High-precision signal conditioning system and conditioning method

Abstract

The invention relates to the technical field of high-precision signals and signal processing and optimization, in particular to a high-precision signal conditioning system and method, and the system comprises a signal conversion module which comprises a current frequency conversion circuit, a voltage frequency conversion circuit and a capacitor voltage conversion circuit; the filtering processing module comprises a low-pass filter and a band-pass filter; noise and interference in the signals are removed, and the quality and stability of the signals are improved; the isolation and amplification module is used for isolating and amplifying the signals; the VAE-LSTM joint algorithm module fuses the advantages of a variational auto-encoder VAE and a long short-term memory network LSTM, and carries out depth feature extraction, noise filtering and time sequence modeling and prediction on signals. According to the method, the deep mode of the time sequence can be effectively captured, and a stable and reliable basis is provided for anomaly detection and signal prediction. The robustness of dynamic signal processing is enhanced; real-time anomaly detection can be realized, and reliable support is provided for equipment state monitoring and fault early warning. And the intelligent level of signal processing is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of signal processing and optimization, and in particular to a high-precision signal conditioning system and a conditioning method. Background Art

[0002] With the development of modern signal processing technology, the precise conditioning and optimization of dynamic signals play an important role in industrial control, communication systems, and medical monitoring. However, traditional signal conditioning methods mainly rely on filters with fixed parameters and simple mathematical models, which are difficult to deal with complex nonlinear signals and time series characteristics. In particular, there are significant deficiencies in the following aspects: 1. Noise interference problem: Signals in complex environments are usually interfered by various noises, such as random noise and structured noise in sensor acquisition signals. Existing methods often cause signal feature loss while effectively denoising. 2. Traditional signal conditioning systems lack real-time response capabilities for dynamically changing signals, especially in the processing of signals with significant time series characteristics (such as industrial vibration signals and medical ECG signals), showing insufficient prediction capabilities. 3. Limited anomaly detection capabilities: In key application scenarios such as industry and medicine, signal anomaly detection is crucial. However, the existing technology relies on fixed thresholds for the detection of abnormal signals and lacks an adaptive detection mechanism based on deep features, resulting in high false alarm and missed alarm rates.

[0003] In recent years, signal processing technology based on machine learning has gradually emerged. Among them, variational autoencoders (VAE) have significant advantages in signal feature extraction and noise reduction; long short-term memory networks (LSTM) have excellent performance in time series modeling and prediction. However, the research on the joint application of VAE and LSTM in dynamic signal conditioning is still in its infancy. How to achieve software and hardware coordinated optimization with the support of hardware conditioning modules and improve the real-time and robustness of the system is still an important problem that needs to be solved by existing technologies. Summary of the invention

[0004] 1. Technical problems to be solved

[0005] The purpose of the present invention is to solve the problems of noise interference, insufficient dynamic response capability and poor abnormality detection effect in dynamic signal conditioning in the prior art, and to provide a high-precision signal conditioning system and conditioning method.

[0006] 2. Technical solution

[0007] The present invention provides a high-precision signal conditioning system, comprising:

[0008] Signal conversion module: includes current frequency conversion circuit, voltage frequency conversion circuit and capacitor voltage conversion circuit; converts signals of different forms into a standard form that is convenient for subsequent processing.

[0009] The filtering processing module includes a low-pass filter and a band-pass filter; it is used to remove noise and interference in the signal and improve the quality and stability of the signal.

[0010] The isolation amplifier module is used to isolate and amplify the signal, thereby improving the safety, stability and anti-interference ability of the system.

[0011] VAE-LSTM joint algorithm module: includes: VAE module and LSTM module; VAE module is used to extract the potential features of the signal and filter out noise; LSTM module is used to perform time series modeling and prediction on the potential features extracted by the VAE module. The VAE-LSTM joint algorithm module combines the advantages of variational autoencoder (VAE) and long short-term memory network (LSTM), and can perform deep feature extraction, noise filtering, and time series modeling and prediction on the signal.

[0012] As an optional solution of the present invention, the current-frequency conversion circuit converts the input current signal into a frequency signal, has the characteristics of high precision, high stability and strong anti-interference ability, and can accurately convert tiny current changes into recognizable frequency changes.

[0013] The voltage-frequency conversion circuit converts the input voltage signal into a frequency signal, which not only improves the signal's anti-interference ability, but also facilitates subsequent digital signal processing and transmission.

[0014] The capacitance-to-voltage conversion circuit converts the capacitance change output by the capacitance sensor into a voltage signal. It converts the small change of capacitance into a measurable voltage signal.

[0015] As an optional solution of the present invention, the low-pass filter allows low-frequency signals to pass through and suppresses high-frequency signals.

[0016] A bandpass filter allows only signals within a specific frequency range to pass through, while suppressing signals at other frequencies.

[0017] Furthermore, the VAE module (Variational Autoencoder) extracts the latent features of the signal and performs noise filtering. The VAE module consists of two parts: an encoder and a decoder. The encoder maps the input signal to the mean and variance in the latent space, and generates a latent vector from the latent space by random sampling; the decoder decodes this latent vector into a reconstructed signal. By minimizing the error between the reconstructed signal and the original signal and the difference between the distribution of the latent vector and the prior distribution, VAE can learn the effective feature representation of the signal and filter out noise to a certain extent.

[0018] The LSTM (Long Short-Term Memory Network) module performs time series modeling and prediction on the potential features extracted by the VAE module. The LSTM module receives the potential feature sequence output by the VAE module and learns and models these features through its internal memory units and gating mechanism. LSTM predicts future signal trends based on historical feature information, providing strong support for further processing and analysis of the signal.

[0019] As an optional solution of the present invention, the VAE module generates potential distribution parameters of the signal through the encoder, and the mean and logarithmic variance are:

[0020] μ=f μ (x;θ e ),logσ 2 =f σ2 (x;θ e ), where μ and logσ 2 are the mean and log variance of the underlying distribution, θ e is the trainable parameter of the encoder. x represents the signal input to the encoder of the VAE module; f u (◆) is a function determined by the encoder, which takes the input signal x and the encoder’s trainable parameters θ e As input, it outputs the mean u of the underlying distribution. σ2 (◆) is a function determined by the encoder, which takes the input signal x and the trainable parameters θ of the encoder e As input, it outputs the log variance of the underlying distribution.

[0021] As an optional solution of the present invention, the VAE module calculates the latent feature z by sampling the latent space:

[0022]

[0023] Where, ∈ is the standard normal distribution noise of random sampling; σ=exp(0.5·logσ 2 ). z represents the latent feature calculated by sampling in the latent space. N(0,I) represents the multivariate standard normal distribution. Among them, 0 represents the mean vector, indicating that the mean of the distribution is 0 in all dimensions; I is the unit matrix, which represents the covariance matrix, which means that each dimension is independent of each other and the variance is 1.

[0024] As an optional solution of the present invention, the LSTM module updates the time series state of the signal through the following steps:

[0025] Forget gate: f t =σ(W f ·[h t-1 ,z t ]+b f );

[0026] In the formula, f t is the output of the forget gate, with a value range of [0,1], indicating the cell state C at the previous time step t-1 The forgotten proportion; W f is the weight matrix of the forget gate; b f is the bias vector of the forget gate; h t-1 is the hidden state of the previous time step; z t is the input feature of the current time step; is the sigmoid activation function

[0027] Input gate: i t =σ(W i ·[h t-1 ,z t ]+b i ),

[0028] In the formula, i t It is the output of the input gate, with a value range of [0,1], indicating the degree of influence of the current input on the cell state; is the candidate value of the cell state at the current time step; W i ,W c are the weight matrices of the input gate and cell state candidate values ​​respectively; b i ,b c are the bias vectors for the candidate values ​​of the input gate and cell state, respectively; is the hyperbolic tangent activation function, with a value range of [-1,1];

[0029] Cell Status: In the formula, C t is the cell state at the current time step; C t-1 is the cell state at the previous time step.

[0030] Output gate: o t =σ(W o ·[h t-1 ,z t ]+b o ),h t =o t tanh(C t );

[0031] In the formula, o t is the output of the output gate, with a value range of [0,1], indicating the cell state C t For the hidden state h t Output ratio; h t is the hidden state of the current time step; W f,W i ,W o ,W c , and b f ,b i ,b c ,b o are the trainable parameters of LSTM.

[0032] As an optional solution of the present invention, the LSTM module uses the hidden state h t Predict the output value of the signal

[0033]

[0034] Among them, W y is the weight matrix of the prediction layer, b y is the bias parameter. t represents the time step. h t is the hidden state of the LSTM module at time step t. y is the weight matrix of the prediction layer. b y is the bias vector of the prediction layer.

[0035] As an optional solution of the present invention, the potential space dimension z of the VAE module is compressed to 10%-30% relative to the dimension of the input signal, and the potential space dimension z is calculated by the following formula:

[0036]

[0037] As an optional solution of the present invention, the LSTM module uses a sliding window technology to dynamically model the time series signal, and the window length is T w , each update is based on the following formula:

[0038]

[0039] Where Tw is the length of the sliding window; t is the time step index; y t Indicates that at time step

[0040] ^

[0041] The actual value of t, that is, the actual observation data of the time series signal at that moment. t represents the predicted output at time step t; L LSTM is the loss function value based on the LSTM model. ||·|| represents a norm operation.

[0042] As an optional solution of the present invention, the LSTM module uses a threshold to judge abnormal signals based on the deviation between the prediction error and the actual signal. The deviation calculation formula is:

[0043] If δ t >δ 阈值 , then the abnormal signal is determined. In the formula, δ t Represents the deviation between the prediction error and the actual signal at time step t. 阈值 Indicates a preset fixed value, which serves as the standard limit for judging whether a signal is abnormal.

[0044] As an optional solution of the present invention, the VAE-LSTM joint algorithm module dynamically adjusts the parameters of the filter module through the following feedback adjustment mechanism:

[0045] Optimize the filter frequency range based on the signal prediction results of LSTM;

[0046] Based on the VAE reconstructed signal, adjust the zero bias compensation of the signal conversion module.

[0047] As an optional solution of the present invention, it also includes an oscillator module for generating sine wave, sawtooth wave and square wave signals to provide a reference signal;

[0048] As an optional solution of the present invention, a constant current source module is included to provide a stable bidirectional current to ensure the current stability of signal processing;

[0049] The present invention provides a conditioning method for a high-precision signal conditioning system, comprising the following steps:

[0050] S1. Signal acquisition and conversion: The input signal is converted through the signal conversion module, including current frequency I / F conversion, voltage frequency V / F conversion or capacitance voltage C / V conversion;

[0051] S2, signal preprocessing: input the converted signal into the filtering processing module to remove high-frequency noise or interference signals;

[0052] S3, signal amplification and isolation: use the isolation amplifier module to electrically isolate and amplify the pre-processed signal to enhance signal strength and prevent interference;

[0053] S4, feature extraction and denoising: Use the VAE module to extract the potential features of the signal, generate the potential distribution parameters of the signal through the encoder; generate feature representation through latent space sampling, and reconstruct the signal through the decoder to filter out noise;

[0054] S5, Time Series Modeling and Prediction: The feature representation generated by the VAE module is input into the LSTM module, and the hidden state is used to predict the future value of the signal based on the dynamic modeling formula of the time series;

[0055] S6, abnormal detection and feedback optimization: determine the abnormal signal according to the deviation between the predicted value and the actual value; adjust the filter frequency range or the zero bias compensation of the signal conversion module according to the feedback result;

[0056] S7, signal output: output optimized signal for external devices to use.

[0057] 3. Beneficial effects:

[0058] The present invention combines the variational autoencoder (VAE) with the long short-term memory network (LSTM) to construct a high-precision signal conditioning system that can effectively solve the problems of long-term feature capture, noise suppression and anomaly detection in complex dynamic signal processing. The system has significant technical effects in the following aspects:

[0059] 1. The ability to capture long-term features is significantly improved: In the LSTM module of the present invention, the cell state is used as the core storage unit of long-term memory, which can dynamically record important information in time series signals. This ensures the long-term retention of key features in the signal, can effectively capture the deep patterns of the time series, and provides a stable and reliable foundation for anomaly detection and signal prediction.

[0060] 2. Enhanced robustness of dynamic signal processing: The present invention combines the feature extraction capability of the VAE module. The VAE maps the input signal to the latent feature space through the encoder and decoder and reconstructs it. The reconstruction process of the VAE retains the key features of the signal while removing complex noise. The LSTM module uses the latent features extracted by the VAE and combines historical state information for time series modeling and prediction. It can maintain high-precision output in complex and noisy environments, and is particularly suitable for scenes with obvious dynamic characteristics such as industrial vibration signals, communication signals, and medical signals.

[0061] 3. Real-time anomaly detection can be achieved: the system can quickly identify signal anomalies, significantly reduce false alarm rates and missed alarm rates, and provide reliable support for equipment status monitoring and fault warning.

[0062] 4. Improved intelligence level of signal processing: The present invention realizes dynamic optimization of signal conditioning through feedback mechanism: adjusting the frequency range of the filter according to the prediction results of LSTM; adjusting the zero bias compensation of the signal conversion module based on the signal characteristics reconstructed by VAE. The filtering and signal conversion modules of the system can adapt to the dynamic environment to ensure the quality of signal output. The system can automatically adjust parameters according to the real-time changes of the signal environment, improving the intelligence and stability of signal processing.

[0063] 5. Improve the overall technical effect of the system: The system can be applied to various scenarios such as industrial monitoring, communication systems, medical signal processing and navigation systems, providing efficient solutions for complex dynamic signal processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a structural block diagram of the present invention;

[0065] Figure 2 It is a flow chart of the high-precision signal conditioning method of the present invention. DETAILED DESCRIPTION

[0066] Example 1: Reference Figure 1 The present invention provides a high-precision signal conditioning system, comprising:

[0067] Signal conversion module: including current frequency conversion circuit, voltage frequency conversion circuit and capacitor voltage conversion circuit; responsible for converting signals of different forms into a standard form that is convenient for subsequent processing.

[0068] Filter processing module: including low-pass filter and band-pass filter; used to remove noise and interference in the signal and improve the quality and stability of the signal.

[0069] Isolation amplifier module: used to isolate and amplify the signal to improve the safety, stability and anti-interference ability of the system. In many practical applications, the signal source and the subsequent processing circuit may be at different potentials, or there may be large electrical interference. The isolation amplifier module uses isolation technology, such as photoelectric isolation and magnetic isolation, to electrically isolate the input signal from the output signal, effectively preventing mutual interference between signals and the propagation of electrical faults. At the same time, the module amplifies the signal to meet the requirements of the subsequent processing circuit for the signal amplitude. The isolation amplifier module consists of an input circuit, an isolation unit and an output circuit. The input circuit is responsible for preprocessing the input signal, such as impedance matching and filtering; the isolation unit uses isolation devices such as photoelectric couplers and transformers to convert the input signal into an optical signal or a magnetic signal for transmission, and then converts it back into an electrical signal; the output circuit amplifies and adjusts the isolated signal to output a signal that meets the requirements.

[0070] VAE-LSTM joint algorithm module: includes: VAE module and LSTM module; VAE module is used to extract the potential features of the signal and filter out noise; LSTM module is used to perform time series modeling and prediction on the potential features extracted by the VAE module. The VAE-LSTM joint algorithm module combines the advantages of variational autoencoder (VAE) and long short-term memory network (LSTM), and can perform deep feature extraction, noise filtering, and time series modeling and prediction on the signal.

[0071] In this technical solution, the high-precision signal conditioning system can achieve high-precision processing, analysis and prediction of various complex signals through the collaborative work of the signal conversion module, filtering processing module, isolation amplification module and VAE-LSTM joint algorithm module, providing strong technical support for applications in many fields.

[0072] Furthermore, the core function of the current-frequency conversion circuit is to convert the input current signal into a frequency signal. In practical applications, many sensors output current signals, such as photoelectric sensors, pressure sensors, etc. In order to facilitate the processing and transmission of digital systems, these current signals need to be converted into frequency signals. Its working principle is based on the combination of current-voltage conversion and voltage-frequency conversion. First, the input current is converted into a voltage signal through a precision resistor, and then the voltage signal is converted into a frequency signal using a special voltage-frequency conversion chip. This conversion method has the characteristics of high precision, high stability and strong anti-interference ability, and can accurately convert small current changes into recognizable frequency changes.

[0073] The voltage-to-frequency conversion circuit is mainly used to convert the input voltage signal into a frequency signal. In some analog signal acquisition systems, the voltage signal output by the sensor may have a wide range and high accuracy requirements. This circuit accurately converts the input voltage signal into a frequency signal through a specific circuit structure and algorithm. The conversion process is based on the principle of integrator and comparator. By integrating the input voltage, a pulse signal is generated when the integral value reaches a certain threshold, thereby realizing the conversion from voltage to frequency. This conversion method can not only improve the anti-interference ability of the signal, but also facilitate the subsequent digital signal processing and transmission.

[0074] The capacitance-to-voltage conversion circuit is used to convert the capacitance change output by the capacitance sensor into a voltage signal. Capacitance sensors are often used in the measurement of some physical quantities, such as displacement, pressure, humidity, etc. These sensors reflect the changes in the measured quantity through changes in capacitance. The capacitance-to-voltage conversion circuit converts small changes in capacitance into measurable voltage signals through clever circuit design. Common implementation methods include bridge circuits and charge amplifier circuits. The bridge circuit uses the change in capacitance to change the balance state of the bridge, thereby outputting a voltage signal related to the capacitance change; the charge amplifier circuit converts the capacitance change into a voltage change by amplifying the charge on the capacitor.

[0075] Furthermore, the function of the low-pass filter is to allow low-frequency signals to pass through while suppressing high-frequency signals. In the actual signal acquisition process, high-frequency noise is often mixed into the signal due to factors such as environmental noise and power supply interference. The low-pass filter filters out these high-frequency noises through its specific frequency response characteristics, retaining only low-frequency useful signals. Common types of low-pass filters include Butterworth filters, Chebyshev filters, and elliptical filters. The Butterworth filter has a flat passband and a gradually decreasing stopband characteristic, which is suitable for occasions where the signal amplitude in the passband is required to be relatively flat; the Chebyshev filter has an equal ripple characteristic in the passband or stopband, which can improve the selectivity of the filter to a certain extent; the elliptical filter has the narrowest transition band and the smallest stopband attenuation, which is suitable for occasions where high filter performance is required.

[0076] The function of a bandpass filter is to allow only signals within a specific frequency range to pass through, while suppressing signals of other frequencies. In some applications, it is necessary to extract signals within a specific frequency range, such as carrier signals in communication systems, specific frequency components in biomedical signals, etc. By designing the appropriate center frequency and bandwidth, the bandpass filter can accurately extract the required signal and effectively suppress interference from other frequencies. The bandpass filter can be composed of a combination of a low-pass filter and a high-pass filter, or a special bandpass filter design method can be used, such as a twin-T network, a Wien bridge circuit, etc.

[0077] Furthermore, the main function of the VAE module (Variational Autoencoder) is to extract the latent features of the signal and filter out the noise. The variational autoencoder is a deep learning architecture based on a probabilistic model. It learns the latent feature representation of the signal by mapping the input signal into a low-dimensional latent space. In this process, the VAE module can automatically model and filter out the noise in the signal to obtain a purer feature representation. The VAE module consists of two parts: an encoder and a decoder. The encoder maps the input signal to the mean and variance in the latent space, and generates a latent vector from the latent space by random sampling; the decoder decodes this latent vector into a reconstructed signal. By minimizing the error between the reconstructed signal and the original signal and the difference between the distribution of the latent vector and the prior distribution, VAE can learn the effective feature representation of the signal and filter out the noise to a certain extent.

[0078] The LSTM (Long Short-Term Memory) module is used to perform time series modeling and prediction on the potential features extracted by the VAE module. The LSTM is a recurrent neural network specifically designed for processing time series data. It can effectively solve the gradient vanishing and gradient exploding problems in traditional recurrent neural networks, thereby better capturing the long-term dependencies in time series data. In the high-precision signal conditioning system, the LSTM module receives the potential feature sequence output by the VAE module, and learns and models these features through its internal memory units and gating mechanisms. LSTM predicts future signal trends based on historical feature information, providing strong support for further processing and analysis of the signal. For example, in industrial process monitoring, LSTM can predict the operating status of equipment based on historical signal data and detect potential fault hazards in advance.

[0079] Embodiment 2:

[0080] The difference between this embodiment and embodiment 1 is that the VAE module generates potential distribution parameters of the signal through the encoder, and the mean and logarithmic variance are respectively:

[0081] μ=f μ (x;θe ); logσ 2 =f σ2 (x;θ e ), where μ and logσ 2 are the mean and log variance of the underlying distribution, θ e is the trainable parameter of the encoder. x represents the signal input to the VAE module encoder. This signal can be various types of data, such as time series signals, image data, etc. It is the starting input of the entire processing flow. The encoder will generate the corresponding potential distribution parameters based on this input signal. u (◆) is a function determined by the encoder, which takes the input signal x and the encoder’s trainable parameters θ e As input, it outputs the mean u of the potential distribution. The specific form of this function depends on the network structure and design of the encoder. Its function is to map the input signal x to the mean position of the latent space through a specific calculation method. σ2 (◆) is a function determined by the encoder, which takes the input signal x and the trainable parameters θ of the encoder e As input, it outputs the logarithmic variance of the potential distribution. Its specific implementation depends on the architecture of the encoder and is used to calculate the logarithmic variance of the distribution corresponding to the input signal x in the latent space.

[0082] Preferably, the VAE module computes the latent feature z by sampling the latent space:

[0083]

[0084] Where ∈ is the standard normal distribution noise of random sampling. z represents the potential feature calculated by sampling in the latent space. It is obtained by adding the mean u to the product of the standard deviation σ and the random noise ∈. As the key intermediate variable connecting the encoder and the decoder in the VAE model, z contains the feature information of the input data. The decoder will reconstruct data similar to the original input based on z. N(0,I) represents the multivariate standard normal distribution. Where 0 represents the mean vector, indicating that the mean of the distribution is 0 in all dimensions; I is the unit matrix, representing the covariance matrix, which means that each dimension is independent of each other and the variance is 1. In practical applications, if the latent space is multidimensional, then ε is sampled from this multivariate standard normal distribution, and z will also take values ​​in the multidimensional latent space accordingly.

[0085] Preferably, the LSTM module updates the time series state of the signal by the following steps:

[0086] Forget gate: f t =σ(W f ·[h t-1 ,z t ]+bf );

[0087] Input gate: i t =σ(W i ·[h t-1 ,z t ]+b i ),

[0088] Cell Status:

[0089] Output gate: o t =σ(W o ·[h t-1 ,z t ]+b o ),h t =o t tanh(C t ), where t represents the time step, which is used to identify different moments in the time series. When processing time series data, the data is input into the LSTM module in sequence, and each time step has corresponding input, state update and other operations. t-1 Represents the hidden state of the LSTM module at time step t-1. The hidden state carries the information of the previous time step and will participate in various calculations of the current time step to help the model remember past information. t-1 is the cell state of the LSTM module at time step t-1. The cell state is an important carrier for long-term memory information in LSTM. It transfers information between different time steps and retains key information during the update process. f is the input x t to the weight matrix of the forget gate, which determines the input x t The degree of influence on the output of the forget gate. b f is the bias vector of the forget gate, which is used to provide a learnable constant term for the calculation of the forget gate to help the model better fit the data. i is the input x t The weight matrix to the input gate controls the input x t Effect on the input gate output. b i is the bias vector of the input gate, adding a learnable constant offset to the calculation of the input gate. c is the input x t The weight matrix to the cell state update part affects the contribution of the input to the cell state update. c It is the bias vector of the cell state update part, which provides a constant term for the calculation of cell state update. o is the input x t The weight matrix to the output gate controls the input x tImpact on the output of the output gate. b o is the bias vector of the output gate, providing an adjustable constant term for the calculation of the output gate. t is the output of the forget gate at time step t. The forget gate determines the cell state c at the previous moment t-1 Which information needs to be retained and which information needs to be forgotten, the value range is between [0-1], calculated by the Sigmoid function. The closer to 1, the more information is retained, and the closer to 0, the more information is forgotten. t is the output of the input gate at time step t. The input gate controls the current input x t Which information in the cell state needs to be added? It is calculated by the Sigmoid function, with a value range of 0-1, which is used to determine the degree of retention of the input information. ～ t is the candidate cell state generated at time step t. It is based on the current input x t and the previous hidden state h t-1 The calculated value is transformed by a tanh function, and its value range is between [-1,1], which contains the new information of the current input. t is the cell state after the time step t is updated. It is calculated by the forget gate output, the cell state at the previous moment, the input gate output and the candidate cell state. It inherits part of the cell state information at the previous moment and adds the new information of the current input. It is the main carrier of long-term memory in the LSTM module. t The output of the output gate at time step t. The output gate determines the current cell state c t Which information needs to be output as the hidden state of the current time step. It is calculated by the Sigmoid function, with a value range of [0-1], which is used to control the proportion of output information. t is the hidden state of the LSTM module at time step t. It is gated by the output gate o t and the cell state c after tanh transformation t The calculated information contains useful information of the current time step, which will be used as one of the inputs of the next time step and can also be used as the output of the model for subsequent tasks.

[0090] Preferably, the LSTM module uses the hidden state h t Predict the output value of the signal Where t represents the time step. t is the hidden state of the LSTM module at time step t. y is the weight matrix of the prediction layer, which is a trainable parameter of the prediction layer. Its function is to convert the hidden state h of the LSTM module at time step tt Perform linear transformation by matrix multiplication W y h t To adjust the weights of each feature in the hidden state, so that the model can better capture the relationship between the hidden state and the final output value, and thus make a reasonable prediction of the output value. y is the bias vector of the prediction layer and is a trainable parameter of the prediction layer. y It provides a learnable constant offset for the prediction result. y For the hidden state h t After linear transformation, add the bias vector b y , which enables the model to more flexibly adjust the prediction results under different data distributions and characteristics, and helps to improve the prediction accuracy and generalization ability of the model.

[0091] Preferably, the potential space dimension z of the VAE module is compressed to 10%-30% relative to the dimension of the input signal, and the potential space dimension z is calculated by the following formula:

[0092]

[0093] Preferably, the LSTM module uses sliding window technology to dynamically model the time series signal, the window length is Tw, and each update is based on the following formula:

[0094] Where Tw is the length of the sliding window, which determines the number of time series data involved in LSTM module modeling each time. t is the time step index, which is used to traverse each time point in the sliding window. t represents the actual value at time step t,

[0095] ^

[0096] That is, the real observation data of the time series signal at that moment. t Represents the predicted output at time step t, which is the predicted value made by the LSTM module for the current time step t based on the historical data in the sliding window. LSTM is the loss function value based on the LSTM model. ||·|| represents a certain norm

[0097] ^

[0098] operation, here used to measure the actual value y t and the predicted value y t The distance or difference between.

[0099] Preferably, the LSTM module uses a threshold to judge abnormal signals based on the deviation between the prediction error and the actual signal. The deviation calculation formula is:

[0100] If δ t >δ 阈值 , then the abnormal signal is determined. In the formula, δ t Represents the deviation between the prediction error and the actual signal at time step t. 阈值 Indicates a preset fixed value, which serves as the standard limit for judging whether a signal is abnormal.

[0101] Preferably, the VAE-LSTM joint algorithm module dynamically adjusts the parameters of the filter module through the following feedback adjustment mechanism:

[0102] Signal prediction results based on LSTM Optimizing the filter frequency range: During signal processing, the frequency range of the filter directly affects the filtering effect of the signal. Signal components of different frequencies may contain different information, and there may also be noise interference. The prediction results of LSTM not only include estimates of future signal values, but also imply the frequency characteristics of the signal. The process of optimizing the filter frequency range is a dynamic and adaptive process. The system first analyzes the possible frequency components and interference in the signal based on the prediction results of LSTM. Then, through specific algorithms and control logic, the parameters of the filter are adjusted to achieve the optimization of the frequency range. This process involves adjusting parameters such as the type of filter (such as low-pass filter, high-pass filter, band-pass filter, etc.), order, and cutoff frequency. For example, if the prediction results show that there is interference in a specific frequency range in the signal, the system may select a band-pass filter of a suitable order and accurately adjust its cutoff frequency so that it can effectively filter out the interference in this frequency range while retaining the useful components of the signal.

[0103] Adjusting the zero-bias compensation of the signal conversion module based on the VAE reconstructed signal: During the signal conversion process, due to the influence of factors such as the non-ideal characteristics of circuit elements and environmental noise, the signal may have a zero-bias error, that is, the output signal has a fixed deviation relative to the input signal. This zero-bias error will affect the accuracy and reliability of the signal, especially in some applications with high signal accuracy requirements, such as sensor signal acquisition and precision measurement. The signal reconstructed by VAE can be used as a reference to detect and correct this zero-bias error. By comparing the original signal and the reconstructed signal, the system can calculate the magnitude and direction of the zero-bias error. If the reconstructed signal shows that there is a zero-bias error, the system will adjust the zero-bias compensation parameters of the signal conversion module based on this error information. The method of adjusting the zero-bias compensation involves adjusting the hardware parameters or software algorithms in the signal conversion module. In terms of hardware, it may be necessary to adjust the gain, bias voltage and other parameters of the amplifier to eliminate the zero-bias error. In terms of software, it may be necessary to use digital signal processing algorithms, such as adaptive filtering and calibration algorithms, to correct the signal. For example, in a sensor signal acquisition system, the analog signal output by the sensor may have a zero-bias error after analog-to-digital conversion (ADC). The system can reconstruct the converted digital signal through VAE, and then adjust the ADC bias voltage or digital correction coefficient according to the difference between the reconstructed signal and the original signal, thereby compensating for the zero bias error and improving the accuracy of signal acquisition.

[0104] Preferably, an oscillator module is also included to generate sine wave, sawtooth wave and square wave signals to provide a rich variety of reference signals for the system.

[0105] Preferably, a constant current source module is also included to provide a stable bidirectional current to ensure the current stability of signal processing and meet the requirements of different signal processing circuits.

[0106] Embodiment 3:

[0107] Reference Figure 2 The present invention provides a high-precision signal conditioning method, comprising the following steps:

[0108] S1. Signal acquisition and conversion: The input signal is converted through the signal conversion module, including current frequency I / F conversion, voltage frequency V / F conversion or capacitance voltage C / V conversion;

[0109] S2, signal preprocessing: input the converted signal into the filtering processing module to remove high-frequency noise or interference signals;

[0110] S3, signal amplification and isolation: use the isolation amplifier module to electrically isolate and amplify the pre-processed signal to enhance signal strength and prevent interference;

[0111] S4, feature extraction and denoising: Use the VAE module to extract the potential features of the signal, generate the potential distribution parameters μ and logσ2 of the signal through the encoder; generate the feature representation z = μ + σ*∈ through latent space sampling, and reconstruct the signal through the decoder to filter out the noise;

[0112] S5. Time series modeling and prediction: The feature representation z generated by the VAE module is input into the LSTM module, and the dynamic modeling formula based on the time series is: h t =σ(W o ·[h t-1 ,z t ]+b o )·tang(C t ) Using the hidden state h t Predict future values ​​of a signal

[0113] S6. Anomaly detection and feedback optimization: Based on the predicted value and the actual value y t Deviation If δ t >δ 阈值 , then determine the abnormal signal; adjust the filter frequency range or the zero bias compensation of the signal conversion module according to the feedback result;

[0114] S7, signal output: output optimized signal for external devices to use.

[0115] Example 4: High-precision conditioning and anomaly detection of industrial vibration signals. Background: Key equipment in a factory generates vibration signals during operation. Equipment failures are usually manifested as abnormal changes in the frequency and amplitude of the vibration signals. In order to detect the operating status of the equipment and predict potential failures, it is necessary to perform high-precision conditioning, noise reduction, modeling and anomaly detection on the vibration signals.

[0116] Input signal: The collected vibration signal is as follows:

[0117] x=[0.3,0.28,0.35,0.4,0.42,0.41,0.38,0.6,0.95,1.2,0.85];

[0118] The signal is mixed with high-frequency noise and periodic interference.

[0119] The processing process includes:

[0120] 1. Signal acquisition and conversion (signal conversion module): The input signal is converted into current frequency (I / F) through the signal conversion module. The converted signal is standardized to adapt to subsequent processing.

[0121] 2. Signal preprocessing: Use a low-pass filter (cut-off frequency is 200Hz) through the filtering module to remove high-frequency noise. The filtered signal is:

[0122] x filtered =[0.3,0.28,0.35,0.4,0.42,0.41,0.38,0.6,0.8,0.9,0.75];

[0123] 3. Signal amplification and isolation (isolation amplifier module): The filtered signal is isolated and gain adjusted through the isolation amplifier module to enhance signal strength while reducing external interference.

[0124] 4. Feature extraction and denoising (VAE module): The encoder generates the potential distribution parameters and uses the encoder to calculate the mean and logarithmic variance of the latent space:

[0125] μ=f μ (x filtered ) = [0.31, 0.29, 0.36, 0.41, 0.43, 0.42, 0.39, 0.61, 0.81, 0.91, 0.76];

[0126]

[0127] Latent space sampling: According to the formula z = μ + σ*∈, where Sampling results in:

[0128] z=[0.31,0.29,0.36,0.41,0.43,0.42,0.39,0.61,0.81,0.91,0.76];

[0129] Decoder reconstructs the signal: The decoder reconstructs the signal based on the latent feature z:

[0130] Through reconstruction, the signal noise has been significantly reduced.

[0131] 5. Dynamic modeling and prediction (LSTM module): Input potential features. Input the potential feature z extracted by VAE into the LSTM module. LSTM updates the state according to the following formula:

[0132] Forget gate: f t =σ(W f ·[h t-1 ,z t ]+b f );

[0133] Input gate: i t =σ(W i ·[h t-1 ,zt ]+b i ),

[0134] Cell Status:

[0135] Output gate: o t =σ(W o ·[h t-1 ,z t ]+b o ),h t =o t tanh(C t );

[0136] Future signal prediction: using hidden state h t Predicting future signal values

[0137] The predicted future signals are:

[0138]

[0139] 6. Anomaly detection and feedback optimization: Based on the predicted value and the actual value y t Deviation Determine whether it is abnormal:

[0140] δ t =[0.01,0.01,0.01,0.06,0.07,0.08,0.09,0.05,0.03,0.02,0.03]; Set threshold δ 阈值 =0.2. The results show:

[0141] δ t <δ 阈值 , no abnormal signal; optimize the filter frequency range according to the feedback to improve the signal processing accuracy.

[0142] 7. Signal output: According to the prediction deviation of LSTM, adjust the parameters of the filter (increase the cutoff frequency to 220Hz), and output the final optimized signal:

[0143] x output =[0.31,0.29,0.36,0.41,0.43,0.42,0.39,0.61,0.81,0.91,0.76].

[0144] The technical effects achieved by the high-precision signal conditioning system of the present invention are: 1. Improved noise reduction capability: The VAE module significantly reduces noise interference and retains the key features of the signal. 2. The time series modeling LSTM accurately captures the timing characteristics of the signal, and the prediction results are highly accurate. 3. The anomaly detection results are accurate, without false positives or false negatives. 4. Adaptive optimization, the system adjusts the filtering parameters in real time to improve the signal output quality.

[0145] This example demonstrates the excellent performance of the present invention in complex dynamic signal conditioning and provides an efficient solution for operation monitoring and fault warning of industrial equipment.

[0146] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A high-precision signal conditioning system, comprising a signal conversion module, a filtering processing module, an isolation amplification module and a VAE-LSTM joint algorithm module, characterized in that: Signal conversion module: including current frequency conversion circuit, voltage frequency conversion circuit and capacitance voltage conversion circuit; converts signals of different forms into a standard form that is convenient for subsequent processing; Filter processing module: including low-pass filter and band-pass filter; remove noise and interference in the signal to improve the quality and stability of the signal; Isolation and amplification module: isolate and amplify the signal; improve the safety, stability and anti-interference ability of the system; VAE-LSTM joint algorithm module: includes VAE module and LSTM module; VAE module is used to extract potential features of signals and perform noise filtering; LSTM module is used to perform time series modeling and prediction on potential features extracted by VAE module, integrating the advantages of variational autoencoder VAE and long short-term memory network LSTM, and performing deep feature extraction, noise filtering, and time series modeling and prediction on signals.

2. The high-precision signal conditioning system according to claim 1, characterized in that: The VAE module generates the potential distribution parameters of the signal through the encoder, and the mean and log variance are: μ=f μ (x;θ e ), Among them, μ and logσ 2 are the mean and log variance of the underlying distribution, θ e is the trainable parameter of the encoder; x represents the signal input to the encoder of the VAE module; f u (.) is a function determined by the encoder, with the input signal x and the trainable parameters θ of the encoder e As input, output the mean u of the underlying distribution; f σ2 (.) is a function determined by the encoder, with the input signal x and the trainable parameters θ of the encoder e As input, it outputs the log variance of the underlying distribution.

3. The high-precision signal conditioning system according to claim 2, characterized in that: The VAE module calculates the latent feature z by sampling the latent space: Where, ∈ is the standard normal distribution noise of random sampling; σ=exp(0.5·logσ 2 ); z represents the potential features calculated by sampling in the latent space; N(0, I) represents the multivariate standard normal distribution; where 0 represents the mean vector, indicating that the mean of the distribution is 0 in all dimensions; I is the unit matrix, representing the covariance matrix, which means that each dimension is independent of each other and the variance is 1.

4. The high-precision signal conditioning system according to claim 1, characterized in that: The LSTM module updates the time series state of the signal by following the steps below: Forget gate: f t =σ(W f ·[h t-1 , z t ]+b f );where f t is the output of the forget gate; W f is the weight matrix of the forget gate; b f is the bias vector of the forget gate; h t-1 is the hidden state of the previous time step; z t is the input feature of the current time step; is the sigmoid activation function; t represents the time step; Input gate: i t =σ(W i ·[h t-1 , z t ]+b i ) In the formula, i t is the output of the input gate; is the candidate value of the cell state at the current time step; W i , W c are the weight matrices of the input gate and cell state candidate values ​​respectively; b i , b c are the bias vectors for the candidate values ​​of the input gate and cell state, respectively; is the hyperbolic tangent activation function, with a value range of [-1, 1]; Cell Status: In the formula, C t is the cell state at the current time step; C t-1 is the cell state at the previous time step; Output gate: o t =σ(W o ·[h t-1 , z t ]+b o ), h t =o t tanh(C t );where o t is the output of the output gate; h t is the hidden state of the current time step; b o is the bias vector for the output gate.

5. The high-precision signal conditioning system according to claim 4, characterized in that: The LSTM module uses the hidden state h t Predict the output value of the signal Among them, W y is the weight matrix of the prediction layer, b y is the bias parameter; t represents the time step; h t is the hidden state of the LSTM module at time step t.

6. The high-precision signal conditioning system according to claim 2, characterized in that: The potential space dimension z of the VAE module is compressed to 10%-30% relative to the dimension of the input signal. The potential space dimension z is calculated by the following formula:

7. The high-precision signal conditioning system according to claim 1, characterized in that: The LSTM module uses sliding window technology to dynamically model the time series signal, and the window length is T w , each update is based on the following formula: in, is the predicted output, y t is the actual value; t is the time step index; L LSTM is the loss function value based on the LSTM model, and · represents the norm operation.

8. The high-precision signal conditioning system according to claim 1, characterized in that: The LSTM module uses a threshold to judge abnormal signals based on the deviation between the prediction error and the actual signal. The deviation calculation formula is: If δ t >δ 阈值 , then the abnormal signal is determined; where δ t Indicates the deviation between the prediction error and the actual signal at time step t, δ 阈值 Indicates a preset fixed value, which serves as the standard limit for judging whether a signal is abnormal.

9. The high-precision signal conditioning system according to claim 1, characterized in that: The VAE-LSTM joint algorithm module dynamically adjusts the parameters of the filter module through the following feedback adjustment mechanism: Optimize the filter frequency range based on the signal prediction results of LSTM; Based on the VAE reconstructed signal, adjust the zero bias compensation of the signal conversion module.

10. A high-precision signal conditioning method based on the high-precision signal conditioning system according to any one of claims 1 to 9, characterized in that: The following steps are involved: S1. Signal acquisition and conversion: The input signal is converted through the signal conversion module, including current frequency I / F conversion, voltage frequency V / F conversion or capacitance voltage C / V conversion; S2, signal preprocessing: input the converted signal into the filtering processing module to remove high-frequency noise or interference signals; S3, signal amplification and isolation: use the isolation amplifier module to electrically isolate and amplify the pre-processed signal to enhance signal strength and prevent interference; S4, feature extraction and denoising: Use the VAE module to extract the potential features of the signal, generate the potential distribution parameters of the signal through the encoder; generate feature representation through latent space sampling, and reconstruct the signal through the decoder to filter out noise; S5, Time Series Modeling and Prediction: The feature representation generated by the VAE module is input into the LSTM module, and the hidden state is used to predict the future value of the signal based on the dynamic modeling formula of the time series; S6, Abnormal detection and feedback optimization: Determine abnormal signals based on the deviation between the predicted value and the actual value; adjust the filter frequency range or zero bias compensation of the signal conversion module based on the feedback results; S7, signal output: output optimized signal for external devices to use.

Images