ANN chip current fault prediction algorithm based on near field scanning

Abstract

The invention relates to the technical field of current fault prediction, in particular to an ANN chip current fault prediction algorithm based on near-field scanning, and the ANN chip current fault prediction algorithm is realized based on a near-field scanning system. The ANN chip current fault prediction algorithm comprises the following steps: carrying out white-box testing and analysis on a model and a current estimation principle, constructing and training an artificial neural network model, and carrying out fault detection application. The current distribution of the power supply network is estimated by solving the quadratic programming problem, the current flowing direction, the wire current amplitude, the via hole current amplitude and the difference of the estimated current distribution with defects are obtained respectively, and the current distribution of the power supply network is estimated by detecting the coaxial through hole defects and the significant decline position of the current amplitude at the open circuit position of the line. According to the invention, the open circuit problem existing in the coaxial through hole or the circuit can be clearly detected.

Description

Technical Field

[0001] The present invention relates to the technical field of current fault prediction, and in particular to an ANN chip current fault prediction algorithm based on near-field scanning. Background Art

[0002] In recent years, VLSI have generally adopted low-power supply voltage, so the impact of power supply noise has become increasingly significant.

[0003] In digital circuits, inrush current synchronized with the clock signal often causes power supply noise in the on-chip system. This current causes voltage drop due to parasitic elements in the power supply network, resulting in changes in delay and even functional failures due to setup / hold violations. The measurement method of the prior art mainly analyzes the mechanism of the damaged area after the integrated circuit is opened. This test method is irreversible, and the test time and economic cost are high. Therefore, there are currently three main problems with the current fault detection problem of integrated circuits: the first problem is that the detection is an invasive destructive detection problem; the second problem is that the detection cost is expensive; and the third problem is that the detection results are not universal. A method is needed to measure and evaluate the actual power supply network after manufacturing, so as to evaluate the failure analysis method of chip current operation in printed circuit boards.

[0004] In summary, in order to meet the above characteristics, the present invention proposes an ANN current fault prediction algorithm based on the near-field scanning technology, so as to efficiently detect the current fault problem caused by the coaxial through-hole or the short circuit of the connection of the chip. Summary of the invention

[0005] The purpose of the present invention is to provide an ANN chip current fault prediction algorithm based on near-field scanning to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] An ANN chip current fault prediction algorithm based on near-field scanning, the ANN (artificial neural network) chip current fault prediction algorithm is implemented based on a near-field scanning system, and the ANN chip current fault prediction algorithm includes the following steps:

[0008] S1. White box test and analysis model: Prepare a DUT with known structure, including open-cover and packaged samples, use a near-field scanning system to obtain electromagnetic interference cloud map information on the surface of the DUT, and estimate the device surface current through the field strength component;

[0009] S2. Current estimation principle: Based on the Biot-Savart law, the magnetic field generated by the current segment at the measuring point is calculated, the induced electromotive force is calculated by Faraday's law of electromagnetic induction, and then the magnetic field is calculated. The relationship between the magnetic field matrix and the current segment matrix is ​​established, and the current distribution is solved through the measurement data;

[0010] S3. Build and train the artificial neural network model: construct the input layer, hidden layer, and output layer, fully connect the neurons in adjacent layers, use the training set to optimize the parameters, the validation set to avoid overfitting, and the test set to test the preliminary results, and use the ANN to predict the field intensity components and estimate the current distribution;

[0011] S4. Perform fault detection application: inject test current into the power supply network of the object under test to induce magnetic field interference emission, apply current estimation method to detect the current distribution of the power supply network, and locate coaxial through-hole defects or open circuit problems by detecting the significant drop position of current amplitude.

[0012] Preferably, the near-field scanning system comprises a detection structure, a transmission structure and an output module;

[0013] The detection structure is used to sense the magnetic field passing vertically through the loop;

[0014] The transmission structure and the detection structure are connected by an oblique inner buckle to eliminate the resonance signal at the connection;

[0015] The output module is used to transmit the radio frequency signal to the SMA output terminal for measurement by a spectrum analyzer or a network analyzer.

[0016] Preferably, the implementation process of the white box test and analysis model in step S1 is as follows:

[0017] S1.1. First, when the physical routing of the object under test is known (i.e., white box condition), prepare the uncapped and packaged samples to be tested and analyzed with the same fault respectively;

[0018] S1.2. Then, a near-field scanning system is used to obtain electromagnetic interference cloud map information on the surface of the object under test;

[0019] S1.3. Finally, the field intensity component on the surface of the object under test is obtained through the near-field scanning system, thereby estimating the current on the surface of the device;

[0020] S1.4. Use a near-field probe to measure the EMI interference source emission of the circuit under test;

[0021] S1.5. Characterize the EMI two-dimensional field strength diagram of the object under test through near-field scanning technology;

[0022] S1.6. Use the graph of the field intensity components captured by the near-field probe to estimate the distribution of current on the object under test.

[0023] Preferably, the process of constructing and training the artificial neural network model in step S3 is as follows:

[0024] S3.1. Construct an artificial neural network including an input layer, one or more hidden layers, and an output layer, ensuring that neurons in adjacent layers are usually fully connected;

[0025] S3.2. The dipole moment is implicitly represented by the weights, biases, and activation functions of the artificial neural network;

[0026] S3.3. Use artificial neural network mapping to replace the matrix-vector multiplication operation between the free-space Green's function and the dipole moment in the traditional dipole method;

[0027] S3.4. After adding hidden layers to the neural network, the mapping ability of the neural network is used to process indirect waves;

[0028] S3.5. Perform artificial neural network training: the training data is randomly divided into three groups, 70% of the rows are selected as the training set for optimizing the artificial neural network parameters, 15% of the rows are selected as the validation set to avoid overfitting, and the last 15% of the rows are used as the test set to test the preliminary results of the winner model;

[0029] S3.6. The weights and biases are updated by using the training set and the back-propagation learning algorithm, and the corresponding field strength components are predicted by the artificial neural network algorithm.

[0030] Preferably, the process of performing fault detection application in step S4 is as follows:

[0031] S4.1. Inject a test current between VDD (digital power pin) and GND (0V line) of the power supply network;

[0032] S4.2. Place a test current source between VDD and GND on the object under test, inject an appropriate test current to induce magnetic field interference emission, and estimate the surface current of the power network;

[0033] S4.3. The current estimation method in step S3.1 to step S3.2 is applied to fault detection in the power supply network;

[0034] S4.4. estimating the current distribution of the power supply network by solving a quadratic programming problem;

[0035] S4.5. Obtain the current flow direction, conductor current amplitude and via current amplitude respectively, and the difference in estimated current distribution of defects by detecting the location of significant drop in current amplitude at the coaxial through-hole defect and the open circuit of the line.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention predicts the electromagnetic field data of the surface of the object to be tested based on the near-field scanning technology combined with the ANN equivalent source reconstruction algorithm, and characterizes the equivalent current of each area through the Biot-Savart law, thereby evaluating the failure analysis method of the chip current operation in the printed circuit board. The current estimation method of the present invention is applied to the fault detection of the power supply network, and estimates the current distribution of the power supply network by solving the quadratic programming problem, and obtains the current flow direction, the conductor current amplitude and the via current amplitude respectively. The difference in the estimated current distribution of the defect exists, and the position where the current amplitude significantly decreases at the coaxial through-hole defect and the line open circuit occurs can be detected by the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is the main flow chart of the ANN chip current fault prediction algorithm model of the present invention;

[0038] Figure 2 The schematic diagram of the calculation of the Biot-Savart law of the present invention;

[0039] Figure 3 It is a schematic diagram of the surface current estimation of the object to be measured based on the near-field scanning technology of the present invention;

[0040] Figure 4 An artificial neural network structure for the equivalent source reconstruction problem of the present invention;

[0041] Figure 5 It is a schematic diagram of estimating the power supply network of the object under test of the present invention; DETAILED DESCRIPTION

[0042] 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.

[0043] Example 1

[0044] See also Figure 1 , Figure 2 , Figure 3 as well as Figure 4 The present invention proposes an ANN chip current fault prediction algorithm based on near-field scanning. The ANN (artificial neural network) chip current fault prediction algorithm is implemented based on a near-field scanning system. The ANN chip current fault prediction algorithm includes the following steps:

[0045] S1. White box test and analysis model: Prepare a DUT with known structure, including open-cover and packaged samples, use a near-field scanning system to obtain electromagnetic interference cloud map information on the surface of the DUT, and estimate the device surface current through the field strength component;

[0046] Specifically, the process is as follows:

[0047] S1.1. First, when the physical routing of the object under test is known (i.e., white box condition), prepare the uncapped and packaged samples to be tested and analyzed with the same fault respectively;

[0048] S1.2. Then, a near-field scanning system is used to obtain electromagnetic interference cloud map information on the surface of the object under test;

[0049] S1.3. Finally, the field intensity component on the surface of the object under test is obtained through the near-field scanning system, thereby estimating the current on the surface of the device;

[0050] S1.4. Use a near-field probe to measure the EMI interference source emission of the circuit under test;

[0051] S1.5. Characterize the EMI two-dimensional field strength diagram of the object under test through near-field scanning technology;

[0052] S1.6. Using the graph of the field intensity components captured by the near-field probe, the distribution of the current on the object under test can be estimated;

[0053] S2. Current estimation principle: Based on the Biot-Savart law, the magnetic field generated by the current segment at the measuring point is calculated, the induced electromotive force is calculated by Faraday's law of electromagnetic induction, and then the magnetic field is calculated. The relationship between the magnetic field matrix and the current segment matrix is ​​established, and the current distribution is solved through the measurement data;

[0054] S3. Build and train the artificial neural network model: construct the input layer, hidden layer, and output layer, fully connect the neurons in adjacent layers, use the training set to optimize the parameters, the validation set to avoid overfitting, and the test set to test the preliminary results, and use the ANN to predict the field intensity components and estimate the current distribution;

[0055] Specifically, the process is as follows:

[0056] S3.1. Construct an artificial neural network including an input layer, one or more hidden layers, and an output layer, ensuring that neurons in adjacent layers are usually fully connected;

[0057] S3.2. The dipole moment is implicitly represented by the weights, biases, and activation functions of the artificial neural network;

[0058] S3.3. Use artificial neural network mapping to replace the matrix-vector multiplication operation between the free-space Green's function and the dipole moment in the traditional dipole method;

[0059] S3.4. After adding hidden layers to the neural network, the mapping ability of the neural network is used to process indirect waves;

[0060] S3.5. Perform artificial neural network training: the training data is randomly divided into three groups, 70% of the rows are selected as the training set for optimizing the artificial neural network parameters, 15% of the rows are selected as the validation set to avoid overfitting, and the last 15% of the rows are used as the test set to test the preliminary results of the winner model;

[0061] S3.6. Update the weights and biases using the training set and back propagation learning algorithm, and predict the corresponding field strength components using the artificial neural network algorithm

[0062] S4. Fault detection application: inject test current into the power supply network of the object under test to induce magnetic field interference emission, apply current estimation method to detect the current distribution of the power supply network, and locate the coaxial through-hole defect or open circuit problem by detecting the significant drop position of the current amplitude;

[0063] Specifically, the process is as follows:

[0064] S4.1. Inject a test current between VDD (digital power pin) and GND (0V line) of the power supply network;

[0065] S4.2. Place a test current source between VDD and GND on the object under test, inject an appropriate test current to induce magnetic field interference emission, and estimate the surface current of the power network;

[0066] S4.3. The current estimation method in step S3.1 to step S3.2 is applied to fault detection in the power supply network;

[0067] S4.4. estimating the current distribution of the power supply network by solving a quadratic programming problem;

[0068] S4.5. Obtain the current flow direction, conductor current amplitude and via current amplitude respectively, and the difference in estimated current distribution of defects by detecting the location of significant drop in current amplitude at the coaxial through-hole defect and the open circuit of the line.

[0069] Example 2

[0070] In actual application, the design method and application steps of the high-performance magnetic field near-field probe based on the ANN chip current fault prediction algorithm based on near-field scanning are as follows:

[0071] See also Figure 1The application of the artificial intelligence algorithm of the present invention in the white box reconstruction of EMI interference sources is to predict the field strength of the object under test through the artificial intelligence algorithm and estimate the actual current of the object under test by solving the inverse problem of the field strength component. It can be applied to the fault analysis and location of the object under test. Its main flow chart is as follows: Figure 1 As shown, firstly, the electromagnetic field components of the surface of the object under test with a known structure at a certain height are obtained through a high-performance near-field scanning system, and then the equivalent dipole array is obtained and its EMI component information is reconstructed using an artificial neural network model. Finally, the current on the surface of the white-box EMI interference source is estimated using the reconstructed EMI component information and used to analyze the disconnection positions in the coaxial through holes and connections of the object under test.

[0072] The transmission structure and the detection structure are connected through an oblique inner buckle structure. The detection loop part of the probe senses and detects the magnetic field vertically passing through the loop through a structure composed of a designed second-layer coil, a third-layer coil and a through hole. The change of the magnetic field vertically passing through the detection loop is deduced according to the expression of the magnetic flux to obtain the corresponding frequency response. During the transmission process, six coaxial through-hole arrays are used at the connection between the first layer and the second layer to eliminate the resonant signal at the connection. Finally, the RF signal is transmitted to the SMA output end, so that the data is measured by a spectrum analyzer or a network analyzer. Specifically, the following steps are included;

[0073] 1) White box testing and analysis model of EMI interference sources of the DUT based on near-field scanning technology:

[0074] First, when the physical routing of the object under test is known (i.e., white box condition), prepare the uncapped and packaged samples to be tested and analyzed with the same fault respectively;

[0075] Then, a near-field scanning system is used to obtain electromagnetic interference cloud map information on the surface of the object under test;

[0076] Finally, the field strength component on the surface of the object under test is obtained through the near-field scanning system, so as to estimate the current on the surface of the device. The specific estimation principle is as follows:

[0077] like Figure 2 As shown in the figure, the magnetic field measured at point P is described as the vector sum of the magnetic fields induced by all current paths on the object under test. When estimating the current, the power supply network is divided into multiple current segments, each of which is modeled as a zero-width line with uniform current. The magnetic field B generated by the current segment l measured at point P is l,p According to the Biot-Savart law, it is expressed as formula (1):

[0078]

[0079] In the formula, i l is the current on line segment l, μ is the magnetic permeability of air, bl,p is the unit vector of the direction of the magnetic field generated by the current segment l at point p, r l,p is the vertical distance from current segment l to point p, θ 1(l,p) and θ 2(l,p) is the angle formed by the current line segment l and the vector from the starting (ending) point of the current line segment to point p. The normal unit vector of the near-field probe at the measuring point is defined as d. The component of the magnetic field in the measuring direction is expressed as formula (2):

[0080]

[0081] Then the total magnetic field B generated by L current segments at point p is d,p Formula (3):

[0082]

[0083] In the equation, Y d,l,p It is expressed as formula (4):

[0084]

[0085] The magnetic flux φ obtained by the near-field scanning system at the measuring point p p (t) is expressed as formula (5):

[0086] φ p (t) = N∫B d (t)dS (5);

[0087] In the above formula, N is the number of turns of the coil, Bd(t) is the component of the magnetic field passing through the coil in the measuring direction, and S is the total area of ​​the coil. In this case, the measuring point means the center of the coil. The induced electromotive force v of the near-field probe at the measuring point p is p (t) is expressed as equations (6) and (7) through Faraday's law of electromagnetic induction:

[0088]

[0089] in:

[0090]

[0091] Because the induced electromotive force V can be measured p The amplitude of the magnetic field B is calculated d,p , so the relationship between the magnetic field B matrix and the current segment I matrix can be expressed as formula (8):

[0092]

[0093] Where P is the total number of measurement points. The current estimation flow chart of the physical wiring of the object under test using the above analysis model is as follows: Figure 3 As shown, the EMI interference source emission of the circuit under test is measured by using a near-field probe, and the EMI two-dimensional field strength pattern of the object under test is characterized by the near-field scanning technology. The pattern of the field strength component captured by the near-field probe is used to estimate the distribution of the current on the object under test;

[0094] 2) EMI interference source white box reconstruction model based on artificial intelligence algorithm:

[0095] See also Figure 4 , an artificial neural network EMI interference source prediction model is established for the equivalent source reconstruction problem. The artificial neural network includes an input layer, one or more hidden layers and an output layer. The neurons in adjacent layers are usually fully connected. w, c and f are weights, biases and activation functions respectively. The dipole moment is implicitly represented by the weights, biases and activation functions of the artificial neural network. The artificial neural network mapping is used to replace the matrix-vector multiplication operation between the free space Green's function and the dipole moment in the traditional dipole method. After adding hidden layers to the neural network, the mapping ability of the neural network can be used to process indirect waves. For the training of the artificial neural network, the entire training data is randomly divided into three groups. 70% of the rows are selected as the training set for optimizing the parameters of the artificial neural network, 15% of the rows are selected as the validation set to avoid overfitting, and the last 15% of the rows are used as the test set to test the preliminary results of the winner model. The weight w, bias c and activation function f are three unknown parameters in the artificial neural network. The weights and biases are updated by using the training set and the back propagation learning algorithm (gradient descent based on a suitable cost function). The corresponding field strength components can be predicted by the artificial neural network algorithm.

[0096] 3) Fault detection applications based on current estimation, in order to control the magnetic field emission of the power supply network for testing purposes, such as Figure 5 As shown, a test current with a frequency of f is injected between VDD and GND of the power supply network. S3.2. The test current source is placed between VDD and GND on the object under test, and an appropriate test current is injected to induce magnetic field interference emission, and the surface current of the power supply network is estimated;

[0097] In order to accurately detect through-hole defects or open circuit problems, the above current estimation method is applied to fault detection of the power supply network. The current distribution of the power supply network is estimated by solving the quadratic programming problem, and the current flow direction, conductor current amplitude and via current amplitude are obtained respectively. The difference in the estimated current distribution of defects is detected by detecting the location of a significant decrease in the current amplitude at the coaxial through-hole defect and the open circuit of the line. The open circuit problem in the coaxial through-hole or circuit can be clearly detected through the estimation method of the above analysis model.

[0098] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An ANN chip current fault prediction algorithm based on near-field scanning, the ANN chip current fault prediction algorithm is implemented based on a near-field scanning system, characterized in that: The ANN chip current fault prediction algorithm includes the following steps: S1. White box test and analysis model: Prepare a DUT with known structure, including open-cover and packaged samples, use a near-field scanning system to obtain electromagnetic interference cloud map information on the surface of the DUT, and estimate the device surface current through the field strength component; S2. Current estimation principle: Based on the Biot-Savart law, the magnetic field generated by the current segment at the measuring point is calculated, the induced electromotive force is calculated by Faraday's law of electromagnetic induction, and then the magnetic field is calculated. The relationship between the magnetic field matrix and the current segment matrix is ​​established, and the current distribution is solved through the measurement data; S3. Build and train the artificial neural network model: construct the input layer, hidden layer, and output layer, fully connect the neurons in adjacent layers, use the training set to optimize the parameters, the validation set to avoid overfitting, and the test set to test the preliminary results, and use the ANN to predict the field intensity components and estimate the current distribution; S4. Perform fault detection application: inject test current into the power supply network of the object under test to induce magnetic field interference emission, apply current estimation method to detect the current distribution of the power supply network, and locate coaxial through-hole defects or open circuit problems by detecting the significant drop position of current amplitude.

2. The ANN chip current fault prediction algorithm based on near-field scanning according to claim 1 is characterized in that: The near-field scanning system includes a detection structure, a transmission structure and an output module; The detection structure is used to sense the magnetic field passing vertically through the loop; The transmission structure and the detection structure are connected by an oblique inner buckle to eliminate the resonance signal at the connection; The output module is used to transmit the radio frequency signal to the SMA output terminal for measurement by a spectrum analyzer or a network analyzer.

3. The ANN chip current fault prediction algorithm based on near-field scanning according to claim 2 is characterized in that: The implementation process of step S1 white box testing and analysis model is as follows: S1.

1. First, when the physical routing of the object under test is known (i.e., white box condition), prepare the uncapped and packaged samples to be tested and analyzed with the same fault respectively; S1.

2. Then, a near-field scanning system is used to obtain electromagnetic interference cloud map information on the surface of the object under test; S1.

3. Finally, the field intensity component on the surface of the object under test is obtained through the near-field scanning system, thereby estimating the current on the surface of the device; S1.

4. Use a near-field probe to measure the EMI interference source emission of the circuit under test; S1.

5. Characterize the EMI two-dimensional field strength diagram of the object under test through near-field scanning technology; S1.

6. Use the graph of the field intensity components captured by the near-field probe to estimate the distribution of current on the object under test.

4. The ANN chip current fault prediction algorithm based on near-field scanning according to claim 3 is characterized in that ,The process of step S3 to construct and train the artificial neural network model is as follows: S3.

1. Construct an artificial neural network including an input layer, one or more hidden layers, and an output layer, ensuring that neurons in adjacent layers are usually fully connected; S3.

2. The dipole moment is implicitly represented by the weights, biases, and activation functions of the artificial neural network; S3.

3. Use artificial neural network mapping to replace the matrix-vector multiplication operation between the free-space Green's function and the dipole moment in the traditional dipole method; S3.

4. After adding hidden layers to the neural network, the mapping ability of the neural network is used to process indirect waves; S3.

5. Perform artificial neural network training: the training data is randomly divided into three groups, 70% of the rows are selected as the training set for optimizing the artificial neural network parameters, 15% of the rows are selected as the validation set to avoid overfitting, and the last 15% of the rows are used as the test set to test the preliminary results of the winner model; S3.

6. The weights and biases are updated by using the training set and the back-propagation learning algorithm, and the corresponding field strength components are predicted by the artificial neural network algorithm.

5. The ANN chip current fault prediction algorithm based on near-field scanning according to claim 4 is characterized in that: The process of performing fault detection application in step S4 is as follows: S4.

1. Inject a test current between VDD (digital power pin) and GND (0V line) of the power supply network; S4.

2. Place a test current source between VDD and GND on the object under test, inject an appropriate test current to induce magnetic field interference emission, and estimate the surface current of the power network; S4.

3. The current estimation method in step S3.1 to step S3.2 is applied to fault detection in the power supply network; S4.

4. estimating the current distribution of the power supply network by solving a quadratic programming problem; S4.

5. Obtain the current flow direction, conductor current amplitude and via current amplitude respectively, and the difference in estimated current distribution of defects by detecting the location of significant drop in current amplitude at the coaxial through-hole defect and the open circuit of the line.

Images