An electronic selvedge control method and system based on FPGA

Abstract

The present invention provides an FPGA-based electronic reel edge control method and system, which relate to the technical field of electronic reel edge control, including: collecting tension data and thickness data of reel edge fabric, speed data, torque data and temperature data of an operating motor, and preprocessing the collected data; performing feature extraction on the preprocessed data to obtain key features; quickly fusing the key features; constructing a state recognition model based on a deep learning algorithm, classifying and identifying the fused key features to obtain an operating mode and a fabric state; constructing an optimal control decision according to the operating mode and the fabric state; constructing a fault warning model according to the optimal control decision, obtaining a fault risk coefficient, and thereby accurately controlling the electronic reel edge control system; constructing a control optimization self-learning algorithm according to the fault risk coefficient, optimizing the fault warning model, and thereby ensuring the safe and stable operation of the electronic reel edge control system.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic selvedge control, in particular to an electronic selvedge control method and system based on FPGA. Background Art

[0002] Traditional electronic selvedge control systems mainly use Or general-purpose processor implementation, which leads to multi-sensor data processing lags: the frequency of multi-channel data acquisition for tension, thickness, and speed is limited, making it impossible to capture high-frequency fluctuations in real time; state recognition lacks real-time performance: traditional rule-based algorithms struggle to cope with complex working conditions, resulting in delayed detection of defects such as edge wear and thickness anomalies; control decision response is slow: the delay from data acquisition to control output can reach hundreds of milliseconds, which cannot meet the requirements of high-speed production scenarios; and system scalability is poor: traditional architectures make it difficult to integrate deep learning algorithms for model iterative optimization. Therefore, we provide an FPGA-based electronic hem control method and system. Summary of the Invention

[0003] In order to solve the above technical problems, the purpose of the present invention is to provide an electronic hemming control method and system based on FPGA.

[0004] In order to achieve the above object, the present invention provides the following technical solution: an electronic hemming control method based on FPGA, the method comprising:

[0005] Based on multi-channel parallel acquisition technology, the tension data and thickness data of the twisted fabric, the speed data, torque data and temperature data of the running motor are collected; based on data preprocessing technology, the collected data are preprocessed;

[0006] Feature extraction is performed on the pre-processed data to obtain key features of the electronic selvedge control system's operating status and the quality of selvedge fabrics. Based on the FPGA's parallel processing capabilities, key features are quickly integrated.

[0007] Based on the deep learning algorithm, a state recognition model is constructed to classify and identify the fused key features, thereby obtaining the operating mode of the electronic hem control system and the fabric state of the hem fabric.

[0008] According to the operation mode and the fabric state, an optimal control decision is established; based on the optimal control decision, a fault warning model is constructed to obtain a fault risk coefficient of the electronic selvedge control system, thereby accurately controlling the electronic selvedge control system;

[0009] According to the fault risk coefficient, a control optimization self-learning algorithm is constructed; according to the control optimization self-learning algorithm, a fault warning model is optimized, thereby ensuring the safe and stable operation of the electronic selvedge control system.

[0010] Furthermore, the process of extracting features from the pre-processed data to obtain key features of the operating status of the electronic selvedge control system and the quality of the selvedge fabric includes:

[0011] Obtain pre-processed tension data, thickness data, speed data, torque data and temperature data at different acquisition times in the same acquisition cycle;

[0012] According to the preprocessed tension data, thickness data, rotational speed data, torque data and temperature data, key features corresponding to the tension data, key features of the thickness data, key features of the rotational speed data, key features of the torque data and key features of the temperature data are obtained.

[0013] Furthermore, the process of rapidly fusing key features includes:

[0014] Read the key features of tension data, thickness data, speed data, torque data, and temperature data from the FPGA;

[0015] Build based on format and the feature quantification and standardization code of the Z-score technique, for feature quantization and standardization respectively;

[0016] The key features after feature quantization and standardization are respectively concatenated to obtain feature vector sets, including: tension feature vector set, thickness feature vector set, speed feature vector set, torque feature vector set and temperature feature vector set;

[0017] based on The parallel processing capability of the algorithm is used to process the tension feature vector set, thickness feature vector set, speed feature vector set, torque feature vector set and temperature feature vector set at the same time, and obtain the correlation and weight between different feature vector sets.

[0018] Furthermore, the process of obtaining the correlation and weights between different feature vector sets includes:

[0019] Construct a correlation calculation code suitable for FPGA, take the tension feature vector set, thickness feature vector set, speed feature vector set, torque feature vector set and temperature feature vector set as input, calculate the Pearson correlation coefficient between different feature vector sets according to the correlation calculation code, record it as the correlation coefficient, and calculate it according to Array parallel computing, generating The correlation matrix of

[0020] A dynamic weight code for the correlation coefficient of FPGA is constructed, a fault threshold is preset, and weights of different feature vector sets are obtained by taking the correlation matrix and the fault threshold as input.

[0021] Furthermore, based on the deep learning algorithm, the process of building a state recognition model includes:

[0022] Read the comprehensive correlation characteristics of several acquisition cycles from the FPGA , and carry out standardization and format conversion;

[0023] The processed comprehensive correlation features Divide into training set, validation set and test set;

[0024] Construct the long short-term memory network model code, input it into the FPGA, compile the model, and obtain the long short-term memory network model;

[0025] Optimize the long short-term memory network model and save the long short-term memory network model with the highest accuracy in the validation set;

[0026] The optimized long short-term memory network model is trained based on the training set, validation set and test set, and is recorded as a state recognition model.

[0027] Furthermore, the process of parameter tuning for the state recognition model includes:

[0028] Build the model optimization strategy code and call Tuner and initialize Tuner, searching for the number of units of the state recognition model, rate, optimizer type; The number of units in the layer is adjusted from 64 to 128, The rate was adjusted from 0.2 to 0.3, and the Optimizer and applied to state recognition model.

[0029] Furthermore, the fused key features are classified and identified to obtain the operation mode of the electronic selvedge control system and the fabric state of the selvedge fabric, and then the process of constructing the optimal control decision includes:

[0030] Define the status category name;

[0031] Map the state category name to the optimized state recognition model, and then output the operating mode confidence of the corresponding acquisition cycle and fabric state confidence ;

[0032] Preset operating mode confidence threshold and the fabric state confidence threshold ;

[0033] According to the operating mode confidence , Fabric state confidence , operating mode confidence threshold and the fabric state confidence threshold Make judgments and obtain corresponding abnormal operation status and abnormal fabric status;

[0034] Fabric status confidence level based on abnormal operation status and abnormal fabric status and operating mode confidence The control decision corresponding to the larger confidence level is compared, and the control decision corresponding to the abnormal state is solved in turn, which is recorded as the optimal control decision.

[0035] Furthermore, according to the optimal control decision, the process of constructing a fault warning model includes:

[0036] Obtain the optimal control decisions corresponding to the historical acquisition cycles in several groups of different time stages, and build a fault warning model based on the time series convolutional network;

[0037] According to the fault warning model, a fault risk coefficient is obtained; the operation mode of the actual electronic selvedge control system and the fabric state of the selvedge fabric are input into the fault warning model to obtain a real fault risk coefficient label;

[0038] The error between the fault risk coefficient and the true fault risk coefficient label is calculated through the loss function to obtain the loss value ;

[0039] according to Optimizer and loss value Update and optimize the fault warning model parameters.

[0040] The present invention further provides an FPGA-based electronic selvedge control system to implement the above-mentioned FPGA-based electronic selvedge control method, comprising: a data acquisition and preprocessing module, a feature extraction and state recognition module, an intelligent control decision module, and a self-learning module;

[0041] The data acquisition and preprocessing module collects tension data and thickness data of the twisted fabric, speed data, torque data and temperature data of the running motor, and preprocesses the collected data;

[0042] The feature extraction and state recognition module extracts features from the preprocessed data to obtain key features of the electronic selvedge control system's operating status and the quality of selvedge fabrics, and quickly fuses these key features. It also builds a state recognition model to classify and identify the fused key features.

[0043] The intelligent control decision module builds the optimal control decision according to the operation mode and fabric status; based on the optimal control decision, it builds a fault warning model to obtain the fault risk coefficient of the electronic selvedge control system, and then accurately controls the electronic selvedge control system;

[0044] The optimization self-learning module constructs a control optimization self-learning algorithm according to the fault risk coefficient; and optimizes the fault warning model according to the control optimization self-learning algorithm.

[0045] The present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program can be executed by a processor to implement the above-mentioned FPGA-based electronic hemming control method.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1. FPGA-based parallel acquisition and feature extraction technology significantly improves the efficiency of simultaneous processing of multiple data types, such as tension and thickness. High-frequency feature capture is greatly enhanced, enabling early detection of defects such as edge wear before failure occurs, significantly reducing missed detection rates under high-speed conditions. LSTM model deployment technology, through model optimization and hardware adaptation, enables low-latency inference on FPGAs, significantly reducing power consumption compared to traditional GPU solutions. State recognition accuracy is significantly improved, significantly reducing the misjudgment rate for complex conditions such as thickness anomalies.

[0048] 2. Based on the dynamic weighting of the Pearson correlation coefficient and a control optimization self-learning algorithm, the system can adjust the decision priority of characteristics such as tension and speed in real time, significantly improving control response speed. Through self-learning iterations, the accuracy of fault warnings is significantly improved, equipment maintenance cycles are extended, and energy consumption is reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction to the drawings required for use in the embodiments will be given below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0050] Figure 1 The figure is a schematic diagram of the steps of an electronic hemming control method based on FPGA.

[0051] Figure 2 This is a module schematic diagram of an electronic selvedge control system based on FPGA.

[0052] Figure 3 This is a flow chart of an electronic selvedge control system based on FPGA. DETAILED DESCRIPTION

[0053] To make the objectives, technical solutions, and advantages of the present invention more apparent, the technical solutions of the present invention will be described in detail below. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other implementations obtained by those of ordinary skill in the art without inventive effort are within the scope of protection of the present invention.

[0054] like Figure 1 As shown, an electronic hemming control method based on FPGA, the method comprising:

[0055] Based on multi-channel parallel acquisition technology, the tension data and thickness data of the twisted fabric, the speed data, torque data and temperature data of the running motor are collected; based on data preprocessing technology, the collected data are preprocessed;

[0056] Feature extraction is performed on the pre-processed data to obtain key features of the electronic selvedge control system's operating status and the quality of selvedge fabrics. Based on the FPGA's parallel processing capabilities, key features are quickly integrated.

[0057] Based on the deep learning algorithm, a state recognition model is constructed to classify and identify the fused key features, thereby obtaining the operating mode of the electronic hem control system and the fabric state of the hem fabric.

[0058] According to the operation mode and the fabric state, an optimal control decision is established; based on the optimal control decision, a fault warning model is constructed to obtain a fault risk coefficient of the electronic selvedge control system, thereby accurately controlling the electronic selvedge control system;

[0059] According to the fault risk coefficient, a control optimization self-learning algorithm is constructed; according to the control optimization self-learning algorithm, a fault warning model is optimized, thereby ensuring the safe and stable operation of the electronic selvedge control system.

[0060] It should be further explained that, in the specific implementation process, based on the multi-channel parallel acquisition technology, the specific process of collecting the tension data and thickness data of the twisted fabric, the speed data, torque data and temperature data of the running motor includes:

[0061] Setting a data acquisition device, including: a tension data acquisition unit, a thickness data acquisition unit, a speed data acquisition unit, a torque data acquisition unit, and a temperature data acquisition unit; and setting an acquisition cycle, including a number of acquisition moments;

[0062] The tension data acquisition unit deploys several tension sensors at tension measurement points on the hem fabric. Using multi-channel parallel acquisition technology, these sensors operate simultaneously. Each tension sensor converts the collected tension data into an electrical signal, which is then converted to a digital signal via an analog-to-digital converter (ADC) and transmitted via a data bus to a cache module within the FPGA.

[0063] For example, the tension sensors evenly distributed on the twisted edge fabric collect tension data simultaneously during the acquisition cycle. The multi-channel parallel acquisition function of FPGA ensures the real-time and accuracy of the tension data.

[0064] The thickness data acquisition unit arranges several thickness sensors at the edge of the twisted fabric. The thickness sensors convert the thickness data at the edge of the twisted fabric into corresponding electrical signals, which are also converted into digital signals through analog-to-digital converters (ADCs) and transmitted to the cache module inside the FPGA through the data bus.

[0065] For example, a capacitive thickness sensor or an ultrasonic thickness sensor is used to simultaneously collect thickness data at different positions on the edge of the serged fabric to obtain the distribution of the fabric thickness.

[0066] The speed data acquisition unit installs several speed sensors on the shaft of the motor. The pulse signal generated by the speed sensor is proportional to the speed of the motor. Based on multi-channel parallel acquisition technology, the pulse signals of multiple speed sensors are processed simultaneously; by counting and analyzing the pulse signals, the speed data of the motor is calculated and stored in the internal cache module of the FPGA.

[0067] For example, for multiple running motors, their speed data can be collected simultaneously to fully understand the operating status of the motors.

[0068] The torque data acquisition unit installs several torque sensors on the output shaft of the motor. The torque sensors convert torque data into electrical signals. Based on multi-channel parallel acquisition technology, torque data of different motors are collected simultaneously and stored in the internal cache module of the FPGA.

[0069] It should be further explained that by analyzing torque data, the load condition of the motor can be determined. If the torque data of multiple motors is collected simultaneously, it is helpful to analyze the load balance between the motors.

[0070] The temperature data acquisition unit arranges several temperature sensors on the motor. The temperature sensors convert temperature information into electrical signals and transmit them to the cache module inside the FPGA through the data bus.

[0071] For example, thermocouples or thermistors can be selected to simultaneously collect the temperature of motor windings and key components, so as to promptly detect temperature anomalies. Multi-channel parallel acquisition technology can simultaneously collect data from multiple temperature sensors and monitor the temperature status of the system in real time.

[0072] It should be further explained that the multi-channel parallel acquisition technology ensures the synchronous acquisition of thickness data at different positions, thereby improving the acquisition efficiency and accuracy of thickness data.

[0073] It should be further explained that, in the specific implementation process, based on the data preprocessing technology, the specific process of preprocessing the collected data includes:

[0074] The filter core is implemented based on the FPGA hardware description language. Based on the FPGA's multi-channel parallel architecture, tension data, thickness data, speed data, torque data, and temperature data are filtered simultaneously. The filter parameters are independently configured for each type of data channel, including: filter type, filter structure parameters, cutoff frequency, transition band width, stopband attenuation, sampling frequency, passband ripple, and phase response.

[0075] It should be further explained that the filter parameters can be dynamically adjusted according to the actual accuracy requirements.

[0076] According to the channel after configuring the filter parameters, filtered tension data, thickness data, speed data, torque data and temperature data are obtained.

[0077] Preset standard parts;

[0078] The data acquisition device is calibrated based on the standard parts, and the data acquisition Group calibration data ,in, is the standard value; is the corresponding sensor output value;

[0079] The calibration curve is fitted based on the least squares method, which is:

[0080] ;in, is the true value; is the corresponding sensor output value; 、 、 is the calibration coefficient;

[0081] According to the calibration curve, the true value after calibration is obtained ;

[0082] The calibration factor 、 、 The data is stored in the FPGA memory, and the calibration calculation is completed through the hardware multiplier and adder after real-time data acquisition.

[0083] For example, the output deviation of the thickness sensor at 25°C is After temperature compensation calibration, the full temperature range Measurement error .

[0084] Perform multi-channel parallel normalization on the calibrated data, including: calculating the as well as , and stored in the register group. Each channel performs normalization calculation independently, and uses the parallel processing unit of FPGA to simultaneously process the data collected by the tension data acquisition unit, thickness data acquisition unit, speed data acquisition unit, torque data acquisition unit and temperature data acquisition unit to obtain the corresponding normalized data.

[0085] For example, the speed data range of the motor , the range of torque data , after normalization, both are mapped to , eliminating the dimension effect.

[0086] It should be further explained that, in the specific implementation process, the specific process of extracting features from the pre-processed data and obtaining the key features of the operating status of the electronic selvedge control system and the quality of the selvedge fabric includes:

[0087] Perform feature extraction on the preprocessed data to obtain the key features of the preprocessed data, including:

[0088] The specific process of obtaining the key features of the corresponding data of the pre-processed twisted edge fabric is as follows:

[0089] The tension data and thickness data at different acquisition times in the same acquisition cycle are recorded as , ,in, Indicates the corresponding collection time;

[0090] Based on the sliding window algorithm, the mean of tension data and thickness data is obtained as well as , standard deviation as well as , as basic data support, the specific acquisition process belongs to the conventional process in this field and will not be repeated;

[0091] According to the mean of the tension data and standard deviation , to obtain the deflection of the lehenge fabric , kurtosis and pulse factor ; It should be further explained that skewness The distribution of symmetry, negative values ​​indicate left deviation, that is, the risk of sudden tension drop; impulse factor Indicates the impact characteristics of tension data, detects sudden changes in tension, which is a precursor to the breakage of the hem fabric; the deflection ;in, The function represents the mathematical expectation; the kurtosis The pulse factor ;in, The function represents the maximum value, represents the root mean square value. The specific acquisition process belongs to the conventional process in this field and will not be repeated here.

[0092] based on , obtain the fundamental frequency amplitude of the tension data in the frequency domain and harmonic content ; It should be further explained that the fundamental frequency amplitude Used to detect tension fluctuations caused by power supply interference; harmonic content Used to evaluate tension harmonics caused by transmission system nonlinearity, such as gear meshing frequency;

[0093] The harmonic content ;in, Indicates the Subharmonic amplitude; Indicates the upper limit of harmonic order; The specific acquisition process belongs to the conventional process in this field and will not be described in detail.

[0094] based on , obtain the time-frequency spectrum mutation of tension data , wavelet detail coefficient and wavelet energy entropy ; It should be further explained that the time spectrum mutation Indicates the detection of a sudden increase or decrease in a specific frequency component in the time-frequency spectrum, that is, the edge of the hem fabric is broken, such as a sudden increase in frequency noise; wavelet energy entropy It indicates the complexity of tension fluctuations. Too large or too small indicates that the control system fails. The specific acquisition process belongs to the routine process in this field and will not be repeated here.

[0095] Based on time domain feature extraction technology and according to the mean value of thickness data , standard deviation , obtain the time domain characteristics of the hem fabric thickness data, including the mean deviation , local standard deviation , gradient and slope of change It should be further explained that the mean deviation indicates the deviation of the actual collected thickness from the set value, which directly affects the compliance of the product specifications; the local standard deviation is used to detect thickness non-uniformity, such as streak defects; the gradient indicates the thickness change rate, identifying thickness mutations, such as a sudden drop in edge thickness; the change slope is used to determine whether the thickness transition area meets the process requirements, such as a gradual change in coating thickness. The specific acquisition process belongs to the conventional process in this field and will not be repeated here;

[0096] Based on the frequency domain feature extraction technology, the frequency domain features of the hem fabric thickness data are obtained, including the spectrum kurtosis and cutoff frequency. It should be further explained that the spectrum kurtosis is used to detect sudden thickness anomalies, such as spikes caused by impurities; the cutoff frequency is used to evaluate the high-frequency components of thickness fluctuations, reflecting the resolution of the measurement system. The specific acquisition process is a routine process in this field and will not be repeated here.

[0097] The specific process of obtaining the key features of the motor corresponding data after preprocessing:

[0098] The speed data, torque data and temperature data of the acquisition period are recorded as , as well as ,in, Indicates the corresponding collection time;

[0099] Based on the sliding window algorithm, the mean of speed data, torque data and temperature data is obtained 、 as well as , standard deviation 、 as well as , as basic data support;

[0100] Based on the time domain feature extraction technology, the time domain features of the speed data, torque data and temperature data are obtained, which are the speed fluctuation rate, overshoot and speed fluctuation period of the speed data, the torque fluctuation rate, peak torque ratio and torque change rate of the torque data, and the temperature gradient, temperature fluctuation and thermal time coefficient of the temperature data; it should be further explained that the fluctuation of the speed data is too large, resulting in uneven fabric tension, the overshoot indicates the speed overshoot in the step response, and reflects the damping ratio of the control system; the speed fluctuation period is used to identify periodic load fluctuations, such as mechanical resonance; the fluctuation of the torque fluctuation rate of the torque data is too large, resulting in transmission system fatigue, such as coupling wear, the peak torque ratio is used to evaluate the impact load degree, such as the peak torque when the loom is beating the weft, and the torque change rate is used to detect torque mutations, such as the sudden increase in torque caused by fabric jamming. The specific acquisition process belongs to the conventional process in this field and will not be repeated here.

[0101] Based on the frequency domain feature extraction technology, the frequency domain features of the speed data, torque data and temperature data are obtained, which are the rotational frequency, rotational frequency sideband and spectral entropy of the speed data, the gear meshing frequency and torque ripple coefficient of the torque data, the low-frequency temperature fluctuation, temperature spectrum peak and temperature-load correlation of the temperature data; it should be further explained that the rotational frequency of the speed data is used to determine the motor base frequency; the rotational frequency sideband is used to identify load fluctuations, such as the increase in sideband energy caused by conveyor belt slippage; the spectral entropy represents the change in spectral complexity, and too large or too small indicates bearing failure or uneven load; the gear meshing frequency of the torque data is used to detect gear wear; the torque ripple coefficient is used to evaluate the smoothness of the motor output, such as the ripple coefficient , the motor output vibration; the low-frequency temperature fluctuations of the temperature data are used to identify periodic faults of the cooling system, such as fan speed fluctuations; the temperature spectrum peak is used to locate the control system oscillation, such as Temperature oscillation caused by inappropriate parameters; temperature-load correlation is used to evaluate the impact of load changes on temperature. For example, a decrease in temperature-load correlation indicates a failure of the cooling system. The specific acquisition process is a routine process in this field and will not be repeated here.

[0102] It should be further explained that, in the specific implementation process, the specific process of rapidly integrating key features based on the FPGA parallel processing capability includes:

[0103] Read the key features of tension data, thickness data, speed data, torque data, and temperature data from the FPGA;

[0104] Build based on format and the feature quantification and standardization code of the Z-score technique, for feature quantization and standardization respectively;

[0105] For example, the local standard deviation of thickness Convert to Format: , rounded to , the actual value is , if the accuracy standard is , the actual value is , the error is The accuracy of the format meets the quantization requirements.

[0106] The key features after feature quantization and standardization are respectively concatenated with feature vectors to obtain feature vector sets, including: tension feature vector set, thickness feature vector set, speed feature vector set, torque feature vector set and temperature feature vector set.

[0107] based on The parallel processing capability is used to process the tension feature vector set, thickness feature vector set, speed feature vector set, torque feature vector set and temperature feature vector set at the same time, and obtain the correlation and weight between different feature vector sets. The specific process includes:

[0108] It should be further explained that the format of the tension feature vector set is:

[0109] ;

[0110] The format of the thickness feature vector set is: [mean deviation, local standard deviation, gradient, slope, spectrum tilt, cutoff frequency];

[0111] The format of the speed feature vector set is: [fluctuation rate, overshoot, speed fluctuation period, speed frequency, speed frequency sideband, spectrum entropy];

[0112] The format of the torque feature vector set is: [torque fluctuation rate, peak torque ratio, torque change rate, gear meshing frequency, torque ripple coefficient];

[0113] The format of the temperature feature vector set is: [temperature gradient, temperature fluctuation, thermal time coefficient, low-frequency temperature fluctuation, temperature spectrum peak, temperature load correlation].

[0114] Construct a correlation calculation code suitable for FPGA, take the tension feature vector set, thickness feature vector set, speed feature vector set, torque feature vector set and temperature feature vector set as input, calculate the Pearson correlation coefficient between different feature vector sets according to the correlation calculation code, record it as the correlation coefficient, and calculate it according to Array parallel computing, generating The correlation matrix of

[0115] For example, the correlation matrix can be expressed in the form of a table:

[0116] As shown in Table 1:

[0117] Correlation coefficient Tension characteristic vector set Thickness feature vector set Speed ​​feature vector set Torque characteristic vector set Temperature feature vector set Tension characteristic vector set 1.000 0.653 0.127 0.812 0.345 Thickness feature vector set 0.653 1.000 0.328 0.547 0.189 Speed ​​feature vector set 0.127 0.328 1.000 0.763 0.621 Torque characteristic vector set 0.812 0.547 0.763 1.000 0.476 Temperature feature vector set 0.345 0.189 0.621 0.476 1.000

[0118] A dynamic weight code for the correlation coefficient suitable for FPGA is constructed, and a fault threshold is preset. The weights of different feature vector sets are obtained by taking the correlation matrix and the fault threshold as input. It should be further explained that the fault threshold is set by the actual tension feature vector set, thickness feature vector set, speed feature vector set, torque feature vector set and temperature feature vector set.

[0119] For example, in the normal state: tension feature vector set (0.35) + thickness feature vector set (0.25) + speed feature vector set (0.15) + torque feature vector set (0.15) + temperature feature vector set (0.10);

[0120] Abnormal state: When abnormal tension fluctuation is detected, dynamic adjustment is made as follows:

[0121] Tension feature vector set (0.60) + thickness feature vector set (0.15) + speed feature vector set (0.05) + torque feature vector set (0.15) + temperature feature vector set (0.05).

[0122] Based on the weighted average method, key features are quickly integrated to generate comprehensive correlation features ; The comprehensive correlation characteristics ;in, is the correlation coefficient; is the product of the weights corresponding to the two correlation coefficients;

[0123] It should be further explained that, in the specific implementation process, the specific process of building a state recognition model based on the deep learning algorithm includes:

[0124] Read the comprehensive correlation characteristics of several acquisition cycles from the FPGA ;

[0125] Construct data normalization code and save it to FPGA. The data processing unit in FPGA executes data normalization code to synthesize correlation features. Standardization is performed to ensure the stability of model training.

[0126] Construct the format conversion code and save it to FPGA. The data processing unit in FPGA executes the format conversion code to synthesize the correlation features. Convert to The required time series data format.

[0127] For example, setting the time step , the comprehensive correlation features of each acquisition cycle Expanded to include the first 10 time points of the series, forming a three-dimensional array to capture the temporal dependence of the data.

[0128] It should be further explained that the three-dimensional array format is: [number of samples, time steps, number of features].

[0129] Comprehensive correlation features after time series data format conversion Divide into training set, validation set and test set;

[0130] Construct the long short-term memory network model code, input it into the FPGA, compile the model, and obtain the long short-term memory network model;

[0131] Based on the LSTM model, configure a callback function to stop training and restore the optimal weights when the validation set loss stops decreasing for five consecutive rounds to avoid overfitting. When validation set loss stagnates, multiply the LSTM model's learning rate by 0.2 to help the model break out of the local optimum. Save the LSTM model with the highest validation set accuracy to ensure optimal parameters for deployment.

[0132] Combine several comprehensive correlation features The number of samples is set to 64 to balance memory usage and gradient estimation accuracy; training is done for 50 rounds. Code instructions disrupt the data order to avoid periodic deviations.

[0133] The loss and accuracy of the training set and validation set are recorded in real time, and a curve is generated to visualize the training dynamics to determine whether the trained long short-term memory network model has converged or is overfitting; the trained long short-term memory network model is recorded as a state recognition model.

[0134] It should be further explained that the specific process of generating the curve based on the loss and accuracy of the training set and the validation set includes:

[0135] Create Parameters , the parameters Used to set the canvas width and height; create parameters , used to create a subgraph layout, the subgraph layout includes: a left subgraph and a right subgraph; the left subgraph is used to describe the loss curve, and the right subgraph is used to describe the accuracy curve; create parameters , respectively used to determine the line style of the curve;

[0136] based on Instructions, automatically display the loss curve and accuracy curve, and refresh the loss curve and accuracy curve by adding a timer; it should be further explained that the state of the loss curve is determined as follows:

[0137] 1. Ideal situation: The training loss and validation loss decrease simultaneously, and the validation loss is slightly higher than the training loss. The final loss tends to be stable without significant fluctuations.

[0138] 2. Abnormal situation: Overfitting: The training loss continues to decrease, and the validation loss first decreases and then increases;

[0139] Underfitting: Both the training loss and the validation loss are higher than expected and decrease slowly.

[0140] Status judgment of the accuracy curve:

[0141] 1. Ideal situation: The training accuracy and verification accuracy improve simultaneously, and the final verification accuracy is close to the training accuracy;

[0142] 2. Abnormal situations: Data imbalance: The verification accuracy fluctuates greatly, and the sample accuracy is significantly low; Model instability: The accuracy fluctuates irregularly in the later stage.

[0143] Parameters of the state recognition model are tuned, and the specific process includes:

[0144] Build the model optimization strategy code and call Tuner and initialize Tuner, searches for the number of units in the state recognition model ( , step size 16), Rate( , step length ), optimizer type ( ), and then obtain the optimal parameters.

[0145] It should be further explained that to maximize the accuracy of the validation set, we use The algorithm dynamically allocates computing resources and efficiently searches for the optimal parameter combination.

[0146] The optimal parameters obtained by tuning, such as The number of units in the layer is adjusted from 64 to 128, The rate was adjusted from 0.2 to 0.3, and the The optimizer is applied to the state recognition model to improve generalization ability.

[0147] It needs to be further explained that in Layers and fully connected layers are added Regularization, constraining the complexity of the state recognition model; increasing The rate is reduced to 0.3 to further suppress overfitting and ensure the stability of the model on unknown data.

[0148] It should be further explained that, in the specific implementation process, the specific process of classifying and identifying the fused key features, obtaining the operation mode of the electronic selvedge control system and the fabric state of the selvedge fabric, and then constructing the optimal control decision includes:

[0149] Define status category names, including: the operating modes of the electronic leno control system are normal operation, edge fluctuation, high-speed production, and low-speed debugging; the fabric status of leno fabric is qualified, edge wear, thickness abnormality, and comprehensive defects;

[0150] Map the state category name to the optimized state recognition model, and then output the operating mode confidence of the corresponding acquisition cycle and fabric state confidence ; The confidence level of the operating mode , including: normal operation confidence, edge fluctuation confidence, high-speed production confidence and low-speed debugging confidence; fabric state confidence , including: qualified status confidence, edge wear confidence, thickness anomaly confidence and comprehensive defect confidence;

[0151] Preset operating mode confidence threshold and the fabric state confidence threshold ;

[0152] If the actual operating mode confidence , then the operating mode is considered to be high confidence and responds with a green light, indicating no abnormal operating status;

[0153] If the actual operating mode confidence , then the operating mode is considered to be of medium confidence, and the yellow light is displayed, indicating that there is an operating status risk, and the maintenance personnel are reminded to conduct an inspection;

[0154] If the actual operating mode confidence , then the operating mode is determined to be of low confidence, and the red light is responded to, indicating that there is an abnormal operating status, reminding maintenance personnel to implement corresponding control decisions.

[0155] If the actual fabric state confidence , then the fabric state is determined to be high confidence and responds to the green light, and there is no abnormality in the fabric state;

[0156] If the actual fabric state confidence , then the fabric status is considered to be medium confidence, and the yellow light is displayed, indicating that there is a risk of abnormal fabric status, and the maintenance personnel are reminded to check;

[0157] If the actual fabric state confidence , then the fabric state is determined to be low confidence, and the red light is responded to, indicating that the fabric state is abnormal, and the maintenance personnel are reminded to implement the corresponding control decision;

[0158] It should be further explained that the confidence level of the operating mode and fabric state confidence The higher the value, the corresponding operation mode and fabric state.

[0159] Fabric status confidence level based on abnormal operation status and abnormal fabric status and operating mode confidence The control decision corresponding to the larger confidence level is compared, and the control decision corresponding to the abnormal state is executed first. The control decision corresponding to the abnormal state is solved in turn and recorded as the optimal control decision. The control decision corresponding to the state is not repeated here.

[0160] For example, if the edge fluctuation confidence is greater than the high-speed production confidence, the control decision of edge fluctuation is executed first, and then the control decision of high-speed production is executed; if the high-speed production confidence is greater than the edge wear confidence, the control decision of high-speed production is executed first, and then the control decision of edge wear is executed; if the edge fluctuation confidence is equal to the high-speed production confidence, the control decision of the corresponding state is executed arbitrarily.

[0161] It should be further explained that, in the specific implementation process, according to the optimal control decision, the specific process of constructing the fault warning model includes:

[0162] Obtain the optimal control decisions corresponding to the historical acquisition cycles in several different time stages. It should be further explained that the optimal control decisions executed in different time stages are universal, reducing the impact of time factors on the model.

[0163] The optimal control decisions corresponding to the historical acquisition cycles in several groups at different time stages are grouped and labeled as is a natural number;

[0164] Will The optimal control decision corresponding to the historical acquisition period in different time stages of the group is used as sample data, and is less than The natural number of is recorded as the sample set;

[0165] The optimal control decisions of the remaining several groups corresponding to the historical acquisition cycles in different time periods are used as test sets; and a training sample set is formed based on the sample set and the test set;

[0166] Based on the time series convolutional network, a standard fault warning model is constructed;

[0167] The training sample set is input into the standard fault warning model to train the standard fault warning model, and the trained standard fault warning model is recorded as the fault warning model.

[0168] It should be further explained that, in the specific implementation process, based on the fault risk coefficient, a control optimization self-learning algorithm is constructed to optimize the fault warning model. The specific process includes:

[0169] The difference between the predicted operation mode of the electronic hem control system and the fabric state of the hem fabric and the actual operation mode of the electronic hem control system and the fabric state of the hem fabric is measured based on the cross entropy loss function, and the selection The optimizer performs optimization, and the specific process includes:

[0170] According to the fault warning model, a fault risk coefficient is obtained; the operation mode of the actual electronic hem control system and the fabric state of the hem fabric are input into the fault warning model to obtain a real fault risk coefficient label;

[0171] The error between the fault risk coefficient and the true fault risk coefficient label is calculated using the loss function, and the formula is:

[0172] ;in, is the loss value; is the real fault risk factor label, and its value is or ; The category predicted by the fault warning model is probability; The category predicted by the fault warning model is probability;

[0173] according to The optimizer updates the fault warning model parameters based on minimizing the loss function. The update formula is:

[0174] ;in, is the learning rate; is the model parameter of the fault warning model, by updating , optimize the fault warning model to ensure the safe and stable operation of the electronic hemming control system.

[0175] like Figure 2 As shown, an electronic selvedge control system based on FPGA includes: a data acquisition and preprocessing module, a feature extraction and state recognition module, an intelligent control decision module and a self-learning module;

[0176] The data acquisition and preprocessing module collects tension data and thickness data of the twisted fabric, speed data, torque data and temperature data of the running motor, and preprocesses the collected data;

[0177] The feature extraction and state recognition module extracts features from the preprocessed data to obtain key features of the electronic selvedge control system's operating status and the quality of selvedge fabrics, and quickly fuses these key features. It also builds a state recognition model to classify and identify the fused key features.

[0178] The intelligent control decision module builds the optimal control decision according to the operation mode and fabric status; based on the optimal control decision, it builds a fault warning model to obtain the fault risk coefficient of the electronic selvedge control system, and then accurately controls the electronic selvedge control system;

[0179] The optimization self-learning module constructs a control optimization self-learning algorithm according to the fault risk coefficient; and optimizes the fault warning model according to the control optimization self-learning algorithm.

[0180] The above embodiments are only used to illustrate the technical method of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. An electronic selvedge control method based on FPGA, characterized in that: The method comprises: Based on multi-channel parallel acquisition technology, the tension data and thickness data of the twisted fabric, the speed data, torque data and temperature data of the running motor are collected; based on data preprocessing technology, the collected data are preprocessed; Feature extraction is performed on the pre-processed data to obtain key features of the electronic selvedge control system's operating status and the quality of selvedge fabrics. Based on the FPGA's parallel processing capabilities, key features are quickly integrated. Based on the deep learning algorithm, a state recognition model is constructed to classify and identify the fused key features, thereby obtaining the operating mode of the electronic hem control system and the fabric state of the hem fabric. Based on the operating mode and fabric status, the optimal control decision is constructed, including: Define the status category name; Map the state category name to the optimized state recognition model, and then output the operating mode confidence of the corresponding acquisition cycle and fabric state confidence ; Preset operating mode confidence threshold and the fabric state confidence threshold ; According to the operating mode confidence , Fabric state confidence , operating mode confidence threshold and the fabric state confidence threshold Make judgments and obtain corresponding abnormal operation status and abnormal fabric status; Fabric status confidence level based on abnormal operation status and abnormal fabric status and operating mode confidence The control decision corresponding to the larger confidence level is executed first, and the control decisions corresponding to the abnormal state are solved in turn and recorded as the optimal control decision; According to the optimal control decision, a fault warning model is constructed to obtain a fault risk coefficient of the electronic selvedge control system, thereby accurately controlling the electronic selvedge control system; According to the fault risk coefficient, a control optimization self-learning algorithm is constructed; according to the control optimization self-learning algorithm, a fault warning model is optimized, thereby ensuring the safe and stable operation of the electronic selvedge control system.

2. The electronic hemming control method based on FPGA according to claim 1, characterized in that: The process of extracting features from the preprocessed data to obtain the key features of the electronic selvedge control system operating status and selvedge fabric quality includes: Obtain pre-processed tension data, thickness data, speed data, torque data and temperature data at different acquisition times in the same acquisition cycle; According to the preprocessed tension data, thickness data, rotational speed data, torque data and temperature data, key features corresponding to the tension data, key features of the thickness data, key features of the rotational speed data, key features of the torque data and key features of the temperature data are obtained.

3. The electronic hemming control method based on FPGA according to claim 2, characterized in that: The process of rapid fusion of key features includes: Read the key features of tension data, thickness data, speed data, torque data, and temperature data from the FPGA; Build based on format and the feature quantification and standardization code of the Z-score technique, for feature quantization and standardization respectively; The key features after feature quantization and standardization are respectively concatenated to obtain feature vector sets, including: tension feature vector set, thickness feature vector set, speed feature vector set, torque feature vector set and temperature feature vector set; based on The parallel processing capability of the algorithm is used to process the tension feature vector set, thickness feature vector set, speed feature vector set, torque feature vector set and temperature feature vector set at the same time, and obtain the correlation and weight between different feature vector sets.

4. The electronic hemming control method based on FPGA according to claim 3, characterized in that: The process of obtaining the correlation and weights between different feature vector sets includes: Construct a correlation calculation code suitable for FPGA, take the tension feature vector set, thickness feature vector set, speed feature vector set, torque feature vector set and temperature feature vector set as input, calculate the Pearson correlation coefficient between different feature vector sets according to the correlation calculation code, record it as the correlation coefficient, and calculate it according to Array parallel computing, generating The correlation matrix of A dynamic weight code for the correlation coefficient of FPGA is constructed, a fault threshold is preset, and weights of different feature vector sets are obtained by taking the correlation matrix and the fault threshold as input.

5. The electronic hemming control method based on FPGA according to claim 4, characterized in that: Based on the deep learning algorithm, the process of building a state recognition model includes: Read the comprehensive correlation characteristics of several acquisition cycles from the FPGA , and carry out standardization and format conversion; The processed comprehensive correlation features Divide into training set, validation set and test set; Construct the long short-term memory network model code, input it into the FPGA, compile the model, and obtain the long short-term memory network model; Optimize the long short-term memory network model and save the long short-term memory network model with the highest accuracy in the validation set; The optimized long short-term memory network model is trained based on the training set, validation set and test set, and is recorded as a state recognition model.

6. The electronic hemming control method based on FPGA according to claim 5, characterized in that: The process of tuning parameters of the state recognition model includes: Build the model optimization strategy code and call Tuner and initialize Tuner, searching for the number of units of the state recognition model, rate, optimizer type; The number of units in the layer is adjusted from 64 to 128, The rate was adjusted from 0.2 to 0.3, and the The optimizer is applied to the state recognition model.

7. The electronic hemming control method based on FPGA according to claim 6, characterized in that: According to the optimal control decision, the process of constructing a fault warning model includes: Obtain the optimal control decisions corresponding to the historical acquisition cycles in several groups of different time stages, and build a fault warning model based on the time series convolutional network; According to the fault warning model, a fault risk coefficient is obtained; the operation mode of the electronic selvedge control system and the fabric state of the selvedge fabric are input into the fault warning model to obtain a real fault risk coefficient label; The error between the fault risk coefficient and the true fault risk coefficient label is calculated through the loss function to obtain the loss value ; according to Optimizer and loss value Update and optimize the fault warning model parameters.

8. An FPGA-based electronic selvedge control system, implementing the FPGA-based electronic selvedge control method according to any one of claims 1 to 7, characterized in that: include: Data acquisition and preprocessing module, feature extraction and state recognition module, intelligent control decision module and self-learning module; The data acquisition and preprocessing module collects tension data and thickness data of the twisted fabric, speed data, torque data and temperature data of the running motor, and preprocesses the collected data; The feature extraction and state recognition module extracts features from the preprocessed data to obtain key features of the electronic selvedge control system's operating status and the quality of selvedge fabrics, and quickly fuses these key features. It also builds a state recognition model to classify and identify the fused key features. The intelligent control decision module builds the optimal control decision according to the operation mode and fabric status; based on the optimal control decision, it builds a fault warning model to obtain the fault risk coefficient of the electronic selvedge control system, and then accurately controls the electronic selvedge control system; The optimization self-learning module constructs a control optimization self-learning algorithm according to the fault risk coefficient; and optimizes the fault warning model according to the control optimization self-learning algorithm.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and the computer program can be executed by a processor to implement the FPGA-based electronic hemming control method according to any one of claims 1 to 7.

Images