Intelligent control system and method for forklift driving

Abstract

The invention discloses a forklift driving intelligent control system and method, and relates to the technical field of forklift control, and the method comprises the steps: collecting and preprocessing multi-source time sequence data of a forklift under all working conditions; constructing a local electromagnetic interference intensity index in each monitoring window; a milemeter decoupling index is constructed in the lifting action process; after each lifting action, calculating a projection centroid position based on the on-fork weight, the fork frame height, the vehicle transverse and longitudinal inclination angle and the inter-wheel load, so as to construct a ramp centroid significance index; and taking the local electromagnetic interference intensity index, the odometer decoupling index and the ramp centroid significance index as input, labeling data points at each moment on the monitoring time sequence, judging whether to trigger a forklift adaptive control algorithm or not based on a random forest algorithm, and outputting forklift control actions. The method solves the problem that when the project angle problem appears in an intermittent and sudden mode, a traditional control and sensing system is difficult to effectively recognize and self-adapt, and therefore serious potential safety hazards are generated.

Description

Technical Field

[0001] This invention relates to the field of forklift control technology, specifically to an intelligent control system and method for forklift drive. Background Technology

[0002] In modern warehousing and logistics systems, forklifts undertake high-density, repetitive handling tasks. To improve operational efficiency and safety, forklift drive systems are evolving towards intelligence and closed-loop systems, with typical solutions integrating sensors and upper-level task scheduling logic. However, in practical applications, several engineering challenges have long constrained the reliability and safety of the system, especially when these problems occur intermittently and suddenly. Traditional control and sensing systems struggle to effectively identify and adapt to these challenges, leading to serious safety hazards. These include: 1) When modern forklifts are driven by both electric and hydraulic systems, the coil PWM and the start / stop of the large motor often generate transient electromagnetic interference (EMI) with high di / dt, which manifests as CAN frame errors, ADC baseline drift, or short-term controller reset. Because EMI is mostly intermittent, short-lived, and has a complex spectrum, traditional filtering and simple shielding are difficult to completely eliminate it. Moreover, it can cause the loss of sensor / communication information at critical lifting moments, leading to safety hazards such as hydraulic malfunctions or material falling.

[0003] 2) When the forklift completes lifting and lowering and returns the motor shaft to the walking state, mechanical backlash, springback, or brief free rotation may cause the mileage encoder to generate invalid pulses or lose steps, resulting in jumps or accumulated errors in mileage-based positioning.

[0004] 3) When operating on ramps or inclined surfaces, the instantaneous displacement of the center of gravity of the cargo or forklift will significantly reduce the vehicle's stability margin. Mileage / attitude estimation based solely on the plane assumption often underestimates the risk of rollover.

[0005] Therefore, the present invention provides a forklift drive intelligent control system and method. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent control system and method for forklift drive to solve the existing problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a forklift drive intelligent control system, comprising: The data processing module is used to collect and preprocess multi-source time-series data of the forklift under all operating conditions; The transient electromagnetic interference index calculation module is used to construct a local electromagnetic interference intensity index based on PWM activity, coil current transients, main motor current spikes, CAN error rate, and ADC baseline offset within each monitoring window; A wheel speed continuity monitoring module is configured to construct a wheel speed decoupling index based on the residual error between the pump shaft encoder or Hall signal, coil holding current, and wheel encoder or wheel speed during the lifting action. A center of mass mutation and ramp stability evaluation module is configured to calculate the projected center of mass position based on the on-fork weighing, fork height, vehicle roll and pitch angle, and inter-wheel load after each lifting action, and then obtain a stability margin, thereby constructing a ramp center of mass significance index. A predictive control module is configured to take the local electromagnetic interference intensity index, wheel speed decoupling index, and ramp center of mass significance index as inputs, label the data points at each time point in the monitoring time sequence, determine whether to trigger the forklift adaptive control algorithm based on a random forest algorithm, and output the forklift control action.

[0008] The data preprocessing module is further improved to record the following time series signals at a high sampling frequency and construct a multi-channel raw time series dataset indexed by time stamp, including the PWM and current channels, CAN error and delay statistics, IMU and wheel speed encoder data, and to perform integral or window average method for integral and normalization processing on the current and energy channels.

[0009] The construction process of the local electromagnetic interference intensity index includes: taking a monitoring window with a length of as the center at the t th time point, L represents the set acquisition time window, and an electromagnetic monitoring sequence containing N channels is constructed; the CAN error count, ADC offset, PWM energy, and motor current sequences are respectively linearly fitted and the fitting residual sequence is calculated; the local electromagnetic interference index of the window is calculated based on the fitting residual sequence ; the calculation formula of the local electromagnetic interference index is represented as: wherein, represents the local electromagnetic interference index at the t th time point; represents the maximum value of the absolute values of all elements in the fitting residual sequence, represents the standard deviation of the fitting residual sequence, represents the mean value of the absolute values of all elements in the fitting residual sequence; represents the total number of CAN error frames or error rate within the window; , represents a weight parameter for adjusting the contribution of different channels to the EMI index.

[0010] The application further improves that the odometer continuity monitoring module takes the encoder timing before and after the lift start as a monitoring window, least squares the accumulated pulses of the wheel encoder and the IMU+vision-based odometer estimation to obtain a fitting residual sequence; takes the linkage confirmation signal of the pump shaft encoder current as a secondary verification channel; and constructs an odometer decoupling index based on the above sequence , for indicating the degree of discontinuity of the odometer reading in the lift stage and the degree of reduction in reliability.

[0011] The application further improves that the odometer decoupling index is calculated as follows: , wherein, represents the odometer decoupling index at the t-th moment, represents the mean value of the fitting residual of the wheel encoder and the IMU / visual odometer; represents the range of the wheel encoder pulses in the monitoring window; represents the range of the pump shaft / coupling encoder in the same window; is a zero hyperparameter.

[0012] The application further improves that the ramp centroid significance index calculation formula is represented as: , wherein, represents the centroid significance index of the k-th peak point in the peak neighborhood; represents the local centroid offset at the corresponding peak moment; represents the neighborhood width, represents the time distance of the s-th sample from the peak moment; represents the centroid offset of the s-th moment in the peak neighborhood.

[0013] The application further improves that the feature vector input by the prediction control module is taken as a training sample of a random forest algorithm, a group of decision tree models are trained for real-time classification and risk prediction; when it is determined to be low risk and , , all vectors are less than the corresponding set threshold, normal action is allowed and the conventional control law is executed; when it is determined to be medium risk and , , all vectors are less than the corresponding set threshold, protective constraints are executed; when it is determined to be high risk or , , If any vector is greater than or equal to the corresponding set threshold, a hard safety action is immediately executed.

[0014] A further improvement of this invention is that the protective constraints include a forklift adaptive control algorithm implemented based on a lightweight neural network. This algorithm uses the local electromagnetic interference intensity index, the odometer decoupling index, and the ramp centroid significance index as inputs to construct a comprehensive state vector reflecting electromagnetic environment stability, driving estimation reliability, and attitude stability. The lightweight neural network, equipped with a model adaptive feedback strategy, performs nonlinear mapping on the input features, outputting the forklift's driving speed, steering angle, and hydraulic control quantities. The output signals are then linearly mapped and transmitted to the driving, steering, and hydraulic actuation actions to achieve adaptive control of the forklift.

[0015] A further improvement of this invention is that the model adaptive feedback strategy includes training the model during the training phase using a sample-weighted loss function, and calculating the historical sensitivity vector using the correlation coefficient. For samples at time t, construct sample weights. , The normalized local electromagnetic interference intensity index, odometer decoupling index, and ramp centroid significance index are represented by these. Robust loss is used during both training and inference phases, and the gradient pruning threshold is dynamically set based on real-time metrics during backpropagation updates. ,in, Based on the gradient threshold, This represents the scaling factor associated with each respective metric. This represents the normalized value of the indicator; The learning rate of the lightweight neural network model is then dynamically adjusted based on the centroid significance index and the local electromagnetic interference index. The response indicator is expressed as ; in, Indicates the basic learning rate. Indicates the interference sensitivity coefficient; Represents the normalized local electromagnetic interference index. This represents the sensitivity coefficient for the significance of the normalized centroid; When a model performance degradation exceeding a preset threshold is detected, the last layer of the lightweight neural network is triggered with the set minimum learning rate. Perform fine-tuning in a limited number of steps.

[0016] On the other hand, the present invention provides a forklift drive intelligent control method, comprising the following steps: S1. Collect and preprocess multi-source time-series data of forklifts under all operating conditions; S2. Within each monitoring window, a local electromagnetic interference intensity index is constructed based on PWM activity, coil current transients, main motor current spikes, CAN error rate, and ADC baseline offset. S3. During the lifting action, the odometer decoupling index is constructed based on the pump shaft encoder or Hall signal, the coil holding current, and the residual between the wheel encoder or odometer. S4. After each lifting action, the projected centroid position is calculated based on the fork load, fork height, vehicle lateral and longitudinal tilt angles and wheel load, and the stability margin is obtained, thereby constructing the slope centroid significance index. S5. Using the local electromagnetic interference intensity index, odometer decoupling index, and ramp centroid significance index as inputs, label the data points at each moment in the monitoring time series, and based on the random forest algorithm, determine whether the forklift adaptive control algorithm is triggered, and output the forklift control action.

[0017] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention first constructs a local electromagnetic interference intensity index by acquiring multi-source timing data such as PWM activity, coil current transients, motor current spikes, CAN error rate, and ADC baseline offset at high frequency. This index can quantify the transient EMI intensity caused by coil switching, PWM abrupt changes, and motor start-stop in real time and accurately, thus solving the problems of unpredictability of EMI and distortion of key signals. 2. By monitoring the wheel encoders during the lifting process and comparing and fitting them with the mileage estimation based on IMU + vision, and introducing the pump shaft encoder as a secondary verification of the linkage state, the odometer decoupling index is constructed, which solves the positioning drift and control error problems caused by mechanical backlash, springback or slippage. 3. By integrating multi-dimensional information such as fork load, fork height, vehicle tilt angle, and wheel load in real time during lifting and lowering operations, and by constructing a ramp center of gravity significance index, it can capture the degree of abrupt changes in the center of gravity rather than just the position itself; thus solving the problem of overturning risk caused by underestimating abrupt changes in the center of gravity when lifting and lowering on complex inclined ground. 4. Based on the model adaptive feedback strategy, including the weighted loss function with sample weighting, dynamic gradient pruning, dynamic learning rate adjustment and online fine-tuning mechanism, the problems of difficult training, unstable inference and performance degradation of neural network models under strong interference, high dynamic and sample imbalance conditions are solved. Attached Figure Description

[0018] Figure 1 This is a framework diagram of a forklift drive intelligent control system according to the present invention; Figure 2 This is a flowchart of a forklift drive intelligent control method according to the present invention. Detailed Implementation

[0019] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0020] The term "and / or" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three cases: A exists alone, A and B exist simultaneously, and B exists alone.

[0021] Example 1 Figure 1 This embodiment illustrates a framework diagram of a forklift drive intelligent control system, including: The data processing module is used to collect and preprocess multi-source timing data of the forklift under all working conditions. Working conditions include lifting / driving / ramp, etc. The multi-source timing data includes, but is not limited to: CAN error frame count and frame delay, switching coil PWM waveform and current sampling, main drive motor current / voltage, pump shaft encoder / Hall effect sensor, wheel speed encoder, IMU (accelerometer and gyroscope), tilt sensor, fork height sensor, fork weighing / torque sensor, hydraulic pressure sensor and ADC baseline reading. The data preprocessing module records the following timing signals at a high sampling frequency and constructs a multi-channel raw timing dataset indexed by timestamps. This dataset includes PWM and current channels sampled at ≥1kHz (for capturing transients), CAN error and delay statistics sampled at ≥100Hz, and IMU and wheel speed encoder data sampled at ≥200Hz. The current and energy channels are integrated and normalized using integration or in-window averaging methods.

[0022] The transient electromagnetic interference (EMI) index calculation module addresses the challenges of hydraulic linkage and motor drive in forklifts. During operation, the controller typically uses PWM to drive the coil and control the inverter to rotate the motor forward and backward to power the hydraulic pump. The switching on and off of the coil, sudden changes in the PWM duty cycle, and the start-stop and reversal of the large motor generate high transient currents, switching reverse spikes, and ground / return potential fluctuations, resulting in strong common-mode / differential-mode electromagnetic interference (EMI) on the vehicle's power supply and ground lines. The engineering consequences of this transient EMI include: CAN bus frame errors or loss of synchronization, transceiver resets, analog ADC baseline drift or reading jitter, and even transient controller restarts or task loss. EMI is often intermittent, short-lived, and has a complex spectrum with unpredictable timing windows, making it difficult to completely eliminate with traditional filtering or simple hardware shielding. Furthermore, when EMI coincides with lifting or critical operation moments, brief communication interruptions or abnormal sensor readings can cause the controller to lose position, pressure, or attitude information, triggering erroneous power outages or brake releases, potentially leading to hydraulic runaway or material falling, posing serious safety risks.

[0023] Therefore, a local electromagnetic interference intensity index is constructed within each monitoring window based on PWM activity, coil current transients, main motor current spikes, CAN error rate, and ADC baseline offset. The process of constructing the local electromagnetic interference intensity index includes: taking time t as the center, and taking a length of... The monitoring window is configured to form an electromagnetic monitoring sequence containing N channels; linear fitting is performed on the sequences of CAN error count, ADC offset, PWM energy, and motor current, and the fitting residual sequence is calculated; based on the fitting residual sequence, the local electromagnetic interference index of the window is calculated. Used to determine the presence of transient EMI and whether its intensity may interfere with the bus / sensor; the formula for calculating the local electromagnetic interference index is expressed as: in, This represents the local electromagnetic interference index at time t. This represents the maximum absolute value of all elements in the fitted residual sequence. This represents the standard deviation of the fitted residual sequence. This represents the mean of the absolute values ​​of all elements in the fitted residual sequence; This indicates the total number or error rate of CAN error frames within the window. , This represents the weighting parameter, used to adjust the contribution of different channels to the EMI index.

[0024] This embodiment, based on linear fitting of a window and residual calculation, emphasizes short-term abrupt changes and nonlinear transients, enabling the system to distinguish between normal trends (slow changes) and transient interference (sudden pulses); it improves the detection sensitivity of transient EMI and reduces misjudgment of normal signal trends. This is achieved by recording and utilizing the maximum residual value. It can quickly quantify high-amplitude burst pulses; rapidly trigger protection logic (e.g., temporarily isolate CAN, delay critical actions), reducing the risk of malfunctions caused by bursts. This is achieved through averaging. with standard deviation The combination of these features expresses the overall instability of the residual distribution: it takes into account both persistent and discrete disturbance characteristics, avoids misleading judgments by a single extreme value, and improves the detection capability for persistent micro-perturbations or multiple small-amplitude disturbances.

[0025] The odometer continuity monitoring module is used to construct the odometer decoupling index based on the pump shaft encoder or Hall signal, coil holding current, and residual between the wheel encoder or odometer during the lifting operation. The odometer continuity monitoring module uses the encoder timing before and after the lifting start as the monitoring window, and performs least-squares fitting between the cumulative pulses of the wheel encoder and the odometer estimation based on IMU+vision to obtain the fitting residual sequence. The linkage confirmation signal of the pump shaft encoder current is used as a secondary verification channel; the odometer decoupling index is constructed based on the above sequence. This is used to indicate the degree of discontinuity and reduced reliability of odometer readings during the ascent and descent phases.

[0026] The method for calculating the odometer decoupling index is as follows: ,in, This represents the odometer decoupling index at time t. This represents the mean of the fitting residuals between the wheel encoder and the IMU / visual odometry. This indicates the range of wheel encoder pulses within the monitoring window; This indicates the range of the pump shaft / coupling encoder within the same window (if the pump shaft has no encoder, the starting / holding current level of the coil current is used as a substitute indicator). To prevent zero-parameter overshoot, when the wheel encoder generates abnormal pulses but the pump shaft does not change accordingly, An increase indicates a decrease in the reliability of the odometer.

[0027] This embodiment captures the consistency between the odometer and the reference estimate by using the wheel encoder and visual fitting residual. It can identify step counting errors or pulse skipping caused by changes in drive coupling during the lifting linkage stage, thereby avoiding bringing erroneous mileage information into the positioning fusion. A pump shaft encoder / coil current is introduced as a secondary verification method. By determining whether the drive side is actually engaged, legitimate mileage pulses are distinguished from idle or erroneous pulses. This avoids mistaking wheel noise when the pump is not engaged for actual displacement, thus improving detection accuracy. Furthermore, through… Amplify the residual signal during the activity; when the wheel is stationary or moving very little, the residual is meaningless; when the wheel is actually moving but the residual is large, it indicates that the reliability of the mileage reading has decreased, and a rapid response should be taken.

[0028] By driving activity If the driving activity is large (indicating that coupling should occur), but the residual is still large, it indicates that there is indeed an anomaly; if the driving activity is small but the wheels show pulses, the ratio will be amplified, indicating decoupling or abnormal idling.

[0029] The centroid mutation and ramp stability assessment module is used to estimate the projected centroid position and calculate the stability margin after each lifting action based on the fork weighing, fork height, vehicle lateral / longitudinal tilt angle and wheel load, thereby constructing the ramp centroid significance index. When a forklift performs lifting operations on ramps or sloping ground, the slope alters the component of gravity, and the longitudinal or lateral displacement (or slippage) of the cargo during lifting can cause abrupt changes in the vehicle's center of gravity. Relying solely on planar assumptions or insufficient mileage and single tilt angle estimates may underestimate instantaneous stability margins and fail to accurately assess the rollover threshold. Practical consequences include near-instability events such as side tilting or overturning during ramp lifting, turning, or braking, leading to deviations between attitude estimates and actual angles, and consequently misjudging the executable speed / angular velocity and lifting height.

[0030] The formula for calculating the centroid significance index of the ramp is as follows: ;in, This represents the centroid significance index of the k-th peak point within the peak neighborhood; This represents the local centroid offset or asymmetric offset measure at the corresponding peak time. Indicates the neighborhood width. This represents the time distance between the s-th sample and the peak time. This represents the centroid shift metric at time s within the peak neighborhood. If the centroid shift at a certain time is unusually prominent within the neighborhood, then... Large, indicating a potential risk of capsizing.

[0031] By identifying sudden shifts through multi-scale peak detection (AMPD, etc.), we can distinguish between continuous slow drift and short-term sharp shift, reduce false alarms, and ensure high recall for dangerous sudden events.

[0032] Peak significance is measured by the ratio of the peak value to the weighted sum of its neighborhood values. A peak value is significant when the offset at a given moment is significantly greater in its neighborhood than in its surrounding neighborhood. The system can accurately determine a high-risk moment and trigger protective measures such as prohibiting further ascent, deceleration, or mechanical locking; it uses time-distance weighting. It emphasizes the context of recent moments and suppresses the influence of historical noise from distant moments in the neighborhood, thus more accurately reflecting the importance of current mutations.

[0033] The predictive control module takes the local electromagnetic interference intensity index, odometer decoupling index, and ramp centroid significance index as inputs, and labels the data points at each moment in the monitoring time series, including normal / communication interference / mileage discontinuity / centroid anomaly / high-risk instability, among which communication interference / mileage discontinuity / centroid anomaly are low-risk; based on the random forest algorithm, it determines whether to trigger the forklift adaptive control algorithm and outputs the forklift control action.

[0034] The feature vector input to the prediction control module serves as the training sample for the random forest algorithm, training a set of decision tree models for real-time classification and risk prediction; during online operation, the module outputs classification probabilities to the real-time feature vectors. ; ; When determined to be low risk and , , When all vectors are less than the corresponding set threshold, normal operation is allowed and the conventional control law is executed. When it is determined to be medium risk and , , When all vectors are less than the corresponding set threshold, protective constraints are applied. When judged as high risk or , , When any vector is greater than or equal to the corresponding set threshold, a hard safety action is immediately executed. For example: triggering mechanical locking (cutting off the lifting command, activating the locking valve); forcibly stopping / decelerating to a safe position; isolating CAN (if EMI is suspected to be severe) and switching to independent hardware safety PLC control; and sending a manual intervention request (buzzer + remote alarm).

[0035] The protective constraints include a forklift adaptive control algorithm implemented based on a lightweight neural network. This algorithm uses the local electromagnetic interference intensity index, odometer decoupling index, and ramp centroid significance index as inputs to construct a comprehensive state vector reflecting electromagnetic environment stability, driving estimation reliability, and attitude stability. The lightweight neural network, equipped with a model adaptive feedback strategy, performs nonlinear mapping on the input features, outputting the forklift's driving speed, steering angle, and hydraulic control quantities. The output signals are then linearly mapped and transmitted to the driving, steering, and hydraulic actuation actions to achieve adaptive control of the forklift. The model adaptive feedback strategy includes training the model during the training phase using a sample-weighted loss function. when A high value indicates that communication / ADC may be disturbed, and there is a high risk of data noise and false signals. During training, blind strong learning on these samples should be reduced, and during inference, the uncertainty threshold should be increased and automatic actions should be triggered more conservatively.

[0036] when If the mileage information is too high, the reliability of the mileage information will decrease. Therefore, the weights of mileage-related features should be reduced, the weights of visual and IMU features should be increased, and the correctness of detecting mileage anomalies should be emphasized in the training loss.

[0037] when A high score indicates that the control safety boundary is close, and the model's sensitivity to such samples should be improved, with priority given to ensuring recall (reducing false negatives).

[0038] Therefore, based on delivery accuracy, a historical sensitivity vector is calculated using the correlation coefficient. For samples at time t, construct sample weights. , The normalized local electromagnetic interference intensity index, odometer decoupling index, and ramp centroid significance index are represented by these. If a certain type of index has a significant impact on delivery accuracy, that is... If the index is large and the current sample has a high index, then the sample has a large weight, and the model will pay more attention to this type of sample, especially for rare but critical centroid outlier samples. Robust loss (e.g., Huber loss) is used during training and inference to reduce the impact of outliers on the gradient; during backpropagation updates, the gradient pruning threshold is dynamically set based on real-time metrics. ,in, Based on the gradient threshold, The scaling factor (hyperparameter) associated with each of the aforementioned metrics, and This is the normalized value of the indicator; when the indicator value increases (indicating an increase in the noise / anomaly level of the current sample or batch), the... The size of the batch size should be reduced accordingly to suppress potential gradient explosion or anomalous updates during a single backpropagation. For very small batches (online deployment), use LayerNorm / GroupNorm instead of BatchNorm to avoid instability caused by small batches. Using gradient accumulation to simulate larger batches helps smooth gradients.

[0039] The mean and variance of the gradient norm are recorded online at each step; if an anomaly occurs (norm > threshold or variance increases suddenly), online fine-tuning is paused and the model is rolled back to the last checkpoint.

[0040] because It represents the intensity of noise interference at the data channel level (directly altering the input data and measurement baseline), and is the main culprit causing abnormal gradients and noise-driven parameter updates. Therefore, it is usually given priority in learning rate control strategies, and lr is compared with... Linkage can directly suppress training instability caused by external electromagnetic transients.

[0041] The learning rate of the lightweight neural network model is then dynamically adjusted based on the centroid significance index and the local electromagnetic interference index. The response indicator is expressed as ; To achieve reduced light intensity (LR) when noise is high; and to increase attention while maintaining low LR when centroid anomalies are high, whereby... Indicates the basic learning rate. The interference sensitivity coefficient represents the degree to which electromagnetic interference affects the learning rate. (The higher the noise level, the lower the lr value). Represents the normalized local electromagnetic interference index. This represents the sensitivity coefficient for the significance of the normalized centroid; (When the centroid is high, slightly increase the learning step size to learn this type of sample quickly); if the two conflict, prioritize reducing the learning rate.

[0042] When a feature distribution shift (e.g., KL divergence or feature statistical abrupt change) is detected within the sliding window, or when model performance degrades beyond a preset threshold, a low learning rate is applied only to the last layer of the network. Perform fine-tuning in a limited number of steps (e.g., 1-5 times), using EWC (Elastic Weight Consolidation) or L2 regularization terms during fine-tuning to prevent catastrophic forgetting; Normalized Local Electromagnetic Interference Intensity Index Greater than the threshold At the same time, it automatically applies stronger median or window filtering to the raw CAN / ADC data at the input layer, and uses robust loss such as Huber or Tukey during backpropagation to reduce the impact of outliers on the gradient. After offline training, the network undergoes quantization (e.g., int8) and structured pruning (the pruning ratio is adjusted according to the tradeoff between deployment accuracy and latency), and during inference, the latency of a single inference is guaranteed to be less than a preset threshold to meet real-time control requirements.

[0043] The threshold and weight settings can be set by default according to the present invention, or they can be set by those skilled in the art.

[0044] Example 2 Figure 2 This invention illustrates a flowchart of a forklift drive intelligent control method. Based on the same inventive concept as Embodiment 1, this invention provides a forklift drive intelligent control system, comprising: S1. Collect and preprocess multi-source time-series data of forklifts under all operating conditions; S2. Within each monitoring window, a local electromagnetic interference intensity index is constructed based on PWM activity, coil current transients, main motor current spikes, CAN error rate, and ADC baseline offset. S3. During the lifting action, the odometer decoupling index is constructed based on the pump shaft encoder or Hall signal, the coil holding current, and the residual between the wheel encoder or odometer. S4. After each lifting action, the projected centroid position is calculated based on the fork load, fork height, vehicle lateral and longitudinal tilt angles and wheel load, and the stability margin is obtained, thereby constructing the slope centroid significance index. S5. Using the local electromagnetic interference intensity index, odometer decoupling index, and ramp centroid significance index as inputs, label the data points at each moment in the monitoring time series, and based on the random forest algorithm, determine whether the forklift adaptive control algorithm is triggered, and output the forklift control action.

[0045] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0046] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0047] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0048] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0049] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A forklift drive intelligent control system, characterized in that: include: The data processing module is used to collect and preprocess multi-source time-series data of the forklift under all operating conditions; The transient electromagnetic interference index calculation module is used to construct a local electromagnetic interference intensity index based on PWM activity, coil current transients, main motor current spikes, CAN error rate, and ADC baseline offset within each monitoring window; The odometer continuity monitoring module is used to construct the odometer decoupling index based on the pump shaft encoder or Hall signal, coil holding current, and residual between the wheel encoder or odometer during the lifting operation. The centroid mutation and ramp stability assessment module is used to calculate the projected centroid position based on the fork load, fork height, vehicle lateral and longitudinal tilt angles and wheel load after each lifting action, thereby obtaining the stability margin and constructing the ramp centroid significance index. The predictive control module takes the local electromagnetic interference intensity index, odometer decoupling index, and ramp centroid significance index as inputs, labels the data points at each moment in the monitoring time series, and determines whether to trigger the forklift adaptive control algorithm based on the random forest algorithm, and outputs the forklift control action.

2. The forklift drive intelligent control system according to claim 1, characterized in that: The data preprocessing module uses a high sampling frequency to record the following timing signals and constructs a multi-channel raw timing dataset indexed by timestamps, including PWM and current channels, CAN error and delay statistics, IMU and wheel speed encoder data, and uses integration or in-window averaging to perform integration and normalization processing on the current and energy channels.

3. The forklift drive intelligent control system according to claim 2, characterized in that: The process of constructing the local electromagnetic interference intensity index includes: taking time t as the center, and taking a length of... The monitoring window, where L represents the set acquisition time window, constitutes an electromagnetic monitoring sequence containing N channels; linear fitting is performed on the sequences of CAN error count, ADC offset, PWM energy, and motor current, and the fitting residual sequence is calculated; based on the fitting residual sequence, the local electromagnetic interference index of this window is calculated. The formula for calculating the local electromagnetic interference index is as follows: in, This represents the local electromagnetic interference index at time t. This represents the maximum absolute value of all elements in the fitted residual sequence. This represents the standard deviation of the fitted residual sequence. This represents the mean of the absolute values ​​of all elements in the fitted residual sequence; This indicates the total number or error rate of CAN error frames within the window. , This represents the weighting parameter, used to adjust the contribution of different channels to the EMI index.

4. The forklift drive intelligent control system according to claim 3, characterized in that: The odometer continuous monitoring module uses the encoder timing before and after the lifting start as the monitoring window, performs least squares fitting between the cumulative pulses of the wheel encoder and the mileage estimation based on IMU+vision, and obtains the fitting residual sequence; the linkage confirmation signal of the pump shaft encoder current is used as the secondary verification channel. Based on the above sequence, an odometer decoupling index is constructed. This is used to indicate the degree of discontinuity and reduced reliability of odometer readings during the ascent and descent phases.

5. The forklift drive intelligent control system according to claim 4, characterized in that: The method for calculating the odometer decoupling index is as follows: ,in, This represents the odometer decoupling index at time t. This represents the mean of the fitting residuals between the wheel encoder and the IMU / visual odometry. This indicates the range of wheel encoder pulses within the monitoring window; This indicates the range of the coupling encoder within the same window; To prevent zero-over-parameter errors.

6. The forklift drive intelligent control system according to claim 5, characterized in that: The formula for calculating the centroid significance index of the ramp is as follows: in, This represents the centroid significance index of the k-th peak point within the peak neighborhood; This represents the local centroid offset at the corresponding peak time. Indicates the neighborhood width. This represents the time distance between the s-th sample and the peak time. This represents the centroid offset metric at time s within the peak neighborhood.

7. The forklift drive intelligent control system according to claim 6, characterized in that: The feature vector input to the prediction control module is used as the training sample for the random forest algorithm, and a set of decision tree models is trained for real-time classification and risk prediction. When determined to be low risk and , , When all vectors are less than the corresponding set threshold, normal operation is allowed and the conventional control law is executed. When it is determined to be medium risk and , , When all vectors are less than the corresponding set threshold, protective constraints are applied. When judged as high risk or , , If any vector is greater than or equal to the corresponding set threshold, a hard safety action is immediately executed.

8. The forklift drive intelligent control system according to claim 7, characterized in that: The protective constraints include a forklift adaptive control algorithm based on a lightweight neural network. This algorithm uses the local electromagnetic interference intensity index, odometer decoupling index, and ramp centroid significance index as inputs to construct a comprehensive state vector reflecting electromagnetic environment stability, driving estimation reliability, and attitude stability. The lightweight neural network, equipped with a model adaptive feedback strategy, performs nonlinear mapping on the input features, outputting the forklift's driving speed, steering angle, and hydraulic control quantities. The output signals are then linearly mapped and transmitted to the driving, steering, and hydraulic actuation actions to achieve adaptive control of the forklift.

9. A forklift drive intelligent control system according to claim 8, characterized in that: The model adaptive feedback strategy includes training the model using a sample-weighted loss function during the training phase, and calculating the historical sensitivity vector using the correlation coefficient. For samples at time t, construct sample weights. , The normalized local electromagnetic interference intensity index, odometer decoupling index, and ramp centroid significance index are represented by these. Robust loss is used during both training and inference phases, and the gradient pruning threshold is dynamically set based on real-time metrics during backpropagation updates. ,in, Based on the gradient threshold, This represents the scaling factor associated with each respective metric. This represents the normalized value of the indicator; The learning rate of the lightweight neural network model is then dynamically adjusted based on the centroid significance index and the local electromagnetic interference index. The response indicator is expressed as ; in, Indicates the basic learning rate. Indicates the interference sensitivity coefficient; Represents the normalized local electromagnetic interference index. This represents the sensitivity coefficient for the significance of the normalized centroid; When a model performance degradation exceeding a preset threshold is detected, the last layer of the lightweight neural network is triggered with the set minimum learning rate. Perform fine-tuning in a limited number of steps.

10. A forklift drive intelligent control method, used to execute a forklift drive intelligent control system as described in any one of claims 1-9, characterized in that: Includes the following steps: S1. Collect and preprocess multi-source time-series data of forklifts under all operating conditions; S2. Within each monitoring window, a local electromagnetic interference intensity index is constructed based on PWM activity, coil current transients, main motor current spikes, CAN error rate, and ADC baseline offset. S3. During the lifting action, the odometer decoupling index is constructed based on the pump shaft encoder or Hall signal, the coil holding current, and the residual between the wheel encoder or odometer. S4. After each lifting action, the projected centroid position is calculated based on the fork load, fork height, vehicle lateral and longitudinal tilt angles and wheel load, and the stability margin is obtained, thereby constructing the slope centroid significance index. S5. Using the local electromagnetic interference intensity index, odometer decoupling index, and ramp centroid significance index as inputs, label the data points at each moment in the monitoring time series, and based on the random forest algorithm, determine whether the forklift adaptive control algorithm is triggered, and output the forklift control action.

Images