalies, avoiding control jitter; and in the small error area, a suitable feedback signal can be provided for the control system; and the integrated circuit control effect is improved.

[0037] Embodiment six, refer to Figure 1 This embodiment is based on the above embodiment. The integrated circuit control module collects the internal signals of the integrated circuit in real time based on the established integrated circuit control model, and inputs them to the integrated circuit control model. The actual control command obtained based on the model is used for integrated circuit control.

[0038] Embodiment seven, refer to Figure 1 This embodiment is based on the above embodiment. The control command range [U min ,U max ] = [0V, 5V] is selected; the transition bandwidth a = 2.5, the smoothing parameter ε = 1.0; the critical slope; first calculate

[0039] Case A: small amplitude output u j = 1.0, falls in [-a, a], σ(u j ) = 0.7311; In the middle nonlinear interval, the output is not strictly linear, nor is it completely saturated, and it can respond sensitively to fine control signals of the fine-tuning current or temperature threshold;

[0040] Case B: positive mutation u j = 3.0, greater than a, σ(u j ) = 0.9591; Enter the linear tail, the output is very close to the upper limit but not completely saturated, leaving a margin of 0.2V, which helps the hardware to tolerate the limit command;

[0041] Case C: negative mutation u j = -3.0, less than -a, σ(u j ) = 0.0409; Corresponding to the slight lifting near the lower limit, it helps to avoid the hardware dead zone caused by the command completely returning to zero.

[0042] Although embodiments of the present application have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present application.

[0043] The above describes the present application and its embodiments, which are not limited, and the drawings only show one of the embodiments of the present application, and the actual structure is not limited thereto. In general, if a person skilled in the art is inspired thereby, without departing from the purpose of the present application, without creative design, similar structure and embodiments of the technical solution are not creative, and should belong to the protection scope of the present application.

Claims

1. An artificial intelligence-based integrated circuit control system, characterized by: The system comprises a data acquisition module, a feature extraction module, a control decision mapping module, a loss design module, an integrated circuit control model establishment module and an integrated circuit control module. The data acquisition module acquires internal historical signals of the integrated circuit and constructs a historical circuit sample set. The feature extraction module captures local mutations and global long-range trends based on a double-branch parallel architecture and adaptively fuses the features of the two. The control decision mapping module flattens the fused features into a vector, inputs the vector into a lightweight control network, and maps the network output to actual circuit control instructions through an activation function. The loss design module designs a loss function based on actual responses and expected responses of the circuit. The integrated circuit control model establishment module establishes an integrated circuit control model based on the historical circuit sample set, in combination with the feature extraction module, the control decision mapping module and the loss design module. The integrated circuit control module controls real-time integrated circuits based on the established integrated circuit control model. The feature extraction module extracts features from the input X of the historical circuit sample set, including two parallel branches of local fine branches and global receptive branches; the local fine branch adopts a small convolution kernel with a size of , followed by channel attention and spatial reconstruction, represented as: ; wherein, is the local branch output; is the Channel-attention+Spatial-reconstruction module; is the convolution operation with a convolution kernel of ; the global receptive branch adopts a large convolution kernel with a size of , combined with a learnable channel weight, represented as: ; wherein, is the global branch output; is the learnable channel weight vector; is the convolution operation with a convolution kernel of ; prior feature fusion is represented as: ; wherein, is the prior fusion output; the global branch strengthens the abnormal response channel, represented as: ; wherein, is the feature after global branch strengthening; is the global max pooling; is the concatenated k times CIBlock; CIBlock consists of pointwise convolution→PReLU→Dropout→BatchNorm; the local branch promotes channel interaction, represented as: ; wherein, is the fusion feature after local branch processing; the fusion weight is output as the fusion ; The loss design module takes the sensor signal waveform actually exhibited by the circuit after the actual control command is applied as the actual response for the i th sample, and sets the error between the actual response and the expected response as , and constructs the adjustable smooth loss L, using the formula: ; the overall gradient is expressed as: ; wherein N is the total number of samples; k is a sensitivity parameter; and r is the overall error.

2. The integrated circuit control system based on artificial intelligence according to claim 1, characterized in that: The control decision mapping module will The flattened vector x is input into the light control network, denoted as: ; ; where, is the flattening operation; is the function mapping of the light control network; u is the multi-channel control instruction, each component corresponds to a physical control quantity; for the activation function mapping to the actual control instruction, denoted as: ; ; where, is the actual control instruction, mapped to the actual control command range; is the activation function mapping; , a and control the shape and characteristics of the activation function; and are the maximum and minimum values of the control command range, respectively.

3. The integrated circuit control system based on artificial intelligence according to claim 2, characterized in that: Further, the integrated circuit control model establishment module divides the historical circuit sample set into a test set and a training set, constructs an integrated circuit control model based on the feature extraction module, the control decision mapping module and the loss design module, sets an error threshold, and when the integrated circuit control model loss converges or reaches the maximum training number, the integrated circuit control model training is completed.

4. The integrated circuit control system based on artificial intelligence according to claim 3, characterized in that: The data acquisition module acquires internal historical signals of the integrated circuit and determines an expected response, which is used as a data label.

5. The artificial intelligence-based integrated circuit control system according to claim 4, characterized by: The integrated circuit control module acquires internal signals of the integrated circuit in real time based on the established integrated circuit control model, inputs the signals into the integrated circuit control model, and controls the integrated circuit based on the actual control command obtained from the model.

Images