A low-error semiconductor low-frequency noise voltage spectral density measurement method

Abstract

The application relates to the technical field of semiconductor testing and noise characterization, and discloses a low-error semiconductor low-frequency noise voltage spectral density measurement method, which comprises the following steps: establishing a static working point of a measured semiconductor device through a bias supply unit, generating a pseudo-random pilot signal through a pilot injection unit and injecting the pilot signal into an input end as a path characterization reference, capturing a mixed signal after superposition of essential noise and the pilot through a double-channel amplification unit in parallel, synchronously acquiring digital sequences of each channel, performing cross-power spectrum estimation on the sequences and an original pilot reference sequence through a digital processing unit, extracting an instantaneous transfer function reflecting channel complex gain characteristics, performing real-time drift compensation according to the instantaneous transfer function and constructing an environmental coherence factor, stripping environmental correlation noise from compensated data by using the factor, and deducting a pilot self-energy item in combination with the instantaneous transfer function. The application eliminates the influence of system drift and environmental interference, and improves the accuracy of low-frequency noise characterization.

Description

Technical Field

[0001] This invention relates to the field of semiconductor testing and noise characterization technology, specifically to a method for measuring low-frequency noise voltage spectral density in semiconductors with low error. Background Technology

[0002] Semiconductor low-frequency noise is a key physical parameter characterizing internal defects, interface quality, and long-term reliability of devices. Its accurate measurement is of great significance for semiconductor process evaluation and circuit design. Among existing low-frequency noise measurement techniques, dual-channel cross-spectrum analysis is the mainstream method for suppressing the noise floor of amplifier systems. It uses two independent amplification channels to sample the same signal and cancels the uncorrelated thermal noise within each channel through cross-power spectrum calculation.

[0003] However, in actual testing, existing technologies face insurmountable accuracy bottlenecks. First, active amplification devices and bias networks in the measurement system are highly susceptible to environmental temperature fluctuations, causing dynamic drift in the gain and phase response of the measurement path over time. Traditional methods typically rely on offline calibration before and after testing, lacking real-time characterization of the path transfer function during the measurement process, and thus failing to correct system errors introduced by dynamic drift. Second, while the dual-channel cross-spectral method can suppress incoherent noise between channels, it lacks an effective identification and separation mechanism for coherent interference signals that simultaneously enter both acquisition channels via power coupling, spatial electromagnetic induction, etc., leading to environmentally relevant background noise mixing into the measurement results and causing severe distortion of the low-frequency spectral density. Furthermore, existing solutions often neglect the non-ideal nature of the frequency response characteristics of the measurement path and the impact of device impedance matching on signal transmission when processing the energy stripping between the injected reference signal and the intrinsic noise of the measured device, making it difficult to accurately restore the intrinsic noise voltage spectral density of the measured semiconductor device. In long-duration, extremely low-frequency measurement scenarios, the increased environmental complexity and the evolution of system characteristics jointly limit the physical accuracy of noise characterization. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a low-error semiconductor low-frequency noise voltage spectral density measurement method, which solves the problems of limited characterization accuracy and measurement result distortion caused by dynamic drift of system gain phase, environmental electromagnetic coherence interference, and impedance mismatch of measurement path in traditional semiconductor low-frequency noise measurement.

[0005] To achieve the above objectives, the present invention provides a low-error semiconductor low-frequency noise voltage spectral density measurement method. This method is based on a dual-channel cross-power spectrum measurement architecture. By introducing a continuous pseudo-random sequence as a path-dependent characterization benchmark and combining frequency domain complex calculation and feedback compensation algorithms, the method achieves real-time correction of system dynamic gain drift and environmental coherent interference.

[0006] The implementation steps of this measurement method include: providing a DC bias to the semiconductor device under test (DUT) through a bias supply unit to establish a static operating point; generating and injecting a pseudo-random pilot signal into the input terminal of the DUT through a pilot injection unit as a path-in-path characterization reference; capturing the mixed signal of the intrinsic noise generated by the DUT and the pseudo-random pilot signal superimposed by a dual-channel amplification unit in parallel, and simultaneously acquiring the digital sequences of the first and second measurement channels using a data acquisition unit; performing cross-power spectrum estimation on the acquired digital sequences and the original pseudo-random pilot reference sequence using a digital processing unit to extract the instantaneous transfer function reflecting the complex gain characteristics of the measurement path; performing real-time drift compensation on the digital sequences of the first and second measurement channels based on the instantaneous transfer function, and constructing an environmental coherence factor based on the instantaneous transfer function; removing environmentally relevant background noise from the compensated measurement data using the environmental coherence factor, and subtracting the pilot energy term corresponding to the pseudo-random pilot signal using the instantaneous transfer function; and performing normalization reconstruction on the stripped and subtracted data to obtain the intrinsic noise voltage spectral density equivalent to the output terminal of the DUT.

[0007] In the above measurement mechanism, the pseudo-random pilot signal is generated using a linear feedback shift register structure based on a primitive polynomial. The order of the linear feedback shift register is preset according to the measurement bandwidth, and the driving clock frequency is not less than a preset multiple of the maximum measurement frequency. When injecting this signal, it is coupled to the input terminal of the semiconductor device under test through a three-terminal symmetrical injection network, and the injection amplitude is adjusted using a programmable attenuator to keep the injected pilot power spectral density within a preset proportion range of the expected noise spectral density, thereby maintaining the linear operating state of the device bias point and amplification path.

[0008] For the acquired channel data, cross-power spectrum estimation employs a piecewise overlapping averaging method to window the digital sequence. The windowed frequency domain sequence is then multiplied point-by-point with the original pseudo-random pilot reference sequence's frequency domain complex conjugate and statistically averaged to obtain the cross-power spectral density function (CPSD) of each measurement channel relative to the original pseudo-random pilot reference sequence. Subsequently, the CPSD obtained from the cross-power spectral density estimation is divided by the auto-power spectral density of the original pseudo-random pilot reference sequence to calculate the instantaneous transfer function (ITF), which includes complex magnitude and complex phase. A moving average filter is then used to smooth the ITF by dynamically adjusting the window width according to the measurement frequency. This step enables accurate extraction of the physical path response characteristics that vary with temperature or time.

[0009] After obtaining the instantaneous transfer function, the system enters the real-time drift compensation stage. The ratio between the instantaneous transfer function at the current moment and the reference transfer function at the start of the measurement is calculated to construct the drift coefficient matrix. Then, inverse proportional weighting is performed on the digital sequences of the first and second measurement channels at the digital backend to compensate for the gain and phase errors introduced by the temperature drift of the active devices.

[0010] To address common-mode interference caused by space electromagnetic radiation or power supply coupling, the system constructs an environmental coherence factor. The cross-power spectral density of the digital sequences from the first and second measurement channels is divided by the complex product of their respective extracted instantaneous transfer functions. This process backtracks the mixed signal equivalently to the physical output node of the semiconductor device under test (DUT). The presence of external nonlinear coherent interference is determined based on the amplitude of the imaginary part of the backtracked signal. During the environmental noise floor removal phase, narrowband energy abrupt changes in the environmental coherence factor are identified, and a dynamic notch filter is used to eliminate power frequency interference. Simultaneously, multiple measurements are performed by controlling the bias supply unit to change the drain current of the DUT, and differential operations are used to eliminate constant coherent components that do not change with the current.

[0011] In the data reconstruction stage, it is necessary to accurately subtract the pilot energy term corresponding to the pseudo-random pilot signal. The system uses the instantaneous transfer function to calculate the equivalent correlation energy term of the pseudo-random pilot signal after dual-channel amplification. The equivalent correlation energy term is then subtracted from the cross-power spectral density of each channel's digital sequence using a complex vector operation, and the energy matching correction factor is iterated based on the residual energy in the pilot characteristic frequency band. Finally, normalized reconstruction is performed. An impedance matching correction coefficient is constructed using the output impedance of the semiconductor device under test and the input impedance of the dual-channel amplification unit, and voltage division compensation is performed. The spectral slope of the curve corresponding to the obtained intrinsic noise voltage spectral density is checked to ensure its validity within a preset physical range, thereby outputting physically accurate semiconductor ultra-low frequency noise characteristic parameters.

[0012] This invention provides a low-error semiconductor low-frequency noise voltage spectral density measurement method, which has the following beneficial effects: 1. This invention achieves real-time dynamic monitoring of system gain and phase fluctuations by injecting pseudo-random pilot signals and extracting the instantaneous transfer function of the measurement path. By using the calculated drift coefficient matrix to perform inverse proportional weighted compensation on the original sequence, the measurement error caused by temperature drift of active devices is effectively solved, improving the stability of the measurement system under long-term operating conditions.

[0013] 2. This invention utilizes the environmental coherence factor to trace the measured signal back to the physical output node of the semiconductor device under test, which can accurately identify and determine external nonlinear coherent interference. By combining variable current differential operation and dynamic notch filtering technology, it achieves the removal of constant environmental background noise and power frequency interference that does not change with current, thereby improving the signal-to-noise ratio of measurement results under complex electromagnetic environments.

[0014] 3. This invention combines pilot energy complex vector subtraction performed using the instantaneous transfer function with impedance matching correction to ensure that the reconstructed voltage spectral density accurately reproduces the intrinsic physical characteristics of the semiconductor device under test. By performing a spectral slope validity check on the spectral density curve, closed-loop quality control of the measurement data is achieved, eliminating the influence of non-physical energy components on the characterization results and enhancing the reliability of the measurement results. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the feedback logic for dynamic drift monitoring and real-time compensation in this invention; Figure 3 This is a logarithmic graph showing the voltage spectral density as a function of frequency according to the present invention. Detailed Implementation

[0016] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] See attached document Figure 1 This invention provides a low-error semiconductor low-frequency noise voltage spectral density measurement method, which is based on system implementation.

[0018] The system includes an electromagnetic shielding unit, a bias supply unit, a pilot injection unit, a dual-channel amplification unit, a data acquisition unit, and a digital processing unit.

[0019] The electromagnetic shielding unit includes a multi-layer metal shielding box and a device adapter housed inside the multi-layer metal shielding box. The multi-layer metal shielding box is used to block external electromagnetic signals from radiating interference to the internal test circuit. The device adapter is equipped with multiple test slots for fixing and connecting the semiconductor device under test.

[0020] The bias supply unit includes a low-noise linear regulated power supply and a DC bias circuit. The output of the low-noise linear regulated power supply is connected to the input of the DC bias circuit. The output of the DC bias circuit is connected to the pins of the device adapter. The DC bias circuit includes a resistor array and a filter capacitor, used to provide a preset DC bias voltage to the semiconductor device under test, enabling it to enter a specific operating state.

[0021] The pilot injection unit includes a pseudo-random sequence generator and a three-terminal symmetrical injection network. The output of the pseudo-random sequence generator is connected to the input electrode of the semiconductor device under test (DUT) through the three-terminal symmetrical injection network. The three-terminal symmetrical injection network includes a high-impedance coupling branch that couples the pseudo-random pilot signal to the node between the bias circuit and the DUT via a capacitor or resistor. The three-terminal symmetrical injection network maintains physical symmetry to ensure consistent signal attenuation characteristics during coupling.

[0022] The dual-channel amplification unit includes a first preamplifier and a second preamplifier. The inputs of the first and second preamplifiers are connected in parallel to the signal output of the semiconductor device under test. The first and second preamplifiers have independent power supply circuits and physical isolation shields to suppress coherent transduction between the two channels.

[0023] The data acquisition unit includes a multi-channel analog-to-digital converter (ADC) and a data buffer. The inputs of the multi-channel ADC are connected to the outputs of the first and second preamplifiers, respectively. The multi-channel ADC is used to convert the amplified analog noise signal into a digital sequence.

[0024] The digital processing unit is connected to the data acquisition unit. The digital processing unit includes a digital signal processor or a computer, used to receive digital sequences and perform cross-correlation operations, transfer function extraction, and noise floor stripping.

[0025] During system operation, the bias supply unit establishes a static operating point for the semiconductor device under test (DUT). The pilot injection unit superimposes a pseudo-random sequence signal onto the input of the DUT. The noise signal generated by the DUT and the superimposed pseudo-random sequence signal are amplified in parallel by a dual-channel amplification unit.

[0026] The amplified signal is synchronously sampled by the data acquisition unit and converted into a digital signal. The digital processing unit acquires the sampled digital signal and retrieves the original pseudo-random sequence reference signal pre-stored in the memory.

[0027] The digital processing unit calculates the cross-power spectrum between the output sequence of each channel and the original pseudo-random sequence reference signal, and analyzes the instantaneous transfer function of each amplification channel in real time, including the amplitude response and phase response.

[0028] The digital processing unit further performs dynamic gain compensation and phase alignment on the acquired dual-channel noise data using the extracted instantaneous transfer function. Utilizing the orthogonality of pseudo-random sequences in the frequency domain, the digital processing unit identifies and removes relevant interference components introduced by the bias supply unit or the external environment, ultimately reconstructing the intrinsic noise voltage spectral density of the semiconductor device under test.

[0029] This invention provides a method for measuring low-error semiconductor low-frequency noise voltage spectral density, wherein the generation and characterization of pseudo-random pilot signals specifically include the following steps and physical logic: The pseudo-random pilot signal is generated by the logic control module in the digital processing unit and converted into an analog electrical signal through a digital-to-analog converter circuit. In this embodiment, the logic control module uses a linear feedback shift register structure to generate a pseudo-random binary sequence. The linear feedback shift register performs logical operations according to a preset primitive polynomial, and the resulting sequence period length is... satisfy: ; in, In this embodiment, the order of the linear feedback shift register is... The value range is set to 17 to 31 to ensure sufficient non-repeatability of the pilot signal within the measurement time period. The frequency resolution of the pseudo-random pilot signal... Driven by clock frequency With sequence period length The decision is made jointly, and the specific relationship is as follows: ; To ensure that the pilot signal has a flat power spectrum characteristic within the tested frequency band (1Hz to 100kHz), the driving clock frequency is... The frequency is set to be no less than 5 times the maximum measurement frequency. The generated digital sequence, after being output by a high-precision digital-to-analog converter (DAC), exhibits a time-domain waveform as a bipolar rectangular pulse sequence.

[0030] Power spectral density of pseudo-random pilot signal in the frequency domain Follow the following envelope properties: ; in, The voltage amplitude of the pilot signal. For clock cycles, i.e. Because the measurement frequency band of 1Hz to 100kHz is much smaller than the clock frequency. Power spectral density It exhibits a flat distribution of approximately white noise within the measurement bandwidth. This flatness ensures uniform sampling and calibration of the transfer function of the measurement path across the entire frequency band.

[0031] To suppress the disturbance of the pilot signal to the physical characteristics of the semiconductor device under test, the pilot injection unit is equipped with a programmable attenuator. The digital processing unit calculates and sets the attenuation coefficient of the programmable attenuator based on the bias current of the semiconductor device under test and the estimated noise level. The final pilot voltage amplitude output to the three-terminal symmetrical injection network. It should meet the following requirements: ; in, This represents the output amplitude of the digital-to-analog converter. The magnitude is controlled at the microvolt level, so that its power spectral density is between 1% and 10% of the expected noise spectral density of the semiconductor device under test, thereby avoiding nonlinear shift of the device's operating point while ensuring the signal-to-noise ratio.

[0032] Simultaneously with signal generation, the digital processing unit stores a copy of the generated original digital sequence in a local high-speed memory buffer. This copy serves as a reference signal for subsequent cross-correlation calculations and is time-aligned with the mixed signal containing noise and pilot signals uploaded by the data acquisition unit. Due to the sharp autocorrelation characteristics of the pseudo-random sequence, it satisfies the following in the time domain: ; in, Autocorrelation function, The signal at the The original amplitude at each sampling point Indicates that the signal passes through The amplitude after a unit delay.

[0033] When displacement When the correlation value approaches zero, the system can accurately extract the pilot component from the contaminated channel output signal using correlation identification technology at the digital back end, and use this as a reference to calculate the instantaneous gain drift of the amplification path.

[0034] This invention generates multiple sets of mutually orthogonal pseudo-random sequences by controlling the primitive polynomial feedback logic of a linear feedback shift register. In multi-terminal injection mode, different sequences are coupled to the drain, source, and gate of the semiconductor device under test, respectively. Utilizing the zero cross-correlation between different sequences, the digital processing unit can separate the contribution factors of different electrode paths to the total noise power spectrum, thereby achieving accurate modeling of the internal noise distribution of complex multi-terminal devices.

[0035] This invention provides a method for measuring low-error semiconductor low-frequency noise voltage spectral density, the specific process of which includes the following description: The input to the signal flow model begins at the output node of the semiconductor device under test (DUT). At the output node, there are three superimposed electrical signal components: the noise signal from the DUT. Injected pseudo-random pilot signal and environmental interference signals Noise signal of the device under test The physical noise generated by the random motion of charge carriers inside the semiconductor device under test; pseudo-random pilot signal It is a known time-domain sequence preset and introduced by the pilot injection unit; environmentally relevant interference signals. It refers to coherent noise that exists in the measurement circuit simultaneously, introduced by residual ripple of the bias circuit, power frequency electromagnetic coupling, or other external common interference sources.

[0036] The signal flow model is transmitted through a parallel dual-channel structure. The first and second measurement channels capture and amplify the superimposed signal, respectively. In the first measurement channel, the superimposed signal is processed by the transfer function of the first preamplifier and mixed with background noise specific to this channel. The time-domain output signal of the first measurement channel... Expressed as: ; in, Let be the time-domain impulse response function of the first measurement channel, which determines the channel's gain characteristics and phase delay; (symbol) This represents a temporal convolution operation; This is the independent background noise generated by the first measurement channel itself, and this noise source is unrelated to any of the input signals.

[0037] In the second measurement channel, the same superimposed signal is processed by the second preamplifier and mixed with the background noise of the second measurement channel. The time-domain output signal of the second measurement channel... Expressed as: ; in, This is the time-domain impulse response function of the second measurement channel; Independent background noise generated for the second measurement channel. Background noise of the first measurement channel. Background noise of the second measurement channel In statistics, they must satisfy the characteristics of being independent and uncorrelated, which is the mathematical premise for cross-spectral correlation to suppress background noise.

[0038] The signal flow model is transformed into the frequency domain for representation in the digital processing unit. This is achieved by analyzing the time-domain signal... and Perform a Fast Fourier Transform (FFT) to obtain the corresponding frequency domain sequence. and In the frequency domain, convolution is converted to multiplication, and the linear superposition relationship of the signal components remains unchanged. The frequency domain model reflects the noise components at different frequency points. The energy distribution on the surface.

[0039] In signal flow models, the correlation properties between various signal components are explicitly defined. (Noise signal from the device under test) With pseudo-random pilot signal They are mutually uncorrelated orthogonal. The time-domain impulse response function of the first measurement channel. Time-domain impulse response function of the second measurement channel Due to the discrete nature of electronic components, there are differences in amplitude response and phase response. The role of the signal flow model is to analyze the known components in subsequent steps. Extraction, and reverse solution and The real-time deviation.

[0040] The signal flow model further incorporates environmentally relevant interference signals. The coupling factor. Because When applied to both channels, they share the same source, which manifests as a non-zero correlation component in cross-power spectral density calculations. This constitutes the main source of measurement error. This embodiment addresses this by explicitly defining [the correlation] in the model. and The boundary provides logical support for subsequent coherence stripping using pilot sequences as tags.

[0041] The digital processing unit receives the aforementioned dual-channel digital sequence and, in conjunction with a copy of the pilot reference signal pre-stored in memory, establishes a frequency domain array containing the variables to be solved. and known variables The model uses a system of multivariate equations. By statistically averaging multiple sets of sampled data, the model can separate the power spectral density of each noise component from the complex mixed signal.

[0042] This invention provides a method for measuring the low-frequency noise voltage spectral density of semiconductors with low error, wherein the cross-power spectral estimation method specifically includes the following description: The digital processing unit acquires the digital sequence of the first measurement channel from the data acquisition unit. Second measurement channel digital sequence and the original pseudo-random pilot reference sequence pre-stored in memory. The digital processing unit uses a segmented overlapping averaging method to process the above sequence in order to reduce the variance of the spectral estimation and suppress the interference of random noise on the transfer function extraction.

[0043] The digital processing unit will have a length of The numerical sequence is divided into The data consists of several overlapping data segments, with an overlap rate between adjacent segments set to 50% to 75%. For each data segment, the digital processing unit applies a window function. The window function is either the Hanning window or the Blackman window, which is used to suppress spectral leakage caused by sequence truncation and to ensure that the energy distribution on the frequency axis meets the calculation accuracy requirements.

[0044] For each windowed data sequence, the digital processing unit performs a Fast Fourier Transform (FFT) to obtain the complex vector representation of each data segment in the frequency domain. Among these, the... The frequency domain sequence of the first measurement channel signal is denoted as: , No. The frequency domain sequence of the original pseudo-random pilot reference sequence is denoted as .

[0045] The digital processing unit obtains a single-segment cross-power spectrum estimate by performing point-by-point multiplication of the frequency domain sequence of the channel output signal with the complex conjugate of the pilot reference sequence. Specifically, the first... Cross power spectrum of segments Expressed as: ; in, This represents the complex conjugate of the frequency domain vector of the original pseudo-random pilot reference sequence. This operation preserves the amplitude gain and phase shift information of the channel response relative to the pilot signal.

[0046] Digital processing unit The cross-power spectral density function is obtained by statistically averaging the cross-power spectral density values ​​of each data segment. : ; in, The window function energy normalization factor is defined as the sequence of window functions. The sum of squares divided by the segment length. Through cumulative averaging of a large sample size, the noise components of the device under test and the amplifier's noise floor, which are uncorrelated with the pilot signal, tend to zero in the cross-power spectrum, while the coherent system transfer function characteristics are highlighted.

[0047] The digital processing unit uses the same operational logic to calculate the cross-power spectral density function of the second measurement channel signal and the pilot reference sequence. Due to slight differences in the physical wiring between the two measurement channels, and the asynchronous thermal drift of the active components in the preamplifier over time, the cross-power spectrum estimation method provides an instantaneous spectrum estimation result within the corresponding time window for each set of noise sampling data through continuous iterative calculations.

[0048] The digital processing unit further calculates the autopower spectral density of the original pseudo-random pilot reference sequence. The self-power spectral density serves as the system reference and is used for quantization and normalization of the transfer function of each measurement channel in subsequent steps. Since the generation logic of the pilot signal is known and the waveform is stable, its self-power spectrum exhibits a flat distribution within the preset measurement bandwidth.

[0049] During cross-power spectrum estimation, the digital processing unit monitors the calculation results of the coherence function in real time to evaluate the quality of the current sampled data. If the coherence between the pilot signal and the channel output signal falls below a preset threshold due to sudden external interference, the digital processing unit will automatically discard the current data segment and trigger a resampling process to ensure that the acquired cross-power spectrum accurately reflects the real-time physical state of the measurement system.

[0050] This invention provides a method for measuring low-error semiconductor low-frequency noise voltage spectral density, wherein the instantaneous complex gain extraction specifically includes the following description: The digital processing unit utilizes the cross-power spectral density functions of each channel obtained in the aforementioned steps. , and pilot signal self-power spectral density The instantaneous transfer function of the two measurement channels is extracted by complex division. and The instantaneous transfer function characterizes the end-to-end response from the pilot injection node through the output of the semiconductor device under test, the preamplifier, and finally to the input of the data acquisition module.

[0051] At frequency point At this point, the instantaneous transfer function of the first measurement channel The calculation is as follows: ; The digital processing unit will transfer the function instantaneously. Decomposed into instantaneous amplitude gain and instantaneous phase response The instantaneous amplitude gain reflects the voltage amplification factor of the signal at that frequency, while the instantaneous phase response reflects the frequency-dependent phase shift generated after the signal passes through the measurement path.

[0052] The digital processing unit uses the same mathematical logic to calculate the instantaneous transfer function of the second measurement channel. Because the dynamic output impedance of the semiconductor device under test varies with frequency... The instantaneous transfer function changes and forms a voltage divider relationship with the input impedance of the measurement system. and This actually includes the gain correction caused by impedance mismatch. Since the pilot signal and the noise signal share the same physical path, the extracted complex gain can accurately reflect the real-time distortion of the noise signal during transmission.

[0053] During processing, the digital processing unit performs a consistency evaluation on the extracted instantaneous complex gain. The digital processing unit compares the two channels at the same frequency point. Gain difference at and phase difference : ; ; The difference parameter is stored in a temporary register of the system and is used to characterize the asymmetry of the dual-channel amplifier. This asymmetry is usually caused by differences in electrode contact resistance or temperature drift of operational amplifier internal parameters. By extracting the complex gain in real time, the system can accurately compensate for this physical difference in the digital domain.

[0054] The digital processing unit further eliminates the impact of computational noise generated by random processes on the transfer function envelope through internal mapping algorithms or smoothing processes. The smoothing process employs a moving average filter, with its window width adjusted according to the frequency resolution. Dynamic adjustment. In the extremely low frequency range below 1Hz, the window width is increased to suppress fluctuations; in the high frequency range close to 100kHz, the window width is decreased to preserve the cutoff characteristics details of the path.

[0055] Instantaneous complex gain extraction has a dynamically updated property. Throughout the entire noise measurement cycle, the digital processing unit repeatedly performs the above calculation at fixed time steps (e.g., every 10 seconds). The raw noise data within each measurement time window is associated with a specific set of instantaneous transfer functions. and This correlation mechanism ensures that when fluctuations in ambient temperature cause the preamplifier's quiescent operating point to drift, the system can perform real-time normalization correction of the noise power spectral density using the updated complex gain.

[0056] In this embodiment, the extracted instantaneous complex gain is also used to monitor the linearity of the measurement system. The digital processing unit determines whether the measurement path has entered a saturation state or is subject to nonlinear interference by calculating the ratio of the magnitude of the cross-power spectrum to the product of the power of each signal, i.e., the coherence function. If the extracted transfer function experiences unexpected and drastic fluctuations over a wide frequency band, the system will determine that the current measurement environment is overloaded and issue a gain adjustment command to the bias supply unit or amplification unit.

[0057] See attached document Figure 2 This invention provides a method for measuring low-error semiconductor low-frequency noise voltage spectral density, wherein the dynamic drift monitoring mechanism specifically includes the following description: The digital processing unit has a reference parameter memory for storing the start time of the measurement cycle (i.e., Initial transfer function extracted at time (time) and This serves as the static calibration benchmark for the system. During subsequent continuous measurements, the digital processing unit performs the aforementioned cross-correlation calculations at a preset time step. Periodically update the current instantaneous transfer function and And compare it with the static calibration benchmark.

[0058] The digital processing unit calculates the drift coefficient matrix. To quantify the performance evolution of the measurement path, the drift coefficient matrix The mathematical expression is: ; in, The value is either 1 or 2, representing two measurement channels respectively. This drift coefficient is a complex number; its magnitude reflects the change in channel gain, and its argument reflects the phase shift over time. The digital processing unit stores the drift coefficient matrix in a cache and uses it as a real-time compensation weight in the subsequent voltage spectral density reconstruction process.

[0059] During the monitoring process, the digital processing unit is set with a gain drift threshold. and phase drift threshold If at a certain moment Detected If the change in gain exceeds the gain drift threshold, or the change in phase angle exceeds the phase drift threshold, the digital processing unit will trigger a compensation procedure. The compensation procedure offsets gain fluctuations caused by ambient temperature fluctuations or active device aging by performing an inverse proportional weighting on the original sampled sequence within that time period at the digital back end, thereby ensuring the stability of the final reconstructed noise voltage spectral density across time.

[0060] To address the long-term sampling characteristics required for semiconductor low-frequency noise measurement, the dynamic drift monitoring mechanism also includes indirect monitoring of the bias circuit stability. This is due to the injected pseudo-random pilot signal. When the operating point of the semiconductor device under test changes due to DC bias drift at its output terminal, its output dynamic impedance will change synchronously and be reflected in the extracted transfer function. The frequency response curve shape. The digital processing unit analyzes the drift coefficient matrix. The nonlinear distortion level over a wide frequency band is used to identify whether the system has experienced a bias point shift, and a calibration command is sent to the bias supply unit when the bias voltage drift exceeds the preset safety range.

[0061] The digital processing unit also incorporates a statistical property evaluation of the sliding time window. Within each monitoring period, the digital processing unit calculates the variance of the drift coefficient. If the variance value shows an abnormal increase within a specific frequency band, it is determined that there is non-stationary high-frequency pulse interference or electromagnetic shielding failure in the measurement environment. In this case, the digital processing unit automatically reduces the confidence weight of the data for that time period in the final power spectrum synthesis, and suppresses the impact of abnormal data on the measurement results through a weighted statistical averaging algorithm.

[0062] Furthermore, the dynamic drift monitoring mechanism is linked with the aforementioned data acquisition unit. The digital processing unit dynamically adjusts the synchronous clock phase compensation value of the analog-to-digital converter based on the monitored phase drift rate. Since the dual-channel cross-spectral method is highly dependent on the phase consistency of the two channel signals, this real-time phase alignment technology based on pilot extraction results can compensate for the minute time delays introduced by cable bending, vibration, or changes in interface contact resistance, ensuring that the coherent terms near 100kHz do not produce false energy attenuation due to phase mismatch.

[0063] The digital processing unit records the drift trajectory throughout the measurement process as a system log file. This file contains the complex gain records for each frequency point at different timestamps, providing underlying data support for subsequent analysis of the long-term reliability of the semiconductor device under test and evaluation of the baseline stability of the measurement system.

[0064] This invention provides a method for measuring low-error semiconductor low-frequency noise voltage spectral density, wherein the coherence factor... The construction logic specifically includes the following description: The digital processing unit acquires the instantaneous transfer function of the first measurement channel. Instantaneous transfer function of the second measurement channel and the cross power spectral density of the two channel output signals. Then, a quantitative assessment process for environmental coherence characteristics is initiated. Coherence factor. The purpose of this project is to establish a normalized observation platform in the complex domain that can isolate the influence of system gain deviation by leveraging the known physical properties of the pilot signal, thereby identifying the non-random environmental interference components hidden in the cross-spectral data.

[0065] The digital processing unit first performs normalization processing on the cross-spectral data. The cross-power spectral density is then... Dividing by the complex product of the instantaneous transfer functions of the two channels, the measurement result is equivalently traced back from the amplifier output to the output node of the semiconductor device under test. Based on this, the digital processing unit further subtracts the self-power spectral density of the pilot signal pre-stored in memory. This is to eliminate the contribution of the known injected signal to the total power density. Coherence factor The mathematical construct expression is as follows: ; in, This represents the complex conjugate of the instantaneous transfer function of the second measurement channel. The formula is used to calculate... Includes the intrinsic noise power spectral density of the semiconductor device under test and environmentally related interference components Comprehensive information.

[0066] In an ideal, interference-free environment, due to the noise of the semiconductor device under test... The two channels are perfectly correlated, and the background noise of each channel is independent. (Normalized) It should be represented as a pure real number and conform to the physical distribution law of semiconductor noise. The digital processing unit monitors... The severity of environmental interference is assessed by the imaginary part amplitude. If at a specific frequency point, If the imaginary part deviates significantly from zero level, it is determined that there is phase nonlinear coherent interference caused by external electromagnetic coupling in this frequency band.

[0067] The digital processing unit further utilizes the coherence factor Frequency-resolved correlation quality of the signal is assessed. For common power frequency interference in low-frequency noise measurements (such as 50Hz and its harmonics), the coherence factor will produce narrowband spikes at the corresponding frequency axis positions. The digital processing unit constructs a mask function based on the correlation characteristics of the pilot sequence to compare... With ideals The deviation of the curve model. If If the local slope or energy distribution does not match the preset semiconductor noise physical model, then this portion of energy is determined to belong to the environmentally related noise floor. .

[0068] The digital processing unit also through A time-dimensional coherence evolution monitoring mechanism was established. Due to the pilot signal... It is injected continuously, and the digital processing unit calculates the values ​​within adjacent measurement periods. The statistical correlation. If If the coherence of certain frequency components remains highly constant over time, and these components do not change by an order of magnitude with changes in the bias voltage of the measured semiconductor device, then the digital processing unit will label these components as the system's inherent correlated noise floor and remove them from the final... Remove from the calculation results.

[0069] coherence factor The construction logic also involves secondary correction of the phase consistency of the dual channels. Digital Processing Unit Analysis In the high-frequency band, if a phase drift trend is found to be caused by differences in cable length or inconsistent delays in the acquisition channel group, the digital processing unit will extract the phase slope and apply it inversely to the transfer function. and Ensure that in the calculation At this time, the signals of the two channels are mathematically perfectly aligned. This closed-loop correction logic based on result feedback effectively avoids the phenomenon of falsely low measurement values ​​caused by hardware channel mismatch.

[0070] Finally, the digital processing unit will construct the coherence factor. This data is used as the underlying data input to subsequent noise stripping algorithms. Through... By performing real part extraction and narrowband denoising, the system can obtain semiconductor intrinsic noise distribution data with extremely high purity. This construction method not only considers the dynamic changes in channel gain, but also achieves fine-grained management of all coherent components on the measurement path through the tagging effect of pilot signals.

[0071] This invention provides a method for measuring low-error semiconductor low-frequency noise voltage spectral density, wherein the correlated noise... The digital filter stripping specifically includes the following description: The digital processing unit constructs the coherence factor using the aforementioned steps. Activate environmentally relevant background noise The identification and digitization stripping process. Due to the coherence factor Ideally, it should only contain the intrinsic noise power spectral density of the semiconductor device under test. However, in actual measurement environments, residual ripple from the bias circuit and spatial electromagnetic coupling can introduce interference signals that are highly coherent with the measurement channel. ,lead to Non-random correlation noise components are superimposed in the middle. .

[0072] The digital processing unit first identifies [the problem] using spectral refinement analysis technology. The narrowband coherent interference component is identified. The digital processing unit searches for energy abrupt changes on the frequency axis where the coherence function is close to 1 and the bandwidth is extremely narrow. These abrupt changes typically correspond to power frequency interference and its higher harmonics. The digital processing unit establishes a dynamic notch filter model for these specific frequency points, and its normalized amplitude characteristics are analyzed. The definition is as follows: ; in, Exponential function The independent variable represents frequency. The center frequency of the identified coherent interference, This is the interference bandwidth control factor. The digital processing unit applies this filtering characteristic to... This achieves the initial elimination of background noise related to narrowband environments.

[0073] For the broadband distributed environmentally relevant noise floor, the digital processing unit employs an adaptive cancellation algorithm based on pilot references. This is due to the injected pseudo-random pilot signal. Transfer function in each channel and The data has been accurately extracted, allowing the system to establish a coherence reference. The digital processing unit identifies those non-conforming parameters of the cross-power spectrum by calculating the complex phase distribution of the cross-power spectrum. Energy components of noise phase characteristics.

[0074] The digital processing unit further introduces a contrast stripping logic based on bias state modulation. During measurement, the digital processing unit controls the bias supply unit to change the drain current of the semiconductor device under test while keeping the pilot signal injection state unchanged. Due to inherent noise... The magnitude varies with the current, while the associated noise floor originates from environmental coupling. Typically, the digital processing unit does not change with the device's operating point; it processes data under different bias states. Differential operations are performed to extract the constant coherent component that does not change with the current, and this component is defined as the estimated correlation noise floor. .

[0075] The digital processing unit subtracts the estimated correlation floor from the coherence factor. The purified power spectral density components are obtained. The mathematical expression after stripping is: ; in, Pure power spectrum or corrected power spectrum This indicates the extraction of the real part of the result of a complex number operation.

[0076] During this process, the digital processing unit only extracts the real part of the complex number operation result. This is due to the background noise in both measurement channels. and The components are independent of each other, with a cross-spectral expectation of zero, while the intrinsic noise, after transfer function normalization, is accumulated on the real part axis. This stripping algorithm eliminates spurious coherence terms introduced by shared power or ground loops.

[0077] To address potential non-stationary correlation interference in the extremely low frequency band (below 10Hz), the digital processing unit (DMU) performs a dynamic masking process based on time-series variance analysis. The DMU monitors the volatility of the coherence factor within a short time window. If the coherence energy of a certain frequency band exhibits abnormally drastic fluctuations unrelated to the pilot signal, the DMU automatically reduces the weighting factor of that frequency band in the final statistical average. This weighted stripping mechanism effectively suppresses measurement errors caused by transient electromagnetic environment changes.

[0078] After completing the noise floor stripping, the digital processing unit also performs a consistency check on the continuity of the spectral curve. Utilizing the power-law distribution characteristics of semiconductor noise in the frequency domain, the digital processing unit performs local fitting of the stripped data points using the least squares method. If the residual components after stripping cause non-physical steps or dips in the spectral curve at specific frequency bands, the system will automatically fine-tune the step size parameters of the adaptive filter and re-execute the stripping calculation until the spectral envelope conforms to the preset physical logic distribution.

[0079] Through the aforementioned digital filtering and stripping steps, this invention achieves refined processing of complex physical interference at the digital backend. This method no longer relies on extremely stringent hardware isolation schemes, but instead utilizes real-time transfer function information provided by pilot signals, using correlation as the criterion to distinguish intrinsic noise from environmental interference, ensuring the acquisition of high-purity semiconductor low-frequency noise data even in complex industrial environments.

[0080] This invention provides a method for measuring low-error semiconductor low-frequency noise voltage spectral density, wherein the accurate subtraction of the pilot's own energy specifically includes the following description: Due to pseudo-random pilot signal As a known signal actively injected into the measurement loop, its energy is superimposed on the output of the dual-channel amplification unit. To obtain the intrinsic noise power spectral density of the semiconductor device under test, the digital processing unit must accurately subtract the energy component contributed by the pseudo-random pilot signal from the synthesized cross-power spectral data. The accurate subtraction process is based on the principle of energy conservation and the orthogonal decoupling characteristics of the signal.

[0081] The digital processing unit first retrieves the autopower spectral density of the original pseudo-random pilot reference sequence from system memory. .Should During the signal generation stage, the voltage amplitude is preset. and clock cycle The established theoretical baseline value represents the initial energy distribution at the injection node.

[0082] The digital processing unit utilizes the instantaneous transfer function of the first measurement channel extracted in the aforementioned steps. Instantaneous transfer function of the second measurement channel Calculate the equivalent correlation energy term of the pseudo-random pilot signal after dual-channel amplification and transmission. The mathematical formula for calculating the equivalent related energy term is as follows: ; in, This is the complex conjugate of the instantaneous transfer function of the second measurement channel. This calculation process not only considers the amplitude gain of the pilot signal from both channels, but also introduces the instantaneous phase difference between the two channels through complex multiplication. Since the pilot signal originates from the same source when entering the two amplification channels, its coherence at the output is determined by the frequency points. The complex gain product at the point is fully represented.

[0083] The digital processing unit will convert the relevant energy terms into equivalent terms. From the original cross power spectral density Vector subtraction is performed. Since this subtraction process is completed in the complex frequency domain, it can simultaneously cancel the real energy component and the imaginary phase shift generated by the pilot signal. The difference spectrum obtained from the subtraction operation... It reflects the strength of the residual correlation signal after removing known pilot interference.

[0084] During the deduction process, the digital processing unit introduces an energy matching correction factor. The energy matching correction factor is dynamically iterated by monitoring the residual component at the center frequency of the pilot signal. If, after subtraction, a non-physical dip or bulge is found in the resulting spectral line within the pilot characteristic frequency band, it is determined that there is a deviation in the transfer function extraction accuracy. The digital processing unit will then fine-tune the process in reverse by minimizing the residual energy. The weights involved in the calculation are used to ensure that the subtraction operation does not excessively cancel the inherent noise background of the semiconductor device under test.

[0085] To address the potential circular convolution effect that pseudo-random sequences may exhibit at the edges of the sampling window, the digital processing unit performs time delay calibration on the equivalent correlation energy term before subtraction. The calibration process adjusts the time delay based on the physical trace delay of the pilot signal between the injection unit and the acquisition unit. and The phase slope. This physical time alignment ensures that the two terms to be subtracted are strictly synchronized on the time axis when performing subtraction, thus avoiding incomplete subtraction caused by time jitter.

[0086] The digital processing unit completely removes the active tag signal injected for calibration from the final result through this step. The advantage of this subtraction method is that its compensation is based entirely on the real-time acquired system transfer function, rather than relying on fixed factory calibration values. This allows the measurement system to maintain the spectral flatness after pilot subtraction even under conditions of drastic ambient temperature fluctuations and nonlinear shifts in the analog front-end gain.

[0087] Finally, the difference spectrum after precise subtraction of the pilot's own energy is obtained. The signal is input into the normalization and reconstruction module. Since the spectrum no longer contains the energy of the injected signal, subsequent calculations will focus entirely on the intrinsic physical noise characteristics of the semiconductor device under test under specific bias conditions.

[0088] This invention provides a method for measuring low-error semiconductor low-frequency noise voltage spectral density, wherein the normalized reconstruction of the low-error voltage spectral density specifically includes the following description: The digital processing unit acquires the complex cross-power spectrum sequence after pilot energy subtraction and ambient noise stripping. This is to obtain the intrinsic noise voltage spectral density equivalent to the output of the semiconductor device under test. The digital processing unit performs normalization calculations based on the instantaneous transfer function. Normalization calculations eliminate the non-idealities in the frequency response of the amplification path, restoring the measured signal from its digitized quantized value to a squared distribution of the physical voltage amplitude.

[0089] The digital processing unit calculates the intrinsic noise voltage spectral density of the semiconductor device under test using the following reconstruction formula. : ; in, The original cross-power spectral density of the dual-channel output signal; and These are the instantaneous transfer functions of the first and second measurement channels, respectively. for The complex conjugate; Given the self-power spectral density of the pilot signal; This indicates the extraction of the real part of the result of a complex number operation; This is a preset system residual noise correction term.

[0090] When the digital processing unit performs the denominator operation in the above formula, it achieves dynamic compensation for dual-channel gain mismatch and phase inconsistency. Because... and It incorporates the complex gain characteristics of the preamplifier as a function of frequency, and the product operation of the denominator terms can automatically cancel out the phase shift difference generated by the two amplifiers at the high-frequency end. This processing method ensures that the magnitude of the cross power spectrum can be accurately mapped to the true noise energy of the device under test across the entire frequency band (especially in the high-frequency band where the phase shift is significant).

[0091] The digital processing unit performs the real part extraction operation. The technical basis lies in the inherent noise of the semiconductor device under test. The two measurement channels are completely coherent and in phase, and their cross-power spectrum exhibits a purely real function; while the two measurement channels generate independent background noise. and After statistical averaging, the cross-spectral expectation value tends to zero. By extracting the real part, the digital processing unit can suppress the interference of residual incoherent components on the spectral density calculation results and improve the signal-to-noise ratio of the measured values.

[0092] To further reduce the statistical variance of the measurement results, the digital processing unit reconstructs the data. The cumulative averaging of multiple sets of samples is performed. The digital processing unit sets the number of cumulative averaging iterations according to the preset frequency resolution requirements. In the low-frequency range of 1Hz to 100Hz, a larger time window is used for sampling to ensure frequency resolution; in the high-frequency range of 100Hz to 100kHz, the number of cumulative averaging iterations is increased. To reduce random fluctuations. After The voltage spectral density after averaging converges to the true physical quantity value.

[0093] The digital processing unit also performs impedance voltage division compensation during the reconstruction process. This is because the output impedance of the semiconductor device under test... Input impedance of the preamplifier This forms a voltage divider relationship, and the actual acquired signal is the result of the voltage divider. The digital processing unit retrieves the device impedance model pre-stored in memory and corrects the normalized result using coefficients. satisfy: ; The digital processing unit will Multiply by the correction factor This eliminates the system error introduced by the change in impedance matching state with frequency.

[0094] The digital processing unit outputs the final... Before the curve, perform a data validity check. The check includes a logical determination of the spectral slope, for example, in... For the noise-dominant frequency band, determine whether its spectral slope is within the physically reasonable range of -0.8 to -1.2. If an abnormal energy spike or spectral line morphology that does not conform to physical laws is detected, the digital processing unit will mark the data in that frequency band as invalid and automatically adjust the pilot injection power and sampling gain to start a new measurement cycle.

[0095] Finally, the normalized and reconstructed semiconductor low-frequency noise voltage spectral density data is stored in the non-volatile memory of the digital processing unit in the form of a double logarithmic coordinate system of frequency amplitude, and can be exported through an external communication interface. This data not only eliminates the background noise of the hardware system, but also eliminates the dynamic drift error during the measurement period through the pilot feedback mechanism, achieving high-precision noise characteristic characterization.

[0096] The low-error semiconductor low-frequency noise voltage spectral density measurement method provided by this invention achieves real-time correction of non-ideal characteristics of the measurement system through the collaboration of hardware circuits and digital processing algorithms.

[0097] At the hardware deployment level, the system connects to the semiconductor device under test (DUT) via a bias supply unit, providing it with a precise DC bias. The pseudo-random pilot signal generated by the pilot injection unit is coupled to the DUT's input terminal via a three-terminal symmetrical injection network. The dual-channel amplification unit contains two sets of physically symmetrical low-noise preamplifiers used to capture the inherent noise superimposed on the pilot signal in parallel. The data acquisition unit synchronously converts the analog mixed signal into a digital sequence and inputs it to the digital processing unit.

[0098] At the digital processing level, the system first performs cross-power spectrum estimation using the acquired digital sequence and the original pseudo-random pilot reference sequence to extract the instantaneous transfer function at each frequency point. The instantaneous transfer function characterizes the complex characteristics of the path in real time, including preamplifier gain fluctuations, phase shifts, and wiring parasitic parameters. Subsequently, the system calculates the deviation between the current time and the initial reference time based on the instantaneous transfer function, constructs a drift coefficient matrix, and applies an inverse proportional weighting to the original sequence to offset the system drift caused by changes in ambient temperature.

[0099] To address environmental coherent interference, the system constructs an environmental coherence factor using an instantaneous transfer function, tracing the measured signal back to the physical output node. By monitoring the imaginary part of the traced signal, the system identifies and locates power frequency interference and its harmonics. After interference stripping and pilot energy subtraction, the system performs normalized reconstruction to obtain an accurate intrinsic noise voltage spectral density.

[0100] Specific application examples: To verify the effectiveness of the above method, this embodiment conducted low-frequency noise characterization tests on an N-channel enhancement-mode field-effect transistor under typical experimental conditions.

[0101] The measurement system operated continuously at room temperature. The sampling frequency was set to 1 kHz, and the measurement bandwidth covered 0.1 Hz to 500 Hz. A pilot injection unit generated a pseudo-random binary sequence of order 17. During the measurement process, observable system gain drift (simulating active device temperature drift) and 50 Hz power frequency electromagnetic interference were artificially introduced to evaluate the corrective capabilities of the invention.

[0102] The measurement results of this embodiment are as follows: Figure 3 As shown. Figure 3 This is a logarithmic graph of voltage spectral density as a function of frequency, containing three curves with clear technical and physical meanings.

[0103] The red solid line represents the original measurement results affected by drift and environmental interference. As can be seen from the figure, the curve has a significant energy spike at 50Hz, and due to the decay of the system gain over time, the amplitude of the curve in the low-frequency range deviates from the physical expectation, causing distortion in the fitted 1 / f noise slope.

[0104] The thick blue solid line represents the intrinsic noise reference curve under ideal conditions, reflecting the true noise distribution of the tested semiconductor device under no systematic error conditions as the underlying comparison benchmark.

[0105] The thick green dashed line represents the inherent noise result after reconstruction using the method of this invention.

[0106] As can be seen from the comparison, the method of the present invention can accurately identify and suppress... Figure 3 The power frequency interference spike at 50Hz is shown, and the reconstructed curve (thick green dashed line) and the ideal reference curve (thick blue solid line) are highly overlapping and alternate across the entire frequency band.

[0107] Experimental data show that this invention reduces the gain error caused by temperature drift to below 0.1 dB through real-time compensation of the instantaneous transfer function. By utilizing the environmental coherence factor and dynamic notch filtering technology, the residual amount of 50 Hz power frequency interference in the reconstructed spectral density is reduced by approximately 25 dB. The resulting noise spectrum slope is stabilized within a preset physical range, effectively supporting high-precision characterization of internal defects and reliability in semiconductor devices.

[0108] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for measuring low-error semiconductor low-frequency noise voltage spectral density, characterized in that, Includes the following steps: A static operating point is established by providing DC bias to the semiconductor device under test through the bias supply unit; The pseudo-random pilot signal is generated by the pilot injection unit and injected into the input terminal of the semiconductor device under test as a reference for in-path characterization. The mixed signal, which is the superposition of the intrinsic noise generated by the semiconductor device under test and the pseudo-random pilot signal, is captured in parallel by a dual-channel amplification unit, and the digital sequence of the first measurement channel and the digital sequence of the second measurement channel are acquired simultaneously by a data acquisition unit. The digital processing unit performs cross-power spectrum estimation between the acquired digital sequence and the original pseudo-random pilot reference sequence to extract the instantaneous transfer function that reflects the complex gain characteristics of the measurement path. Real-time drift compensation is performed on the digital sequences of the first and second measurement channels according to the instantaneous transfer function, and an environmental coherence factor is constructed according to the instantaneous transfer function. The environmental coherence factor is used to remove the environmental-related background noise from the compensated measurement data, and the instantaneous transfer function is used to subtract the pilot energy term corresponding to the pseudo-random pilot signal. Normalization reconstruction is performed on the stripped and deducted data to obtain the intrinsic noise voltage spectral density equivalent to the output of the semiconductor device under test.

2. The method for measuring low-error semiconductor low-frequency noise voltage spectral density according to claim 1, characterized in that, The steps for generating the pseudo-random pilot signal include: A pseudo-random pilot signal is generated based on a primitive polynomial using a linear feedback shift register structure. The order of the linear feedback shift register is preset according to the measurement bandwidth, and the driving clock frequency is not less than a preset multiple of the maximum measurement frequency.

3. The method for measuring low-error semiconductor low-frequency noise voltage spectral density according to claim 1, characterized in that, The steps of injecting the pseudo-random pilot signal include: The pseudo-random pilot signal is coupled to the input terminal of the semiconductor device under test through a three-terminal symmetrical injection network. The injection amplitude is adjusted by a programmable attenuator so that the injected pilot power spectral density is within the preset proportion range of the expected noise spectral density.

4. The method for measuring low-error semiconductor low-frequency noise voltage spectral density according to claim 1, characterized in that, The steps for cross-power spectrum estimation include: The digital sequence is windowed using a segmented overlapping averaging method. The windowed frequency domain sequence is then multiplied point-by-point with the original pseudo-random pilot reference sequence in the frequency domain and statistically averaged to obtain the cross-power spectral density function of each measurement channel relative to the original pseudo-random pilot reference sequence.

5. The method for measuring low-error semiconductor low-frequency noise voltage spectral density according to claim 1, characterized in that, The steps for extracting the instantaneous transfer function include: The cross-power spectral density function obtained by cross-power spectral estimation is divided by the auto-power spectral density of the original pseudo-random pilot reference sequence to calculate the instantaneous transfer function containing complex magnitude and complex phase. The instantaneous transfer function is then smoothed by dynamically adjusting the window width with the measurement frequency using a moving average filter.

6. The method for measuring low-error semiconductor low-frequency noise voltage spectral density according to claim 1, characterized in that, The real-time drift compensation steps include: The ratio between the instantaneous transfer function at the current moment and the reference transfer function at the start of the measurement is calculated to construct a drift coefficient matrix. Then, inverse proportional weighting is performed on the digital sequences of the first and second measurement channels at the digital back end to offset the gain and phase error introduced by the temperature drift of the active device.

7. The method for measuring low-error semiconductor low-frequency noise voltage spectral density according to claim 1, characterized in that, The steps for constructing the environmental coherence factor include: The cross-power spectral density of the first and second measurement channel digital sequences is divided by the complex product of the extracted instantaneous transfer functions, and the mixed signal is equivalently traced back to the physical output node of the semiconductor device under test. The presence of external nonlinear coherent interference is determined based on the amplitude of the imaginary part of the traced signal.

8. The method for measuring low-error semiconductor low-frequency noise voltage spectral density according to claim 1, characterized in that, The step of stripping away environmentally relevant background noise includes: Narrowband energy abrupt change points in the environmental coherence factor are identified and power frequency interference is eliminated using a dynamic notch filter. Multiple measurements are performed by controlling the bias supply unit to change the drain current of the semiconductor device under test, and differential operations are used to eliminate constant coherence components that do not change with current.

9. The method for measuring low-error semiconductor low-frequency noise voltage spectral density according to claim 1, characterized in that, The steps for deducting the pilot energy term corresponding to the pseudo-random pilot signal include: The equivalent correlation energy term of the pseudo-random pilot signal after dual-channel amplification is calculated using the instantaneous transfer function. The equivalent correlation energy term is then subjected to complex vector subtraction from the cross-power spectral density of each channel digital sequence. Finally, the energy matching correction factor is iterated based on the residual energy of the pilot characteristic frequency band.

10. The method for measuring low-error semiconductor low-frequency noise voltage spectral density according to claim 1, characterized in that, The normalization reconstruction steps include: An impedance matching correction coefficient is constructed using the output impedance of the semiconductor device under test and the input impedance of the dual-channel amplifier unit, and voltage division compensation is performed. The validity of the curve corresponding to the obtained intrinsic noise voltage spectral density is verified by checking whether the spectral slope is within a preset physical range.

Images