Abstract:Millimeter-wave radar gesture recognition based on deep learning has attracted more and more attention due to its characteristics of contact-free, privacy protection and low environmental dependence. However, most of the current learning methods use fully supervised methods, whose performance is limited by the acquisition and annotation of radar data, and their learning samples all come from a single environment, which greatly affects the transfer ability in different scenarios. Therefore, this paper proposes a A cross-domain gesture recognition method based on semisupervised generative adversarial networks. First, through data preprocessing, the dynamic mixed time feature map (DFTM) is extracted to eliminate environmental interference and more comprehensively characterize the dynamic characteristics of gestures; secondly, data enhancement is performed based on the characteristics of millimeter wave signals to further expand the amount of data and improve the model. Generalization ability; thirdly, in order to solve the problem that the labeled data available in practical applications is usually less, an improved semi-supervised generative adversarial network is proposed and constructed. A classifier is added on the basis of the original GAN to help improve the performance by generating data. The discriminative ability of the classifier simultaneously utilizes a small amount of labeled data in the source domain and a large amount of unlabeled data in the target domain to achieve domainindependent gesture recognition. Experimental results show that the average gesture recognition accuracy for new users, new environments and new locations reaches 98.21%, 95.23% and 97.6% respectively. Compared with other existing gesture recognition methods, the method proposed in this article can achieve high cross-domain gesture recognition accuracy even with only a small amount of labeled data, providing new research ideas for subsequent millimeter wave radar human-computer interaction.

Abstract:Aiming at the problems of internal complexity and too hard predicting the dissolved gas concentration of transformer oil，a method combining VMD with CS-SVR was proposed for decomposing, predicting, and reconstructing gas concentration. In this paper, firstly, VMD is utilized to decompose the original dissolved gas concentration into a set of stationary modal components. Subsequently, SVR, which has relatively good predictive performance, was used to predict each modal component separately. Finally, CS is utilized for global search to optimize and select SVR parameters, and the predicted dissolved gas concentration results are overlaid and reconstructed. Through simulation experiments on the H2 content, the root mean square error is 0.124 μL/L and the average absolute percentage error is 1.19%, effectively enhancing prediction accuracy. Further validation of the model′s effectiveness is conducted through modeling and predicting CO and C2H4. The results indicate that the VMD-CS-SVR model has high accuracy and is suitable for predicting dissolved gas concentration in transformer oil.

Abstract:To address the issue of insufficient spatial correlation information for emotion EEG characterization extracted by shallow graph convolution, this paper proposes a deep graph convolutional network model. The model utilizes deep graph convolution to learn the intrinsic relationships among global channels of emotional EEG, applying residual connections and weight self-mapping during the convolutional propagation process to address the problem of node features in deep graph convolution networks converging to a fixed space and failing to learn effective features. Additionally, PN regularization is added after the convolutional layer to expand the distance between different emotional features and improve emotion recognition performance. Experimental results on the SEED dataset show that compared to shallow graph convolution networks, the accuracy of the proposed model has increased by 0.7% while the standard deviation has decreased by 3.15. These results demonstrate the effectiveness of the global brain region spatial correlation information extracted by this model for emotion recognition.

Abstract:The UAV passive positioning system can realize the bistatic passive positioning of the target with the help of radio and television signals, and the performance of the signal processing board directly determines the positioning accuracy. In order to adapt to the light UAV platform and ensure the accuracy of positioning measurement, a miniaturized signal processing board is designed in this paper. Firstly, the traditional architecture of the signal processing board is simplified and optimized, and FPGA+AD9467 is used as the overall architecture of signal processing. In order to solve the signal integrity problem caused by the miniaturized design, the Cadence and HFFS simulation software are used to simulate and analyze the signal integrity of the processing board during the whole design process, and the signal integrity of the processing board is ensured from the aspects of reflection, crosstalk, and electromagnetic shielding. In order to improve the signal resolution ability of the processing board, a four-spectrum line difference algorithm based on Blackman window function was designed on the basis of the traditional FFT spectrum measurement. The algorithm is used to test the spectral characteristics of the received signal of the processing board, which reduces the measurement error caused by spectrum leakage and fence effect and improves the measurement accuracy. The SNR is better than 90 dB and the SFDR is better than 75 dB in the 29.5 kHz bandwidth. The test results show that the signal processing ability of the designed miniaturized signal processing board is excellent, which meets the application scenarios of passive positioning of light UAV.

Abstract:Channel phase inconsistency is one of the important factors affecting the direction-finding accuracy, rectify zero-phase can improve the direction-finding accuracy. In this paper, an outlier recognition method based on zero-phase difference fitting model is proposed, by deducing the linear fitting model of zero-phase, the fitting phase error is established, and the fitting phase error follows normal distribution, so that the outlier recognition of zero-phase is converted into the gross error recognition of fitting phase error, Finally, the gross error is discriminated and eliminated by 3σ criterion and Grubbs criterion. This method can effectively identify and eliminate the outlier recognition of zero-phase. The experimental calculation shows that outlier recognition and elimination method base on zero-phase difference can effectively improve the direction-finding accuracy, and has been applied in engineering.

Abstract:In order to address the problem that the system performance of traditional control strategy of three-level inverter-fed permanent magnet synchronous motor decreases due to parameter changes, a model-free predictive current control method based on extended state observer is proposed to enhance the robustness of the system. Firstly, the ultralocal model of the system is established to predict the future current state, which does not involve any motor parameters and solves the problem that parameter changes affect the system performance. Then, for the total disturbance of the system, dq-axes’ expanded state observers are designed to estimate the nonlinear part of the ultra-local model. Furthermore, the optimal voltage vector is selected by cost function, and the balance control of midpoint potential is realized by using the redundancy characteristic of small vector. Finally, comparative experiments are conducted on a DSP experimental test rig, in terms of the traditional three-level model predictive current control, model-free predictive current control and the proposed method. The experimental results verify that the proposed method has preferable dynamic and steady-state performance and improves the anti-disturbance ability of the three-level drive system effectively.

Abstract:In order to solve the problems of low control accuracy and poor system fitness caused by the difficulty of selecting the weight matrix of quadratic linear regulator (LQR)，this paper was designed a nonlinear decreasing weight particle swarm optimization (NLDW-PSO) algorithm. Based on the two-degree-of-freedom vehicle dynamics model, the lateral tracking error model is constructed, and the LQR steady-state error is eliminated by feedforward control. With lateral deviation, heading deviation and front wheel steering angle as evaluation functions, the system output error state is fed back to NLDW-PSO algorithm, The designed nonlinear decreasing inertia weight factor can improve the particle population optimization performance, which adaptively adjusts the LQR weight coefficient update strategy to form a closed-loop optimization control, and finally obtains the extreme value of objective function of the system. The tracking effect of the designed controllers is compared, the results showed that the proposed NLDW-PSO optimized LQR algorithm has the best tracking control effect, and it′s maximum Lateral error was 0.076m by Carsim/Smulink co-simulation, and the mean Lateral error was reduced by 69.74% compared with the fixed weight coefficient LQR. The tracking control accuracy and adaptive ability of the vehicle are significantly improved, and it has strong robustness to velocity change.

Abstract:Absolute positioning accuracy is an important index to measure robot performance. In order to improve the absolute positioning accuracy of industrial robots, a method of positioning error classification based on kinematic calibration and spatial interpolation is proposed. Firstly, the kinematics model of the robot is established based on D-H method, and the error model of the end position of the robot is established by using differential kinematics theory. The kinematic parameters are identified by using the tolerance ridge estimation combined with IGG3 weight factor function. Then, based on the spatial similarity of robot positioning errors, the remaining errors are compensated by the spatial interpolation method. Finally, the proposed method is verified by experiments. The results show that the robot positioning error RMS decreases from 0.812 mm before compensation to 0.049 mm, and the accuracy is increased by 93.97%. This method can effectively reduce the absolute positioning error of the robot and improve the positioning accuracy.

Abstract:Aiming at the problem that the health status of bearings cannot be directly monitored and predicted, we designed a L1 regularized bidirectional gating recurrent unit model and a health index constructed by Bray Curtis distance, which can directly represent the health status of bearings. Firstly, L1 regularization is used to extract effective features from the current bearing vibration data as degradation features, and the features of the first time window of the vibration data as health features. Then, the Bray Curtis distance between the bearing degradation features and health features is calculated to construct the HI of the bearing. The health status of the bearing is monitored in real time through the cloud monitoring platform, and the future health status is predicted using the BiGRU model. Once the HI of the bearing exceeds the change rate threshold, the platform will alarm immediately, and the health status of the bearing is predicted. The model is compared with bidirectional long short term memory and bidirectional long short term memory with attention models. The results show that the accuracy of this model is 97.52%, much higher than the other two models, and the model is more lightweight, which reflects the practicability of this method.

Abstract:CoaXPress is a new high-speed digital image transmission interface standard suitable for a variety of high-speed and high-bandwidth image transmission applications. In this paper, a high-speed serial transmission system based on CoaXPress interface is designed and implemented. To solve the problem of multi-channel high-speed serial data transmission, scheduling, caching and synchronization, four CoaXPress interfaces are designed for each hardware module. The module is based on FPGA, uses PCIe3.0×8 interface to communicate with the main controller, and uses DDR4 to cache high-speed data. In FPGA firmware logic design, XDMA is used to communicate with the main controller, FDMA is used to complete the data scheduling of DDR4, and GTH is used to send and receive high-speed serial data. Based on the combination of FIFO cache synchronization technology and PXI_TRIG trigger bus technology, the synchronization between 32 transmitting interfaces of 8 CoaXPress transmitting modules is successfully realized. Finally, the CoaXPress interface module and system are tested, and the eye map, code rate, bit error rate and synchronization accuracy of CoaXPress interface meet the requirements. The high-speed serial transmission system based on CoaXPress interface designed in this paper is stable and reliable, and has been applied to the test of load-data transmission link of a new generation of space vehicle.

Abstract:To address the issues of low automation in the gait training platform for bipedal robots and the difficulty in coordinating real-time observation of robot walking and status information during gait debugging by a single operator, this paper proposes a stereo vision and pose recognition-based tracking servo system for celestial orbital robots. Firstly, using a stereo camera setup, the left and right camera images are matched to obtain depth information for each pixel. Based on the depth information obtained from the images, the pose of the bipedal robot is recognized, and the depth information of each joint is extracted. The motion state of the bipedal robot is determined based on the joint depth information, and a tracking strategy is formulated for servo control. By introducing pose recognition, a higher level of automation and tracking protection can be achieved based on the changes in the bipedal robot′s pose. Experimental results demonstrate a high level of automation and dynamic tracking performance.

Abstract:To investigate the texture features of recognition memory EEG and address the issue of structural instability in extracting EEG texture features using vertical symmetric local graph structure (VSLGS) and symmetric local graph structure (SLGS). A recognition memory experiment was designed based on new and old paradigms, and relevant EEG data were collected from 35 medical students and 35 non-medical students. The EEG data were categorized into six different stages: learning medical images, learning non-medical images, recognizing old medical images, recognizing old non-medical images, recognizing new medical images, and recognizing new non-medical images.Firstly, a 2D wavelet transform was applied to obtain three subbands for each participant′s EEG data. Then, an improved integrated local graph structure method was proposed to extract features from the original data and the three subbands. This improved algorithm incorporated extended symmetric local graph structures (ESLGS) and composite local graph structures (CLGS) with stable structures. The features were then normalized to avoid overfitting, and feature matrix columns with correlation coefficients between 0.8 and 1 were selected using Pearson correlation coefficients.The improved algorithm was validated on classifiers such as support vector machines (SVM), and the model was evaluated using four metrics: accuracy, precision, recall, and F1 score. Compared to the original algorithm, the improved algorithm achieved an increase of 3.8%, 0.4%, 0.3%, 1.6%, 5.1%, and 4.2% in classification accuracy on support vector machines for each condition.The classification results indicate significant differences between medical students and non-medical students in the recognition stage of learning medical images. The addition of ESLGS and CLGS demonstrates better classification performance compared to the original algorithm, which utilized VSLGS and SLGS.

Abstract:In view of the influence of nonlinearity, model inaccuracy and complex operating conditions on the control accuracy of the electromagnetic levitation system, an improved sliding mode variable structure control strategy was proposed. Through the modeling, analysis and design of the electromagnetic levitation system, a segmented variable speed approach law is designed based on the position outer loop controller based on the sliding mode variable structure control method and the current inner loop controller based on the PI regulator, which reduces the tracking error of the suspension gap and the jitter of the control system, and introduces the velocity state quantity to reduce the excessive approach velocity of the control system in the early stage of the sliding mode. The simulation and hardware experimental results show that after adopting the improved sliding mode variable structure control strategy, the adjustment time of the system is reduced by 60%, the overshoot is reduced by 77%, and the system has smaller steady-state error and jitter, and has stronger anti-disturbance ability under the same interference situation. This method can not only effectively improve the control performance of electromagnetic levitation system, but also has a good reference value for other nonlinear control systems.

Abstract:This study aims to tackle the challenges encountered by autonomous mobile robots in point-to-point path planning, encompassing issues such as low search efficiency, susceptibility to local optima, and inadequate real-time handling of unknown dynamic and static obstacles. To this end, we have carried out an effective integration of the enhanced A* algorithm with the improved DWA. Within the enhanced A* algorithm, we have introduced obstacle-rate-based weighting factors and a bidirectional optimization strategy, aiming to bolster search efficiency and facilitate the generation of smoother paths. Furthermore, the refined DWA algorithm integrates two novel obstacle evaluation functions and adeptly addresses the local optima issue through the adjustment of weight coefficients. By unifying the enhanced DWA algorithm with the improved A* algorithm, we have enabled proficient real-time obstacle avoidance for unknown dynamic and static obstacles. Simulation results indicate that the improved A* algorithm proposed in this paper, compared with the traditional A* algorithm and the enhanced algorithm from reference ［23］, demonstrates significant performance improvements in four different environments. Specifically, the number of path turns decreased by an average of 30.14% and 18.16%, the search space was reduced by 35.09% and 15.21%, and the planning time was shortened by 82.36% and 38.26%, respectively. Furthermore, when integrated with the improved DWA algorithm, the time required for path planning, the length of the planned path, and the average motion speed were optimized compared to combining with the traditional DWA algorithm and the fusion algorithm from reference ［23］, showing an average reduction of 37.46% and 9.82% in planning time, a decrease of 4.59% and 3.63% in path length, and an increase of 53.49% and 7.09% in average motion speed.

Abstract:As the core device of the electronic valve, the operating state of the micro permanent magnet stepping motor has a direct impact on whether the electronic valve can perform operations normally. In order to accurately grasp the remaining useful life of the motor, a prediction method considering individual differences in the Wiener process is proposed. Firstly, by analyzing the failure process of motor performance degradation, the effective value of phase current is selected as the characteristic quantity of performance degradation. Secondly, due to the simultaneous participation of multiple motors of the same model in the experiment, a motor performance degradation model considering the individual differences in the Wiener process is established, and parameter estimation is carried out based on the EM algorithm.Finally, the number of starting and stopping times is designed as the reference of the motor′s remaining useful life in the experiment and compared with the prediction results based on the traditional Wiener process. The experimental results show that the average prediction error of the proposed method decreases by 3.74%, which has a higher prediction accuracy.

Abstract:In order to meet the requirements of real-time detection of road signs, for the current mainstream target detection algorithms on the image processor there are a large number of model parameters, poor real-time performance, high power consumption and high cost, a real-time detection of road signs based on FPGA is proposed. In order to reduce the number of parameters and improve the detection speed, YOLOv3-tiny is used as the feature extraction network for the training and optimization of the weight parameters; the model floating-point parameters are quantized into 8-bit fixed-point numbers, and the quantized network model is used to complete the deployment experiments on the FPGA. The experimental results show that at the Yolov3-tiny network detection rate, the test frame rate of this system for the experimental dataset can reach 153 fps, the power consumption is 4.92 W, and the peak GOP/s is 115GOP/s. This system can satisfy the requirement of real-time target detection, and it can realize the deployment of the system under low power consumption.

Abstract:A lightweight YOLOv8n-based algorithm for PCB defect detection is proposed to address the trade-off between detection accuracy and model size. Firstly, the large target detection layer is deleted, the small target detection layer is added, and the network structure is adjusted to make the model lightweight and improve detection accuracy. Secondly, the C2f module is combined with GhostConv and DWConv to design the C2f-GhostD module to replace the C2f module, reducing the computational cost of the model. Then, PConv is integrated into the Detect module, resulting in the POne-Detect module, which is applied to the detection network to streamline its structure. Finally, the SimAM attention mechanism is added to the neck network to improve information capture ability. The experimental results show that in the PCB dataset, compared with YOLOv8n, the proposed algorithm reduces the number of parameters by 78.7%, reduces the model size by 73.7%, and improves the mAP0.5 to 98.6%, meeting the hardware deployment requirements of the model.

Abstract:A correction algorithm based on the principle of geometric optics is proposed to address the geometric deformation of part images in machine vision defect detection of polyhedral injection molded parts. Under the condition that the shooting positioning error is not greater than 1 mm, the correction error of the method is theoretically <0.1 mm, which can meet the needs of machine vision defect detection for injection molded parts. Firstly, preprocess the collected images to obtain image edges; Next, the intersection points of the contour lines are determined as part vertices, and different surfaces of the part are segmented based on their positions and mapped onto a two-dimensional plane; Then, calculate the offset of each pixel in the image based on geometric optics; Finally, perform point by point correction on the pixels in the image. Using a set of hexahedral parts to simulate actual working conditions, experiments were conducted under different shooting positioning error states, and the correction algorithm was validated using Matlab. The experimental results show that the error of the proposed method is within 0.1 mm, which is consistent with theoretical analysis and meets the requirements for geometric correction accuracy of part images in machine vision defect detection of injection molded parts.

Abstract:To deal with the issues of poor segmentation accuracy and adaptability in existing thresholding segmentation methods, an image thresholding method guided by maximizing similarity of multi-directional weighted intuitionistic fuzzy is proposed. First, the proposed method utilizes convolution kernels based on first-order derivative of anisotropic Gaussian to perform multi-directional convolution operation and multi-scale product transformation on an input image, which will output four reference images with unimodal histogram in four directions. Then, it constructs the corresponding intuitionistic fuzzy sets by sampling four reference images with a binary contour image. Finally, it utilizes a multi-directional weighting strategy to fuse four intuitionistic fuzzy sets to construct a similarity objective function, and selects the gray level corresponding to the maximum value of this objective function as the segmentation threshold. The proposed method is comprehensively compared with 5 recent segmentation methods, and the experimental results on 8 synthetic images and 88 real-world images show that the proposed method has higher segmentation accuracy and more flexible adaptability, and the average Matthews correlation coefficients are 0.998 and 0.964 for the synthetic images and real-world ones, which outperform the second-best method by 39.90% and 26.22%, respectively.

Abstract:The measurement of key dimensions of power transformers is an important part of their assembly, manufacturing, transportation, and installation processes. Existing measurement methods are cumbersome and inefficient. Therefore, this paper proposes a visual detection method for key dimensions of 110 kV oil-immersed power transformers. This method utilizes the YOLOv5 object detection algorithm and Grabcut image segmentation algorithm to achieve intelligent detection and segmentation of key components. Then, based on the principle of binocular stereo vision, key dimensions such as the distance between bushings and the maximum cross-sectional dimension of the transformer are measured. This paper establishes a scaled model of the appearance of 110 kV oil-immersed power transformers, and analyzes the effects of factors such as shooting distance and angle on visual inspection of key dimensions of power transformers through experiments. The results indicate that this paper achieved visual detection of key dimensions of transformers under different shooting distances and angles based on scaled model experiments, verifying the effectiveness of this method. This method can provide reference for on-site dimension measurement of 110 kV oil-immersed.

Abstract:Tomato yield is affected by diseases, weather and other factors, among which leaf disease is the most critical factor affecting tomato yield. However, in the field of leaf disease detection, the existing models generally have the problem of insufficient generalization ability and high detection rate of small lesions. In this paper, an improved tomato disease early detection algorithm is proposed to improve these problems by optimizing the YOLOv5s network in various aspects, while keeping the model lightweight. Firstly, Mosaic9 data enhancement technology is used to strengthen the detection ability of the model for minor lesions, increase the complexity of the image back-ground, and improve the generalization ability of the model. Secondly, GSConv and Slim-Neck networks are used to lightweight the model and reduce the computational burden while maintaining the accuracy of the model. At the same time, the SimAM attention mechanism was used to capture the features of small lesions on the leaves more accurately, thus reducing the missed detection rate. In addition, in order to further enhance the detection ability of multi-scale targets, adaptive spatial feature fusion is introduced to effectively integrate features of different scales, and improve the detection accuracy of multi-scale targets, especially small targets. The experimental results showed that the model had excellent perfor-mance in early detection of leaf diseases, and the average recognition accuracy, recall rate, F1 score and mAP of leaf mold, early disease, late disease and healthy leaf disease reached 0.951%, 0.918%, 0.934% and 0.948%, respectively. It can be seen that this method has a good detection performance for minor lesions, and improves the problem of insufficient generalization ability of the model and missing detection in the detection process of minor lesions, and further improves the detection effect.

Abstract:As space precision loads increasingly require on-board vibration, the importance of active vibration isolation loads in suppressing micro-vibrations of sensitive loads has gradually become more prominent. In order to realize the collection and suppression of low-frequency and small-amplitude micro-vibrations to meet the vibration environment requirements required by the load, this paper designed a detection system based on FPGA main control board with Qt host computer software. FPGA is used to control multiple Delta-Sigma ADC to complete the synchronous collection of micro-vibration signals, analysis of command protocols, implementation of PID control algorithms and output of drive control signals. At the same time, real-time waveform plotting and spectrum analysis of load acceleration signals are implemented through software. After integrated joint debugging, the system tested the real-time telemetry function and vibration isolation control performance of the active vibration isolation load. The results showed that the acceleration amplitude spectral density integral dropped from 1.73×10-6 g to 1.41×10-7 g, and the vibration isolation suppression ratio reached -25 dB, achieving a good micro-vibration suppression effect and verifying that this system can meet the needs of micro-vibration suppression.

Abstract:To mitigate the issue of positioning deviations in positioning systems caused by non-line of sight (NLOS) errors in Ultra-Wide Ban signals, this study presents an unsupervised clustering method that utilizes the characteristic parameters of the channel impulse response for identifying NLOS signals. The method involves the extraction of eight characteristic parameters from the channel impulse response waveform, followed by the use of the principal component analysis algorithm to reduce the dimension of the multi-dimensional features. An improved K-means clustering algorithm, based on iterative self-organizing data analysis, is then used to select K-values adaptively for distinguishing between LOS and NLOS signals. Finally, the redundancy and correlation of feature parameters are combined to distinguish the classification results. The experimental results demonstrate that this approach effectively identifies NLOS signals with better environmental adaptability and has a recognition accuracy of 95%.

Abstract:The inversion of physical and chemical properties of satellite soil is the most important part of deep space exploration, and thermal properties such as thermal conductivity and heat capacity parameters are the scientific basis for studying the composition of satellite soil, and temperature measurement is an important parameter for in-situ detection of satellite soil based on penetration. In this paper, the surface temperature inversion method of lunar soil probe based on LSTM neural network algorithm is studied to solve the problem that the surface temperature of lunar soil probe can not be measured directly. The penetration process was simulated by ANSYS/LS-DYNA finite element software to obtain the temperature data of multiple groups of reconnaissance warhead. The data were selected according to the finite difference method of discrete heat conduction equation, and the inversion model was established by using the long and short term memory neural network. The root-mean-square error of the inversion curve is 12.9 ℃ and the maximum relative error is less than 10% compared with the experimental curve. The experimental results show that the method proposed in this paper can realize the inversion of the outer surface temperature of the probe.

Abstract:Aiming at the problems of long start-up time, slow response speed and poor stability when the load changes in the hammer mill control system in the feed processing industry. A PID control method based on BP neural network algorithm is proposed. Firstly, establishing the reference model of the transfer function of the combined system of the frequency converter and the hammer mill drive motor and analyzing its stability. Then, based on the analysis of conventional PID and fuzzy PID control algorithms, the adaptive neural network algorithm PID is applied to the control process of the hammer mill drive system. By building a simulation model for the control motor of the hammer mill and it is simulated and analyzed by the Simulink graphical programming function. And based on LABVIEW software, a testing platform for the hammer mill measurement and control system was built for experimental testing and analysis.The results show that the designed BP neural network PID controller can achieve good adaptive tracking for the speed reference model given by the feed crushing system, with faster response to step signals, smaller overshoot, and stronger anti-interference ability. The designed adaptive controller can automatically adjust PID parameters according to changes in working conditions, resulting in an average reduction of 5.16% in electricity consumption per ton of material and an average increase of 2.08% in productivity, The control of the spindle speed of the hammer mill is more precise, with smaller errors, and has high control accuracy and strong robustness, meeting the adaptive control requirements of the feed hammer mill drive system.