Circuit system health status prediction method based on low-frequency noise and deep learning

Abstract

The present invention relates to a circuit system health prediction method based on low-frequency noise and deep learning, which includes the following steps: step 1: assigning stability importance to the circuit system; step 2: extracting low-frequency noise and obtaining steady-state distribution characteristic parameters of the low-frequency noise of the components in the time domain analysis; step 3: processing low-frequency noise signal data and establishing a neural network model; step 4: completing the training of the neural network model and obtaining the prediction result of the circuit life. The present invention realizes circuit health analysis by establishing a deep learning prediction model based on the analysis of the low-frequency noise of the components. It can adapt to circuit systems with different failure criteria and is suitable for accurate diagnosis in the engineering field. Compared with the traditional aging test method based on modeling of a large number of test results, it has obvious advantages; it does not require overstressing during the test process, realizes non-destructive testing of the circuit, and realizes functional prediction of the circuit system timing based on health diagnosis.

Description

Technical Field

[0001] The present application relates to the technical field of electronic component fault monitoring and early warning, and specifically to a method for predicting the health status of a circuit system based on low-frequency noise and deep learning. Background Art

[0002] Electronic components play a crucial role in modern production and life. In the circuit systems of large, complex equipment, failure of key components can lead to serious malfunctions of the entire system, or even to serious accidents, posing unpredictable threats to the overall stability and safety of the system. Therefore, real-time monitoring and analysis of key components, and timely prediction of potential failure risks before they occur, are essential to ensuring the stability of equipment systems.

[0003] Currently, the most widely used methods for diagnosing component failures in circuit systems are qualitative methods such as manual observation and parameter testing. These methods rely on human sensory perception and parameter value determination, providing only a qualitative assessment of the component's operating status. In most cases, they only provide feedback after a failure has occurred, failing to provide pre-emptive warnings. To date, equipment failures and safety incidents caused by component failures in circuit systems remain a frequent occurrence in engineering projects, and existing fault diagnosis and prediction technologies still lack an ideal solution for addressing component failure warnings.

[0004] Therefore, it is essential to research and design a monitoring solution that can predict component failures in advance. With the deepening of low-frequency noise theory research, new ideas and methods are constantly being discovered and refined. The parameters of low-frequency electrical noise in semiconductor components can, to a certain extent, reflect damage to the component's internal materials and structures, and changes in these parameters often precede changes in the electrical parameters that cause component failure.

[0005] By analyzing the parameter variation trend of this noise, it is possible to monitor the components in working state in real time and provide early warning of faults, thereby effectively ensuring the testability and stability of the equipment and its circuit system. Although there are devices that can extract low-frequency noise, such devices are generally large and impractical, and there is currently no method to predict and screen the failure of components and circuit systems based on the low-frequency noise of components. Summary of the Invention

[0006] In order to overcome the shortcomings of the existing technology, the present invention establishes a deep learning prediction model through the extraction and analysis of the low-frequency noise of components to realize circuit health status analysis. It can adapt to circuit systems with various failure criteria and is suitable for accurate diagnosis in multiple engineering fields. Compared with the traditional aging test method based on modeling of a large number of test results, it has obvious advantages; the present invention improves the feasibility and pertinence of the test, does not require overstress during the test process, and realizes non-destructive testing of the circuit.

[0007] To achieve the above objectives, the solution adopted by the present invention is:

[0008] A method for predicting the health status of a circuit system based on low-frequency noise and deep learning, comprising the following steps:

[0009] Step 1: Assign stability importance to the circuit system;

[0010] Calculate the importance of n components in the circuit system, the importance of the i-th component is r i , the results of n components are normalized and sorted, and the method for obtaining the importance of components is as follows:

[0011]

[0012] Where: r i Indicates the importance of the i-th component; Δoutput i Indicates the system output change caused by the output change of the i-th component; i is the component number; n is the total number of components;

[0013] Step 2: Extract low-frequency noise and obtain the steady-state distribution characteristic parameters of the component low-frequency noise time domain analysis noise;

[0014] The steady-state distribution characteristic parameters include key frequency domain distribution characteristic parameters and key time domain distribution characteristic parameters;

[0015] Step 21: Use the optimized power spectrum to obtain key frequency domain distribution characteristic parameters. The optimized power spectrum is processed using a bilinear function fitting process. The frequency domain description model of the low-frequency noise is as follows:

[0016]

[0017] Where: log 10 Indicates the logarithm with base 10; S v (f) represents the key frequency domain distribution curve of the low frequency band; f represents the signal frequency value; f β represents the corner frequency of the bilinear function of the noise signal spectrum; k1 and k2 represent the first and second linear slopes respectively; D1 and D2 represent the first and second intercepts respectively;

[0018] Step 22: Obtain key time-domain distribution characteristic parameters using an optimized three-parameter distribution function. The optimized three-parameter distribution function description model is as follows:

[0019]

[0020] Where: p(x) represents the key time domain distribution characteristic function; x represents the voltage value of the low-frequency noise signal data; γ represents the kurtosis value of the noise signal in the time domain; σ represents the standard deviation value of the noise signal in the time domain; g′(θ) represents the derivative of the probability density coverage function; g(θ) represents the probability density coverage function; θ represents the probability density coverage rate; μ represents the mean value of the noise signal in the time domain; m represents the probability density scaling factor;

[0021] Step 3: Process low-frequency noise signal data and establish a neural network model;

[0022] Obtain the key frequency domain energy distribution concentration parameters and probability distribution characteristic parameters of the low-frequency noise signal data; establish a nonlinear relationship between the steady-state distribution characteristic parameters obtained in step 2 and the system output change caused by the component output change obtained in step 1 through the neural network model, as shown below:

[0023] Δoutput i =noise(f β ,k,μ,σ,γ);

[0024] Where: noise represents the nonlinear function relationship trained by the neural network; k represents the slope of the noise signal spectrum on the logarithmic axis;

[0025] Step 4: Complete the training of the neural network model and obtain the prediction results of the circuit life;

[0026] Given a failure threshold based on the usage environment, determine the weighted results of system output changes caused by component output changes. Based on the weighted results, obtain the comprehensive impact of each module on the system. Complete the training of the neural network model in step 3 and obtain the circuit life prediction result. The specific acquisition method is as follows:

[0027]

[0028] Where: Δoutput represents the life prediction result of the neural network model;

[0029] Preferably, the extraction of low-frequency noise in step 2 is specifically as follows:

[0030] Sort the components according to the importance calculation results obtained in step 1, extract the low-frequency noise of the components with the top 90% importance, and collect the corresponding low-frequency noise signal data through the noise acquisition circuit.

[0031] Preferably, the signal acquisition device in step 2 is a multi-stage low-frequency signal amplification device, the acquisition frequency band of which should include 0.03Hz-500Hz, the signal of which should be a double-ended differential signal, and the first stage amplification should use two processes and similar junction field-effect transistors to amplify by 6-20 times, so as to reduce the interference caused by the noise of the acquisition equipment. The second stage should use multiple amplifiers in parallel for amplification, which is 8-way parallel, and the amplification factor should be 100-1000 times. Finally, the output stage should consider the impedance matching of the sampling circuit and set a suitable amplification factor to obtain a mV-level low-frequency noise signal.

[0032] Preferably, the key frequency domain distribution characteristic parameters and key time domain distribution characteristic parameters of the low-frequency noise signal data obtained in step 3 are optimization results obtained by focusing on analyzing the low-frequency part of the noise distribution model, specifically:

[0033] The full frequency domain distribution function is used to process the low-frequency noise signal data as shown below:

[0034]

[0035] Where: S all (f) represents the full frequency domain distribution curve; f i Indicates the signal frequency value; f0 indicates the characteristic frequency of the noise signal; f γ Indicates; A represents the basic energy of the noise signal; B represents the energy amplitude of the low-frequency band of the noise signal; C i Indicates that different frequency points correspond to different energy amplitudes;

[0036] In order to accurately extract effective features, we mainly analyze low-frequency signals and establish an optimized description model for low-frequency signals, as shown below:

[0037]

[0038] Where: f β Represents the second corner frequency of the bilinear function of the noise signal spectrum.

[0039] Preferably, the loss function of the neural network model in step 3 in the time dimension is as follows:

[0040]

[0041] Where: F loss Represents the loss function calculation result of the neural network model in the time dimension; BCELoss represents the multi-state classification loss calculation function; x pred Represents the activation function value of the neuron at the previous moment; x lobel Represents the activation function value of the neuron at the current moment.

[0042] Preferably, the key time domain distribution characteristic parameters of the low-frequency noise signal data obtained in step 3 are optimization results based on kurtosis information and normal distribution, specifically:

[0043]

[0044] Where: F cdf represents the noise time series cumulative distribution function;

[0045] The probability density distribution function of the optimized noise signal is obtained by taking the derivative of the cumulative distribution. The acquisition method is as follows:

[0046] p(x)=F c ' df (x);

[0047] Where: F c ' df Represents the derivative of the noise time series cumulative distribution function.

[0048] Preferably, the noise acquisition circuit is a circuit output port adapter with a shielding device, which is used to connect the circuit to be tested and the noise measurement circuit. Under the action of the shielding device, other electromagnetic noise signals in the environment can be prevented from being introduced during the signal acquisition process.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] (1) The present invention establishes a deep learning prediction model by extracting and analyzing the low-frequency noise of components to realize circuit health status analysis. It can adapt to circuit systems with various failure criteria and is suitable for accurate diagnosis in multiple engineering fields. Compared with the traditional aging test method based on modeling of a large number of test results, it has obvious advantages.

[0051] (2) The present invention does not require a large number of repeated tests, thereby improving the feasibility and pertinence of the test. The low-frequency noise extracted from a component contains information about its internal lattice damage, pin damage, and other damage. Compared with traditional screening tests and repeated tests, which can damage the device and circuit during testing, the testing method of the present invention extracts the low-frequency noise at both ends of the component without applying overstress, thus achieving non-destructive testing.

[0052] (3) The present invention selects a low-frequency noise signal with degradation characteristics and combines it with the importance allocation method to truly realize the performance mapping from the circuit component level to the circuit system level. At the same time, combined with the neural network data analysis method, it realizes the time series deduction of the low-frequency noise signal, so that this method has the function of life prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1This is a control block diagram of a circuit system health status prediction method based on low-frequency noise and deep learning according to an embodiment of the present invention;

[0054] Figure 2 This is a flow chart of a prediction method according to an embodiment of the present invention;

[0055] Figure 3 This is a functional module decomposition diagram of the analysis process of an embodiment of the present invention;

[0056] Figure 4a-4e This is a specific circuit diagram of the data acquisition device according to an embodiment of the present invention;

[0057] Figure 5 This is a diagram showing a voltage stabilizing circuit system according to an embodiment of the present invention;

[0058] Figure 6 This is a comparison diagram of the improved time domain description model of the voltage stabilizing circuit according to the embodiment of the present invention;

[0059] Figure 7 This is a comparison diagram of the improved time domain description model of the voltage stabilizing circuit according to an embodiment of the present invention before and after improvement;

[0060] Figure 8 A comparison diagram of the improved frequency domain description model of the voltage stabilizing circuit according to an embodiment of the present invention;

[0061] Figure 9 A schematic diagram of the voltage regulation prediction accuracy of the voltage stabilizing circuit according to an embodiment of the present invention using the method;

[0062] Figure 10 This is a schematic diagram of the ripple prediction accuracy of the voltage stabilizing circuit according to an embodiment of the present invention using this method. DETAILED DESCRIPTION

[0063] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings.

[0064] The embodiment of the present invention takes a certain actual voltage-stabilizing circuit system as an example, and establishes a deep learning prediction model by extracting and analyzing the low-frequency noise of components to realize circuit health status analysis. It can adapt to circuit systems with various failure criteria and is suitable for accurate diagnosis in multiple engineering fields. This method does not require a large number of repeated tests, which improves the feasibility and pertinence of the test. The low-frequency noise extracted from a certain component contains information about its internal lattice damage, pin damage and other damage. Compared with traditional screening tests and repeated tests, which will damage the components and circuits during the test, the test method of the present invention realizes non-destructive testing. Figure 1 This is a control block diagram of a circuit system health status prediction method based on low-frequency noise and deep learning according to an embodiment of the present invention.

[0065] The embodiment of the present invention provides a circuit system health prediction method based on low-frequency noise and deep learning, such as Figure 2 The flowchart of the prediction method according to an embodiment of the present invention is shown. To demonstrate the applicability of the present invention, the method is applied to an example of a voltage stabilizing circuit system, specifically comprising the following steps:

[0066] S1: Assign stability importance to the voltage stabilization circuit system;

[0067] The importance of the seven components in the voltage stabilization circuit system is calculated, and the importance of the i-th component is r i , the results of the 7 components are normalized and sorted, and the method for obtaining the component importance is as follows:

[0068]

[0069] Where: r i Indicates the importance of the i-th component; Δoutput i It represents the change in system output caused by the change in the output of the i-th component; i represents the component number; n represents the total number of components.

[0070] According to the importance calculation results of the components obtained in S1, the top 90% of the components are the voltage regulator chip LM324, the voltage regulator tube BTZ52C3 and the transistor 8050DU. The low-frequency noise on them is extracted and the corresponding low-frequency noise signal data is collected through the noise acquisition circuit. The noise acquisition circuit is a circuit output port adapter with a shielding device, which is used to connect the circuit to be tested and the noise measurement circuit. Under the action of the shielding device, other electromagnetic noise signals in the environment can be prevented from being introduced during the signal acquisition process. Figure 3 Shown is a functional module decomposition diagram of the analysis process of an embodiment of the present invention.

[0071] S2: Extract low-frequency noise and obtain the steady-state distribution characteristic parameters of the component low-frequency noise in the time domain analysis;

[0072] The signal acquisition device is a multi-stage low-frequency signal amplification device. Its acquisition frequency band should include 0.03Hz-500Hz. Its signal should be a double-ended differential signal. The first stage of amplification should use two junction field-effect transistors with extremely similar processes to amplify 6-20 times, so as to reduce the interference caused by the noise of the acquisition equipment. The second stage should use multiple amplifiers in parallel, with 8 channels in parallel, and the amplification factor should be 100-1000 times. The final output stage should consider the impedance matching of the sampling circuit and set the appropriate amplification factor to obtain a mV-level low-frequency noise signal. Figure 4a-4e This is a schematic diagram of a specific circuit of an acquisition device according to an embodiment of the present invention. Specifically, Figure 4a This is the front-end acquisition circuit diagram, which uses differential signals and uses diodes with similar processes to reduce interference; Figure 4bis the selected analog-to-digital conversion circuit diagram; Figure 4c This is the first-stage amplifier circuit diagram, which uses junction field-effect transistors to reduce internal noise interference; Figure 4d This is the second stage amplifier circuit diagram, which contains two parallel connections; Figure 4e This is the power supply circuit diagram used.

[0073] Steady-state distribution characteristic parameters include key frequency domain distribution characteristic parameters and key time domain distribution characteristic parameters;

[0074] S21: The key frequency domain distribution characteristic parameters are obtained by optimizing the power spectrum. The optimized power spectrum is processed using a bilinear function fitting process. The frequency domain description model of the low-frequency noise is as follows:

[0075]

[0076] Where: log 10 Indicates the logarithm with base 10; S v (f) represents the key frequency domain distribution curve of the low frequency band; f represents the signal frequency value; f β represents the corner frequency of the bilinear function of the noise signal spectrum; k1 and k2 represent the first and second linear slopes respectively; D1 and D2 represent the first and second intercepts respectively.

[0077] S22: The key time domain distribution characteristic parameters are obtained by optimizing the three-parameter distribution function. The description model of the optimized three-parameter distribution function is as follows:

[0078]

[0079] Where: p(x) represents the key time domain distribution characteristic function; x represents the voltage value of the low-frequency noise signal data; γ represents the kurtosis value of the noise signal in the time domain; σ represents the standard deviation value of the noise signal in the time domain; g′(θ) represents the derivative of the probability density coverage function; g(θ) represents the probability density coverage function; θ represents the probability density coverage rate; μ represents the mean value of the noise signal in the time domain; m represents the probability density scaling factor.

[0080] S3: Process low-frequency noise signal data and build a neural network model;

[0081] The key frequency domain energy distribution concentration parameters and probability distribution characteristic parameters of low-frequency noise signal data are obtained based on the optimization results obtained by focusing on the low-frequency part of the noise distribution model. Specifically, they are:

[0082] The full frequency domain distribution function is used to process the low-frequency noise signal data as shown below:

[0083]

[0084] Where: S all(f) represents the full frequency domain distribution curve; f i Indicates the signal frequency value; f0 indicates the characteristic frequency of the noise signal; f γ Indicates; A represents the basic energy of the noise signal; B represents the energy amplitude of the low-frequency band of the noise signal; C i Indicates that different frequency points correspond to different energy amplitudes.

[0085] In order to accurately extract effective features, we mainly analyze low-frequency signals and establish an optimized description model for low-frequency signals, as shown below:

[0086]

[0087] Where: f β Represents the second corner frequency of the bilinear function of the noise signal spectrum.

[0088] The nonlinear relationship between the steady-state distribution characteristic parameters obtained in S2 and the system output changes caused by the component output changes obtained in S1 is established through the neural network model, as shown below:

[0089] Δoutput i =noise(f β ,k,μ,σ,γ);

[0090] Where: noise represents the nonlinear function relationship trained by the neural network; k represents the slope of the noise signal spectrum on the logarithmic axis.

[0091] The loss function of the neural network model in the time dimension is as follows:

[0092]

[0093] Where: F loss Represents the loss function calculation result of the neural network model in the time dimension; BCELoss represents the multi-state classification loss calculation function; x pred Represents the activation function value of the neuron at the previous moment; x lobel Represents the activation function value of the neuron at the current moment.

[0094] The key time domain distribution characteristic parameters of low-frequency noise signal data are obtained based on the optimization results of kurtosis information and normal distribution, specifically:

[0095]

[0096] Where: F cdf represents the noise time series cumulative distribution function.

[0097] The probability density distribution function of the optimized noise signal is obtained by taking the derivative of the cumulative distribution. The acquisition method is as follows:

[0098] p(x)=F c ' df (x);

[0099] Where: F c ' df Represents the derivative of the noise time series cumulative distribution function.

[0100] S4: Complete the training of the neural network model and obtain the prediction results of the circuit life;

[0101] Given a failure threshold based on the usage environment, determine the weighted results of system output changes caused by component output changes. Based on the weighted results, obtain the comprehensive impact of each module on the system. Complete the training of the neural network model in S3 and obtain the circuit life prediction results. The specific acquisition method is as follows:

[0102]

[0103] Where: Δoutput represents the life prediction result of the neural network model.

[0104] like Figure 5 The figure shows a voltage stabilizing circuit system according to an embodiment of the present invention. A low-frequency noise signal is extracted from the circuit, analyzed in the time and frequency domains, and then input into a neural network to obtain the health status and life prediction results of the circuit system. In the figure, H represents the health status and life prediction.

[0105] like Figure 6 The figure shows a comparison diagram of the improved time domain description model of the voltage stabilizing circuit according to the embodiment of the present invention; Figure 7 The figure shows a comparison diagram of the improved time domain description model of the voltage stabilizing circuit according to the embodiment of the present invention before and after the improvement, proving the accuracy of performance prediction and timing prediction.

[0106] like Figure 8 The figure shows a comparison diagram of the improved frequency domain description model of the voltage stabilizing circuit according to an embodiment of the present invention; the solid line is after improvement, and the dotted line is before improvement, which can be clearly seen by comparing the situation before and after improvement.

[0107] like Figure 9 The figure shows the voltage regulation prediction accuracy of the voltage stabilizing circuit according to the embodiment of the present invention using the method; Figure 10 The figure shows the ripple prediction accuracy of the voltage-stabilizing circuit using this method in an embodiment of the present invention. The analysis of the ripple and voltage prediction accuracy shows a high degree of overlap between the two trends and very similar absolute values, demonstrating the effectiveness of this method in practical applications.

[0108] In summary, the prediction results of the circuit system health status prediction method based on low-frequency noise and deep learning in this case have proven to be excellent.

[0109] (1) The embodiment of the present invention provides an extraction and analysis of the low-frequency noise of components and establishes a deep learning prediction model to realize the analysis of the circuit health status. It can adapt to circuit systems with various failure criteria and is suitable for accurate diagnosis in multiple engineering fields. Compared with the traditional aging test method based on a large number of test results, it has obvious advantages. The verification and analysis of actual cases prove that this method has a good use effect.

[0110] (2) The embodiments of the present invention do not require a large number of repeated tests, thereby improving the feasibility and pertinence of the test. The low-frequency noise extracted from a certain component contains information about its internal lattice damage, pin damage, and other damage. Compared with traditional screening tests and repeated tests, which may damage the device and circuit during the test, the test method of the present invention extracts the low-frequency noise at both ends of the component without applying overstress, thereby realizing non-destructive testing.

[0111] (3) The embodiment of the present invention selects a low-frequency noise signal with degradation characteristics, and combines it with the importance allocation method to truly realize the performance mapping from the circuit component level to the circuit system level. At the same time, combined with the neural network data analysis method, it realizes the time series deduction of the low-frequency noise signal, so that the method has the life prediction function. The analysis of the actual application effect proves that this method can meet the actual application needs.

[0112] The embodiments described above are merely descriptions of preferred implementations of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should fall within the scope of protection determined by the claims of the present invention.

Claims

1. A circuit system health prediction method based on low-frequency noise and deep learning, characterized in that: It includes the following steps: Step 1: Assign stability importance to the circuit system; Calculate the importance of n components in the circuit system, the importance of the i-th component is r i , the results of n components are normalized and sorted, and the method for obtaining the importance of components is as follows: Where: r i Indicates the importance of the i-th component; Δoutput i Indicates the system output change caused by the output change of the i-th component; i is the component number; n is the total number of components; Step 2: Extract low-frequency noise and obtain the steady-state distribution characteristic parameters of the component low-frequency noise time domain analysis noise; The steady-state distribution characteristic parameters include key frequency domain distribution characteristic parameters and key time domain distribution characteristic parameters; Step 21: Use the optimized power spectrum to obtain key frequency domain distribution characteristic parameters. The optimized power spectrum is processed using a bilinear function fitting process. The frequency domain description model of the low-frequency noise is as follows: Where: log 10 Indicates the logarithm with base 10; S v (f) represents the key frequency domain distribution curve of the low frequency band; f represents the signal frequency value; f β represents the corner frequency of the bilinear function of the noise signal spectrum; k1 and k2 represent the first and second linear slopes respectively; D1 and D2 represent the first and second intercepts respectively; Step 22: Obtain key time-domain distribution characteristic parameters using an optimized three-parameter distribution function. The optimized three-parameter distribution function description model is as follows: Where: p(x) represents the key time domain distribution characteristic function; x represents the voltage value of the low-frequency noise signal data; γ represents the kurtosis value of the noise signal in the time domain; σ represents the standard deviation value of the noise signal in the time domain; g′(θ) represents the derivative of the probability density coverage function; g(θ) represents the probability density coverage function; θ represents the probability density coverage rate; μ represents the mean value of the noise signal in the time domain; m represents the probability density scaling factor; Step 3: Process low-frequency noise signal data and establish a neural network model; Obtain the key frequency domain energy distribution concentration parameters and probability distribution characteristic parameters of the low-frequency noise signal data; establish a nonlinear relationship between the steady-state distribution characteristic parameters obtained in step 2 and the system output change caused by the component output change obtained in step 1 through the neural network model, as shown below: Output i =noise(f β ,k,m,s,c); Where: noise represents the nonlinear function relationship trained by the neural network; k represents the slope of the noise signal spectrum on the logarithmic axis; Step 4: Complete the training of the neural network model and obtain the prediction results of the circuit life; Given a failure threshold based on the usage environment, determine the weighted result of the system output change caused by the component output change. Based on the weighted result, obtain the impact of each module on the circuit system. Complete the training of the neural network model in step 3 and obtain the circuit life prediction result. The specific acquisition method is as follows: Where: Δoutput represents the life prediction result of the neural network model.

2. The circuit system health status prediction method based on low-frequency noise and deep learning according to claim 1 is characterized in that: The extraction of low-frequency noise in step 2 is specifically as follows: Sort the components according to the importance calculation results obtained in step 1, extract the low-frequency noise of the components with the top 90% importance, and collect the corresponding low-frequency noise signal data through the noise acquisition circuit.

3. The circuit system health status prediction method based on low-frequency noise and deep learning according to claim 1, characterized in that: The signal acquisition device described in step 2 is a multi-stage low-frequency signal amplification device, and its acquisition frequency band should include 0.03Hz-500Hz. Its signal should be a double-ended differential signal, and the first stage amplification should use two processes and similar junction field-effect transistors to amplify 6-20 times, so as to reduce the interference caused by the noise of the acquisition equipment. The second stage should use multiple amplifiers in parallel amplification, which is 8-way parallel, and the amplification factor should be 100-1000 times. Finally, the output stage should consider the impedance matching of the sampling circuit and set the appropriate amplification factor to obtain a mV-level low-frequency noise signal.

4. The circuit system health status prediction method based on low-frequency noise and deep learning according to claim 1, characterized in that: The key frequency domain distribution characteristic parameters and key time domain distribution characteristic parameters of the low-frequency noise signal data obtained in step 3 are optimization results obtained by focusing on analyzing the low-frequency part of the noise distribution model, specifically: The full frequency domain distribution function is used to process the low-frequency noise signal data as shown below: Where: S all (f) represents the full frequency domain distribution curve; f i Indicates the signal frequency value; f0 indicates the characteristic frequency of the noise signal; f γ Indicates; A represents the basic energy of the noise signal; B represents the energy amplitude of the low-frequency band of the noise signal; C i Indicates that different frequency points correspond to different energy amplitudes; In order to accurately extract effective features, the low-frequency signal is analyzed and an optimized description model of the low-frequency signal is established, as shown below: Where: f β Represents the second corner frequency of the bilinear function of the noise signal spectrum.

5. The circuit system health status prediction method based on low-frequency noise and deep learning according to claim 1, characterized in that: The loss function of the neural network model in step 3 in the time dimension is as follows: Where: F loss Represents the loss function calculation result of the neural network model in the time dimension; BCELoss represents the multi-state classification loss calculation function; x pred Represents the activation function value of the neuron at the previous moment; x lobel Represents the activation function value of the neuron at the current moment.

6. The circuit system health status prediction method based on low-frequency noise and deep learning according to claim 1, characterized in that: The key time domain distribution characteristic parameters of the low-frequency noise signal data obtained in step 3 are the optimization results based on the kurtosis information and the normal distribution, specifically: Where: F cdf represents the noise time series cumulative distribution function; The probability density distribution function of the optimized noise signal is obtained by taking the derivative of the cumulative distribution. The acquisition method is as follows: p(x)=F′ cdf (x); Where: F′ cdf Represents the derivative of the noise time series cumulative distribution function.

7. The circuit system health prediction method based on low-frequency noise and deep learning according to claim 2, characterized in that: The noise acquisition circuit is a circuit output port adapter with a shielding device, which is used to connect the circuit to be tested and the noise measurement circuit. Under the action of the shielding device, electromagnetic noise signals in the environment can be prevented from being introduced during the signal acquisition process.

Images