Voltage monitoring system and method based on a multi-channel successive approximation analog-to-digital converter

Abstract

The application discloses a voltage monitoring system based on a multi-channel successive approximation type analog-digital converter, which comprises RC integrators connected to each analog input end and used for integrating and processing input analog voltage signals; LVDS comparators connected with corresponding RC integrators and used for converting output voltages of the RC integrators into working voltage ranges of the successive approximation type analog-digital converter; successive approximation type analog-digital converters connected with the LVDS comparators and used for analog-digital converting analog signals output by the LVDS comparators and outputting digital signals; and CLB-DSP cooperative processing units connected with the successive approximation type analog-digital converters and used for calibrating the digital signals, generating PWM duty cycles corresponding to target voltage values and sending the PWM duty cycles to a PWM generator. The application can improve the defects of the prior art and improve the precision of analog-digital conversion.

Description

Technical Field

[0002] This invention relates to the field of integrated circuit application technology, and in particular to a voltage monitoring system and method based on a multi-channel successive approximation analog-to-digital converter. Background Technology

[0004] In systems such as servers, personal computers, and industrial control, it is necessary to monitor multiple power supply voltages (such as 12V, 5V, 3.3V, 2.5V, 1.8V, 1.2V, 0.8V, 0.6V, etc.) in real time to ensure stable system operation. Traditional solutions typically use dedicated analog-to-digital converter (ADC) chips, such as 8-channel 12-bit ADC chips. These are costly, increase PCB layout complexity, require a dedicated controller for reading, and have low integration.

[0005] In recent years, some solutions have utilized the LVDS differential input buffer inside FPGAs / CPLDs as analog comparators, combined with external RC circuits to construct simple ADCs. For example, Xilinx's patent US6246258 discloses a method for implementing an ADC using a comparator in an interface unit on a digitally programmable integrated circuit, which can construct a Flash ADC or a SAR ADC. Altera's patent (US7321327) employs a Σ-Δ architecture. However, these solutions typically use external resistor strings or PWM+RC methods, resulting in poor conversion accuracy and high implementation costs for multi-voltage adaptation. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a voltage monitoring system and method based on a multi-channel successive approximation analog-to-digital converter, which can overcome the shortcomings of the prior art and improve the accuracy of analog-to-digital conversion.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows.

[0009] A voltage monitoring system based on a multi-channel successive approximation analog-to-digital converter includes, An RC integrator is connected to each analog input terminal to perform integration processing on the input analog voltage signal; An LVDS comparator, connected to a corresponding RC integrator, is used to convert the output voltage of the RC integrator into the operating voltage range of a successive approximation analog-to-digital converter. A successive approximation analog-to-digital converter, connected to an LVDS comparator, is used to convert the analog signal output from the LVDS comparator into a digital signal. The CLB-DSP coprocessing unit, connected to a successive approximation analog-to-digital converter, is used to calibrate digital signals, generate the PWM duty cycle corresponding to the target voltage value, and send it to the PWM generator.

[0010] A monitoring method for the voltage monitoring system based on a multi-channel successive approximation analog-to-digital converter, as described above, includes the following steps: The RC integrator integrates the analog signal and then inputs it into the LVDS comparator, which converts it into the operating voltage range of the successive approximation analog-to-digital converter (ADC), and then converts it into a digital signal through the successive approximation ADC. Outlier detection and removal are performed on digital signals, and the remaining normal data is standardized. The standardized digital signal is subjected to global coarse correction to obtain the optimal gain correction coefficient and the optimal zero-point correction coefficient, and coarse correction data is generated. Local fine correction is performed on the coarse correction data to predict the correction residuals and generate fine correction data; the correction residuals are then fed back to the global coarse correction as new correction targets to update the correction parameters and obtain the final correction data. After denormalization, the PWM duty cycle corresponding to the target voltage value is generated, and the PWM generator is controlled to output the corresponding voltage signal.

[0011] As a preferred method, outlier detection and removal of digital signals includes the following steps: Let the original output digital signal be ,in This represents the original digital signal of the nth sample, the tth time step, and the cth channel. A sliding window is set on the timing of the signal output from each channel, and the mean and standard deviation of the data within the sliding window are calculated. , , Where W is the length of the sliding window, when When this occurs, it is marked as an outlier. Otherwise set ,Will Construct an outlier mask .

[0012] As a preferred method, the data standardization method is , and These are all normal numbers of the c-th channel. The mean and standard deviation of the data.

[0013] Preferably, global coarse correction of digital signals includes the following steps: Let the objective function be ,in Initialize the gray wolf population; For each gray wolf Calculate the objective function Sort the gray wolves by fitness value from smallest to largest, and obtain the positions of the top three gray wolves. The convergence factor of the g-th generation is The adaptive weights of the g-th generation are Update the location of the gray wolves: , in For random coefficients, As the swing factor, , A random number between [0, 1]; If a gray wolf's position exceeds the search space, it is pulled back to the boundary. Every 10 iterations, a chaotic local search is performed on the gray wolf with the highest fitness to generate a new position. If the fitness of the new position is higher than that of the original position, the position of the gray wolf is updated. The optimal solution obtained after iteration is the optimal gain correction coefficient. and zero-point correction parameters Substitute into the correction formula Coarse correction data were calculated. .

[0014] As a preferred method, the initial population is generated using a Logistic chaotic mapping. Mapping chaotic sequences to parameter search space Only P is generated Gray Wolf's starting position .

[0015] Preferably, performing local fine correction includes the following steps: An encoder is constructed, comprising a channel attention module and a feedforward neural network; the channel attention module is as follows: , in , Where h is the model dimension and h is the number of attention heads. Multi-scale channel dimension features of coarse-corrected data; The attention weight between channels is Multi-channel attention , , To output the weight matrix; Perform residual connections and layer normalization Feedforward neural networks are ,in , , , These are learnable parameters; The encoder output is Error correction via linear layer prediction Local fine correction data are .

[0016] As a preferred method, one-dimensional convolutional layers of different scales are used to extract multi-scale channel-dimensional features from the coarse-corrected data. , express One-dimensional convolution; Optimal global correction parameters Embedded into features as prior information , , It is an N×T matrix of all 1s.

[0017] As a preferred method, the encoder is trained using a multi-objective hybrid loss function. ,That Medium correction error loss ,aisle Consistency loss ,Lu Rod loss

[0018] , δ is the threshold for Huber loss. This is the loss weighting coefficient.

[0019] Preferably, the residual R predicted by the attention mechanism is used as the new correction target to update the objective function of the global coarse correction. The optimal global correction parameters obtained from global coarse correction. For the initial population At the center, a new global coarse correction is performed to obtain the final correction parameters. Final data verification Positive result is Then The corrected data is denormalized to restore it to its original scale. .

[0020] The beneficial effects of adopting the above technical solution are as follows: Addressing the shortcomings of existing technologies, this invention employs a fully digital correction architecture, combining global coarse correction and local fine correction to solve zero-point and gain drift problems, as well as nonlinear distortion issues, and achieves high-precision correction through bidirectional feedback. This invention can improve the performance of a low-cost, simple ADC from 8 bits to 12 bits, achieving industrial-grade high precision. Its static error and dynamic performance fully meet the accuracy requirements of most industrial applications. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of a specific embodiment of the present invention. Detailed Implementation

[0022] Reference Figure 1 This invention implements a voltage monitoring system based on FPGA. Each analog input of the voltage monitoring system is connected to an LVDS comparator via an RC integrator. The LVDS comparator converts the output voltage of the RC integrator into the operating voltage range of a successive approximation analog-to-digital converter (ADC). The output of the LVDS comparator is connected to the successive approximation ADC, which performs analog-to-digital conversion and outputs a digital signal. The output of the successive approximation ADC is connected to a CLB-DSP co-processing unit. The CLB-DSP co-processing unit calibrates the digital signal and generates a PWM duty cycle corresponding to the target voltage value. The CLB-DSP co-processing unit sends the generated PWM duty cycle to a PWM generator, which outputs a voltage signal.

[0023] To improve the accuracy of voltage conversion, this invention specifically designs a correction algorithm that runs on the CLB-DSP co-processing unit. After compiling and simulating this algorithm on a computer, it is written into the FPGA, and the CLB-DSP co-processing unit can then implement this correction process.

[0024] Let the original output digital signal be , This represents the original digital signal of the nth sample, the tth time step, and the cth channel.

[0025] A sliding window is set on the timing of the signal output from each channel, and the mean and standard deviation of the data within the sliding window are calculated. , , Where W is the length of the sliding window, when When this occurs, it is marked as an outlier. Otherwise set ,Will Construct an outlier mask .

[0026] Standardize the raw digital signal , and denoted as the mean and standard deviation of all normal data in the c-th channel, respectively.

[0027] Next, the Grey Wolf optimization algorithm is used to perform global coarse correction on the standardized digital signal. The original Grey Wolf algorithm employs a linear convergence strategy and lacks a dedicated local search mechanism. Its convergence result is highly sensitive to the initial population; multiple runs yield significantly different correction parameters, failing to guarantee consistency and repeatability. Furthermore, the initial population distribution in the original Grey Wolf algorithm relies entirely on a random number generator, failing to guarantee a uniform distribution across the entire parameter search space. The individual position update mechanism lacks a clear objective and is prone to getting trapped in local optima. Therefore, the original Grey Wolf optimization algorithm performs poorly on multi-channel signal data. In view of this, this invention makes the following series of improvements to the Grey Wolf optimization algorithm.

[0028] The goal of the global coarse correction in this invention is to find the optimal gain correction coefficient. and zero-point correction parameters This minimizes the error between the corrected data and the reference data (usually the median of all channels). Correction formula The formula is The weighted mean square error is used as For the objective function ,That middle .

[0029] The initial population is generated using a Logistic chaotic mapping. Mapping the chaotic sequence Shooting into the parameter search space Generate P gray wolves initial position .

[0030] For each gray wolf Calculate the objective function Adapt the gray wolf Sort the degree values ​​from smallest to largest and get the positions of the first three gray wolves. The gth generation of the collection Convergence factor is The adaptive weights of the g-th generation Heavy Update the location of the Grey Wolves: , in For random coefficients, As the swing factor, , It is a random number between [0,1].

[0031] If a gray wolf's position exceeds the search space, it is pulled back to the boundary. Every 10 iterations, a chaotic local search is performed on the gray wolf with the highest fitness to generate a new position. If the fitness of the new position is higher than that of the original position, the position of that gray wolf is updated. The optimal solution obtained after the iteration is the optimal gain correction coefficient. and zero-point correction parameters Substituting the values ​​into the correction formula, we obtain the coarse correction data. .

[0032] After completing the global coarse correction, we then perform local fine correction. Here, we use an attention mechanism for correction. However, the traditional attention mechanism calculates only the time dimension, while the global coarse correction mentioned above has already solved the overall drift problem in the time dimension. Therefore, it cannot be used in conjunction with the Grey Wolf optimization algorithm to complete local fine correction. Moreover, the traditional attention mechanism is a data-driven black-box model that can only learn features from the input data and cannot incorporate external prior information or posterior feedback, which prevents it from utilizing the results of the global coarse correction. Therefore, we have optimized and improved the attention mechanism.

[0033] One-dimensional convolutional layers of different scales are used to extract multi-scale channel-dimensional features from the coarse-corrected data. , This represents a one-dimensional convolution. The optimal global correction parameters are then used. Embedded into features as prior information , , It is an N×T matrix of all 1s.

[0034] An encoder is built that includes a channel attention module and a feedforward neural network. Channel attention module: , in , For model The dimension is h, where h represents the number of attention heads.

[0035] The attention weight between channels is Multi-channel attention

[0036] , , This is the output weight matrix.

[0037] Perform residual connections and layer normalization Feedforward neural networks are... ,in , , , These are learnable parameters.

[0038] The encoder output is Error correction is predicted using linear layers. Local fine correction data are .

[0039] The encoder is trained using a multi-objective hybrid loss function. ,That Medium correction error loss ,aisle Consistency loss Robust Sexual loss

[0040] , δ is the threshold for Huber's loss. This is the loss weighting coefficient.

[0041] Finally, the residual R predicted by the attention mechanism is used as the new correction target to update the objective function of the Grey Wolf algorithm. The optimal global correction parameters obtained from global coarse correction. Using the initial population center as the starting point, re-run the gray wolf optimization algorithm process described above to obtain the final correction parameters. The final data correction result is: Then, the corrected data will be denormalized to restore it to the original data scale. .

[0042] A comparative experiment was conducted using a pure hardware circuit composed of an FPGA and an external RC circuit (comparative example) and a circuit with the correction method of this invention written into the FPGA based on this circuit (exemplary example). The FPGA used was an Artix-7 XC7A35T, and the signal source was an Agilent 33220a. Performance parameters such as zero-point error, gain error, inter-channel standard deviation, and sample processing time were compared, and the results are as follows:

[0043] It is evident that this invention, while meeting the real-time requirement of 1kSPS (a sampling period much smaller than 1ms), significantly outperforms existing analog-to-digital converters in all major performance parameters.

[0044] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A voltage monitoring system based on a multi-channel successive approximation analog-to-digital converter, characterized in that: include, An RC integrator is connected to each analog input terminal to perform integration processing on the input analog voltage signal; An LVDS comparator, connected to a corresponding RC integrator, is used to convert the output voltage of the RC integrator into the operating voltage range of a successive approximation analog-to-digital converter. A successive approximation analog-to-digital converter, connected to an LVDS comparator, is used to convert the analog signal output from the LVDS comparator into a digital signal. The CLB-DSP coprocessing unit, connected to a successive approximation analog-to-digital converter, is used to calibrate digital signals, generate the PWM duty cycle corresponding to the target voltage value, and send it to the PWM generator.

2. A monitoring method for a voltage monitoring system based on a multi-channel successive approximation analog-to-digital converter as described in claim 1, characterized in that... Includes the following steps: The RC integrator integrates the analog signal and then inputs it into the LVDS comparator, which converts it into the operating voltage range of the successive approximation analog-to-digital converter (ADC), and then converts it into a digital signal through the successive approximation ADC. Outlier detection and removal are performed on digital signals, and the remaining normal data is standardized. The standardized digital signal is subjected to global coarse correction to obtain the optimal gain correction coefficient and the optimal zero-point correction coefficient, and coarse correction data is generated. Local fine correction is performed on the coarse correction data to predict the correction residuals and generate fine correction data; the correction residuals are then fed back to the global coarse correction as new correction targets to update the correction parameters and obtain the final correction data. After denormalization, the PWM duty cycle corresponding to the target voltage value is generated, and the PWM generator is controlled to output the corresponding voltage signal.

3. The monitoring method for a voltage monitoring system based on a multi-channel successive approximation analog-to-digital converter according to claim 2, characterized in that: Outlier detection and removal for digital signals Includes the following steps, Let the original output digital signal be ,in This represents the original digital signal of the nth sample, the tth time step, and the cth channel. A sliding window is set on the timing of the signal output from each channel, and the mean and standard deviation of the data within the sliding window are calculated. ， ， Where W is the length of the sliding window, when When this occurs, it is marked as an outlier. Otherwise set ,Will Construct an outlier mask .

4. The monitoring method for a voltage monitoring system based on a multi-channel successive approximation analog-to-digital converter according to claim 3, characterized in that: The method for standardizing data is , and denoted as the mean and standard deviation of all normal data in the c-th channel, respectively.

5. The monitoring method for a voltage monitoring system based on a multi-channel successive approximation analog-to-digital converter according to claim 2, characterized in that: Global coarse correction of digital signals includes the following steps: Let the objective function be ,in Initialize the gray wolf population; For each gray wolf Calculate the objective function Sort the gray wolves by fitness value from smallest to largest, and obtain the positions of the top three gray wolves. The convergence factor of the g-th generation is The adaptive weights of the g-th generation are Update the location of the gray wolves: ， in For random coefficients, As the swing factor, , A random number between [0, 1]; If a gray wolf's position exceeds the search space, it is pulled back to the boundary. Every 10 iterations, a chaotic local search is performed on the gray wolf with the highest fitness to generate a new position. If the fitness of the new position is higher than that of the original position, the position of that gray wolf is updated. The optimal solution obtained after the iteration is the optimal gain correction coefficient. and zero-point correction parameters Substitute into the correction formula Coarse correction data were calculated. .

6. The monitoring method for a voltage monitoring system based on a multi-channel successive approximation analog-to-digital converter according to claim 5, characterized in that: The initial population is generated using a Logistic chaotic mapping. Mapping chaotic sequences to parameter search space Generate the initial positions of P gray wolves. .

7. The monitoring method for a voltage monitoring system based on a multi-channel successive approximation analog-to-digital converter according to claim 5, characterized in that: Performing local fine correction includes the following steps: An encoder is constructed, comprising a channel attention module and a feedforward neural network; the channel attention module is as follows: ， in , Where h is the model dimension and h is the number of attention heads. Multi-scale channel dimension features of coarse-corrected data; The attention weight between channels is Multi-channel attention , , To output the weight matrix; Perform residual connections and layer normalization Feedforward neural networks are ,in , , , These are learnable parameters; The encoder output is Error correction via linear layer prediction Local fine correction data are .

8. The monitoring method for a voltage monitoring system based on a multi-channel successive approximation analog-to-digital converter according to claim 7, characterized in that: One-dimensional convolutional layers of different scales are used to extract multi-scale channel-dimensional features from the coarse-corrected data. , Represents one-dimensional convolution; the optimal global correction positive parameter Embedded into features as prior information , , It is an N×T matrix of all 1s.

9. The monitoring method for a voltage monitoring system based on a multi-channel successive approximation analog-to-digital converter according to claim 8, characterized in that: The encoder is trained using a multi-objective hybrid loss function. Among them, the correction error loss Channel consistency loss Robustness loss , δ is the threshold for Huber loss. This is the loss weighting coefficient.

10. The monitoring method for a voltage monitoring system based on a multi-channel successive approximation analog-to-digital converter according to claim 9, characterized in that: The residual R predicted by the attention mechanism is used as the new correction target to update the objective function of the global coarse correction. The optimal global correction parameters obtained from global coarse correction. Using the initial population as the center, a new global coarse correction is performed to obtain the final correction parameters. ; final The data correction result is Then The corrected data is denormalized to restore it to its original scale. .

Images