A forklift heat balance automatic control method

Abstract

This invention relates to an automatic thermal balance control method for forklifts, belonging to the field of forklift thermal management technology. The method includes: acquiring historical operating conditions and temperature data of the forklift to train an LSTM neural network and obtain a temperature prediction model; collecting real-time operating condition data for a preset time period before the current moment and inputting it into the model to obtain the predicted component temperature at a preset future moment; determining the target control mode of the cooling system based on a dual-dimensional judgment of the component's real-time temperature and predicted temperature; generating control commands based on the target control mode and sending them to the cooling actuator to adjust its operating parameters. This invention achieves accurate prediction of future temperatures through deep mining of time-series operating condition data using an LSTM network. Based on the dual judgment of "real-time + predicted" temperature, it intelligently allocates multiple control paths such as predictive intervention and forced full cooling, solving the problems of delayed response and untimely heat dissipation in traditional threshold control and achieving accurate prediction and dynamic adjustment of thermal balance.

Description

Technical Field

[0001] This invention relates to the field of forklift thermal management technology, specifically to an automatic control method for forklift thermal balance. Background Technology

[0002] Forklifts are indispensable loading and unloading equipment in logistics warehousing, ports, and factory workshops, operating under high intensity and complex, variable conditions. Whether traditional internal combustion forklifts or the rapidly growing electric forklifts, their power systems (engines or motors), hydraulic systems, and battery systems generate a significant amount of heat during operation. If this heat is not dissipated promptly or effectively managed, it will lead to an imbalance in the vehicle's thermal balance, directly causing a series of problems such as engine overheating, decreased hydraulic oil viscosity leading to increased leakage, reduced motor efficiency, accelerated battery aging, and even thermal runaway. These issues severely impact the forklift's operating efficiency, equipment lifespan, and operational safety.

[0003] Early forklift cooling designs relied heavily on passive radiators (such as water tanks and oil coolers) coupled with fans operating at a fixed speed. For internal combustion forklifts, the high-temperature return oil from the hydraulic system typically flows directly back to the oil tank, relying solely on natural heat dissipation from the tank walls. This resulted in extremely low cooling efficiency, easily leading to oil temperature spikes and damage to hydraulic components under high-pressure, high-frequency operation. While some improvements added throttling devices or independent oil temperature radiators, their control logic was simple, often relying on mechanical switches or manual judgment, making it impossible to dynamically adjust cooling capacity based on real-time operating conditions. Most existing thermal balance control methods are based on "feedback regulation," meaning that cooling equipment (such as fans and water pumps) is only activated to lower the temperature after a temperature sensor detects that the temperature of a component (such as the battery or hydraulic oil) has exceeded a threshold. This lag in control leads to large temperature fluctuations, and in certain secondary operating conditions or continuous high-load operations, the thermal balance system is prone to instantaneous collapse because the response speed of the cooling system cannot keep up with the rate of temperature rise. Although wide-temperature-range calibration technology for batteries has emerged in recent years, there is still a technological gap at the vehicle level in terms of "feedforward prediction" and "cooperative control" for thermal balance under different operating conditions (such as full-load climbing, high-frequency lifting, and standby charging). Therefore, an automatic thermal balance control method for forklifts is proposed to address the above problems. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an automatic thermal balance control method for forklifts. The system can accurately predict the temperature trend at a preset time in the future based on historical operating condition sequences before the actual temperature rises, thereby advancing the control timing from "starting after overheating" to "intervening before overheating". This achieves accurate prediction and active suppression of sudden temperature rises, significantly improving the thermal balance robustness of forklifts under harsh operating conditions.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an automatic control method for forklift thermal balance, comprising the following steps: S1. Collect the operating condition data sequence and corresponding component temperature data sequence of the forklift during historical operating periods to build a training sample set; S2. Collect multi-source data in real time to form a real-time dataset, and divide the current operating state into the first operating condition, the second operating condition or the third operating condition by threshold judgment. S3. Construct a dual-path LSTM model and execute the corresponding prediction path based on the working condition judgment result: When the first operating condition is identified, the preprocessed sample data is input into LSTM path one, the component prediction temperature at the first preset time in the future is output, and step S5 is executed. When the second working condition is identified, the preprocessed sample data is enhanced to form an extended feature vector. The extended feature vector is then input into LSTM path two, and step S4 is executed. When the third operating condition is identified, skip the prediction step and proceed directly to step S5 and execute the forced full cooling mode. S4. Construct a sample weighted loss function to calculate the gradient of the LSTM model, and construct a two-layer early stop mechanism based on the peak error index in the second working condition. Adaptively balance the LSTM model by improving the learning rate. The LSTM model outputs the component temperature at the second preset time. Execute step S5. S5. Obtain the real-time component temperature of the forklift at the current moment, and combine it with the component predicted temperature output in step S4 or step S5 to determine the target control mode. The target control mode includes predictive intervention mode, normal following mode, forced full cooling mode and energy-saving sleep mode. Generate corresponding control commands according to the target control mode, and send the control commands to the corresponding heat dissipation actuator to adjust the working parameters of the heat dissipation actuator.

[0006] Furthermore, the data acquisition is defined by... Output feature vector at time step for: ; in: These are operation signals, including accelerator pedal opening, hoist handle displacement, and steering angle; , which is a power signal, including traction motor current and voltage, hoisting motor current and power supply, and hydraulic system pressure; These are status signals, including current vehicle speed, cargo weight, and ambient temperature; Differentiated preprocessing paths are adopted based on the different physical properties of the features: Operation signal preprocessing ; ; in, The throttle opening is the result of moving average filtering, and W is the width of the sliding window. The rate of change is used to capture sudden changes in the driver's intentions; Power signal preprocessing in, The current after first-order low-pass filtering. These are the filter coefficients. For the past The amount of electrical energy consumed within seconds; Status signal preprocessing ; The ambient temperature was filtered using median filtering to remove occasional noise interference.

[0007] Furthermore, classifying the current operating status into different operating condition categories includes the following steps: Define working condition identification vector ; , ; in for real-time load rate for Actual power output at any given time This refers to the rated power of the forklift. The load rate threshold, This is the duration threshold; , ; for The slope factor at any given time. for Slope angle at any time Slope threshold; for The current rate of temperature rise at any given moment. for Real-time temperature of the component. for Real-time component temperature at any given moment. The sampling time interval; Based on the above identification indicators, the operating conditions are divided into three types, and corresponding routes are executed: ; First working case, input to the first path of the LSTM model; The second working condition is input into the second path of the LSTM model; In the third operating condition, the current temperature rise rate has exceeded the critical value, so the forced full cooling mode is activated directly, skipping the prediction stage.

[0008] Furthermore, the calculation of enhanced features and the formation of extended feature vectors includes the following steps: The second working condition sample is selected from the original data. For the selected second working condition sample, the following four additional enhancement features are calculated: Cumulative heat generation in, for The cumulative heat generated over time, For the current moment, For summation index variables, Total heating power, The sampling period is Corresponding to a cumulative window of 60-120 seconds, This represents the total heat generation power at the k-th sampling point in the past. The sampling period; Heat accumulation index ; in Forgetting factor, To accumulate the number of sampling points within the window, For 60-120 seconds, Here are the weighting coefficients for the k-th step; Temperature rise acceleration ; ; for Component temperature at any given time The component temperature at the previous sampling time. Sampling period, time The rate of temperature rise, The temperature rise rate at the previous sampling time. For a moment The acceleration of temperature rise; Heat dissipation margin ; for Heat dissipation margin at all times This is the system's maximum heat dissipation power. for The actual heat dissipation power at any given time; After calculating the enhanced features, all features are concatenated to form an extended feature vector: ; in, These are operation signals, power signals, and status signals, respectively. These are cumulative heat generation, heat accumulation index, temperature rise acceleration, and heat dissipation margin, respectively.

[0009] Furthermore, the process of constructing the sample weighted loss function in the second LSTM path is as follows: The weighted loss function formula is as follows: ; in, The number of training samples. Let be the true temperature value of the i-th sample. To predict temperature values ​​for the model, This is the L2 regularization term, used to prevent the model from overfitting. is the regularization coefficient, set to 0.001; Sample weights The calculation formula is as follows: ; in This is the upper limit of normal temperature. This is the critical temperature threshold. Weight amplification factor, when hour, That is, the weight of the normal temperature sample is 1, when hour, , to reach the maximum weight value.

[0010] Furthermore, the two-layer early stopping mechanism in the second LSTM path splits the traffic to different early stopping processing paths based on the judgment result. The specific process is as follows: A special early shutdown judgment condition for the second operating condition is constructed, which is based on the peak error index unique to the second operating condition: ; in: , is the peak error of the model on the validation set after the e-th round of training, which is the maximum value of the prediction error among all samples; , represents the historical peak error; This is the peak error tolerance threshold. For patience value, take 10-15 rounds.

[0011] Furthermore, the early stopping strategy 1 and early stopping strategy 2 are specifically as follows: Early Stop Strategy 1: when Early stopping strategy 1 is triggered, training is terminated directly, and the optimal model is selected from the historical training rounds. The model selection criteria are as follows: ; in, Subtract the number of early stops , To trigger the early stop round, Weighted loss of the validation set in round e. Peak error in round e, The peak error balance coefficient is set to 0.2-0.3. Early Stop Strategy 2: when This triggers early stop strategy 2, allowing for continued training and further adjustments. The basic early stop judgment formula is: ; That is, when the weighted loss of the validation set no longer decreases for several consecutive rounds, the basic early stop is triggered.

[0012] Furthermore, before the basic early stopping condition is met, when the validation set experiences continuous loss... If the learning rate does not improve after each round, then implement learning rate decay: ; in, The learning rate after decay. This is the learning rate decay factor. This is the current learning rate; Training is terminated when the basic early stopping condition is met, and the model with the minimum loss on the validation set is selected. ; When the peak error of the model has not deteriorated in the second operating condition, the model is allowed to continue training to pursue higher overall progress, and the learning rate decay mechanism helps the model break through local optima.

[0013] Compared with the prior art, the technical solution of this application has the following beneficial effects: I. This invention relates to an automatic thermal balance control method for forklifts. Traditional control methods rely on fixed temperature thresholds and are reactive adjustments. When faced with sudden, high-intensity secondary operating conditions (such as fully loaded climbing or frequent lifting), thermal inertia often leads to untimely heat dissipation response and temperature overshoot. This invention introduces a long short-term memory neural network to deeply explore the inherent coupling law between the timing data of driver operation signals, power output, etc., and component temperature rise. The system can accurately predict the temperature trend at a preset time before the actual temperature rise, based on historical operating condition sequences. This advances the control timing from "starting after over-temperature" to "intervening before over-temperature," achieving accurate prediction and active suppression of sudden temperature rises, and significantly improving the thermal balance robustness of the forklift under harsh operating conditions.

[0014] II. This invention presents an automatic thermal balance control method for forklifts. It abandons a simplistic heat dissipation logic and establishes a decision-making model based on a dual-dimensional fusion of "current real-time temperature" and "predicted future temperature." By dynamically judging temperature change trends (rapid rise, stable, dangerous over-limit, rapid drop), the system can intelligently match four control modes: predictive intervention, normal following, forced full cooling, or energy-saving hibernation. This not only ensures that critical components will not exceed the safe temperature limit under any operating condition, solving the safety pain point of thermal balance; but also, when it predicts that the temperature is about to naturally drop (such as at the end of a heavy-load downhill operation), the system can actively reduce heat dissipation power, avoiding energy waste caused by excessive cooling, which is of positive significance for extending the operating time of electric forklifts.

[0015] Third, this invention comprehensively covers the entire process from multi-source heterogeneous data acquisition, LSTM-based temporal feature extraction, fuzzy logic trend judgment, to precise adjustment of the heat dissipation actuator. This automatic control method requires no manual intervention and can adapt to different driving habits, different load weights, and different ambient temperatures, realizing a high degree of intelligence and automation in the forklift thermal management system, and significantly reducing the complexity of developing and calibrating the overall vehicle control strategy. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a flowchart of the working condition judgment process of the present invention; Figure 3 This is a diagram illustrating the judgment process of the dual-layer early stop mechanism of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1: S1. Collect the forklift's operating condition data sequence and corresponding component temperature data sequence during historical operating periods to construct a training sample set; the operating condition data sequence includes operation signals, power signals, and status signals, and preprocesses them using moving average filtering, first-order low-pass filtering, or median filtering according to the physical properties of different signals; data acquisition is performed by defining... Output feature vector at time step for: ; in: These are operation signals, including accelerator pedal opening, hoist handle displacement, and steering angle; , which is a power signal, including traction motor current and voltage, hoisting motor current and power supply, and hydraulic system pressure; These are status signals, including current vehicle speed, cargo weight, and ambient temperature; Differentiated preprocessing is used based on the different physical properties of the features: Operation signal preprocessing ; ; in, The throttle opening is the result of moving average filtering, and W is the width of the sliding window. The rate of change is used to capture sudden changes in the driver's intentions; Power signal preprocessing ; ; in, The current after first-order low-pass filtering. These are the filter coefficients. For the past The amount of electrical energy consumed within seconds; Status signal preprocessing ; The ambient temperature is filtered using a median filter with M ranging from 5 to 11 to remove occasional noise interference.

[0019] Based on the different physical properties of the signals, differentiated preprocessing paths are employed. For the operation signal, a moving average filter is used to remove high-frequency noise, while the rate of change is calculated to capture sudden changes in driver intent; the moving window width is 5 to 10 sampling periods. For the power signal, a first-order low-pass filter is used to smooth spike noise in the current signal; the filter coefficient is 0.2 to 0.5, and the cumulative energy consumption over the past 60 to 120 seconds is calculated to reflect the cumulative effect of heat. For the ambient temperature in the status signal, a median filter is used to remove occasional noise interference; the median filter window is 5 to 11 sampling points.

[0020] S2. Based on real-time collected multi-source data, the current operating state is divided into the first operating condition, the second operating condition, or the thermal runaway early warning condition by threshold judgment; wherein, the second operating condition is when the load rate is greater than the first load threshold, or the ramp rate is greater than the first ramp threshold, or the current temperature rise rate is greater than the first temperature rise rate threshold, and the duration exceeds the first time threshold. The steps involved in classifying the current operating status into different operating condition categories are as follows: Define working condition identification vector ; , ; in for real-time load rate for Actual power output at any given time This refers to the forklift's rated power; the first operating condition is the heavy-duty condition. The preferred value is 0.8. The preferred duration is 10 seconds; , ; for The slope factor at any given time. for Slope angle at time, the first The preferred value is 0.15; for The current rate of temperature rise at any given moment. for Real-time temperature of the component. for Real-time component temperature at any given moment. The sampling time interval, This is the critical temperature rise rate threshold. Based on the above identification indicators, the operating conditions are divided into three types, and corresponding routes are executed: ; General mode, the first working condition input is a general LSTM model; Extreme heavy load mode, second working condition, high load continuous ramp, input to the second working condition LSTM model path two; Thermal runaway early warning mode, third condition: the current temperature rise rate has exceeded the critical value, directly triggering the forced full cooling mode and skipping the prediction stage.

[0021] Based on real-time collected multi-source data, the current operating status is divided into the first operating condition, the second operating condition, or the third operating condition by threshold judgment.

[0022] Operating condition identification is achieved through three core indicators: load rate, slope factor, and current temperature rise rate. The load rate is the ratio of actual power output to the forklift's rated power; the slope factor is the sine of the current slope angle; and the current temperature rise rate is the rate of change of component temperature over time. When the load rate is greater than 0.8, the slope factor is greater than 0.15, or the current temperature rise rate is greater than 3°C / second, and these conditions persist for more than 10 seconds, it is classified as the second operating condition. When the current temperature rise rate exceeds the critical threshold and continues to deteriorate, it is classified as a thermal runaway warning condition. All other conditions are classified as the first operating condition.

[0023] Based on the identification results, the corresponding prediction path is executed: for the first operating condition, the data is input into the general long short-term memory neural network model; for the second operating condition, the data is input into the second operating condition-specific long short-term memory neural network model; for the thermal runaway early warning operating condition, the prediction step is skipped, and the subsequent control steps are directly entered and the forced full cooling mode is executed.

[0024] S3. Execute the corresponding prediction path based on the working condition identification results: When the first operating condition is identified, the preprocessed data is input into the general long short-term memory neural network model, which outputs the component's predicted temperature at the first preset time in the future. When the second operating condition is identified, the preprocessed data is input into the second path of the LSTM model. According to the two-layer early stop mechanism, the predicted temperature of the component at the second preset time in the future is output. When a thermal runaway warning condition is identified, the prediction step is skipped and the process proceeds directly to step S5 and the forced full cooling mode is executed. S4. Construct a sample weighted loss function to calculate the gradient of the LSTM model, and construct a two-layer early stop mechanism based on the peak error index in the second working condition. Adaptively balance the LSTM model by improving the learning rate. The LSTM model outputs the component temperature at the second preset time. Execute step S5. In the training process of the LSTM model in the second scenario, a two-layer early stopping mechanism is adopted. The first layer judges based on the peak error metric. When the peak error deteriorates beyond the tolerance threshold for several consecutive epochs, early stopping strategy 1 is triggered, training is terminated directly, and the optimal model that balances the overall loss and peak error is selected. When the first layer's judgment condition is not met, early stopping strategy 2 is triggered, training continues, and the basic early stopping judgment is adopted. When the weighted loss on the validation set no longer decreases for several consecutive epochs, early stopping is triggered. During the training process, when the loss does not improve for several consecutive epochs, the learning rate is decayed to help the model break through local optima.

[0025] The second working condition sample is selected from the original data. For the selected second working condition sample, the following four additional enhancement features are calculated: Cumulative heat generation ; in, for The cumulative heat generated over time, For the current moment, For summation index variables, Total heating power, The sampling period is Corresponding to a cumulative window of 60-120 seconds, This represents the total heat generation power at the k-th sampling point in the past. The sampling period; Heat accumulation index ; in Forgetting factor, To accumulate the number of sampling points within the window, For 60-120 seconds, Here are the weighting coefficients for the k-th step; Temperature rise acceleration ; ; for Component temperature at any given time The component temperature at the previous sampling time. Sampling period, time The rate of temperature rise, The temperature rise rate at the previous sampling time. For a moment The acceleration of temperature rise; Heat dissipation margin for Heat dissipation margin at all times This is the system's maximum heat dissipation power. for The actual heat dissipation power at any given time; After calculating the enhanced features, all features are concatenated to form an extended feature vector: ; in, These are operation signals, power signals, and status signals, respectively. These are cumulative heat generation, heat accumulation index, temperature rise acceleration, and heat dissipation margin, respectively.

[0026] For the first operating condition, the preprocessed feature sequence is input into the LSTM model path one. This model adopts a three-layer LSTM stacked structure and outputs the predicted temperature of the component 30 seconds later.

[0027] For the second operating condition, samples for that condition are first selected from the original data, and four additional enhanced features are calculated: cumulative heat generation, heat accumulation index, temperature rise acceleration, and heat dissipation margin. The cumulative heat generation is the integral of the total heat output over the past 60 to 120 seconds; the heat accumulation index is the weighted heat generation with a forgetting factor ranging from 0.95 to 0.98; the temperature rise acceleration is the rate of change of the temperature rise rate; and the heat dissipation margin is the ratio of the difference between the maximum and actual heat dissipation power to the maximum heat dissipation power. After calculating the enhanced features, all features are concatenated to form an extended feature vector, which is then input into path two of the second operating condition LSTM model to output the predicted component temperature 10 to 20 seconds later.

[0028] During the model training phase, the second LSTM model path for the second operating condition employs a sample-weighted loss function, assigning higher weights to high-temperature samples. The upper limit of normal temperature is set at 85°C, and the critical temperature threshold is set at 95°C. When the sample temperature exceeds the upper limit of normal temperature, its weight is amplified linearly, with an amplification factor of 3 to 5. Simultaneously, L2 regularization is used to prevent model overfitting, with a regularization coefficient of 0.001.

[0029] The weighted loss function formula is as follows: ; in, The number of training samples. Let be the true temperature value of the i-th sample. To predict temperature values ​​for the model, This is the L2 regularization term, used to prevent the model from overfitting. is the regularization coefficient, set to 0.001; Sample weights The calculation formula is as follows: ; in This is the upper limit of normal temperature. This is the critical temperature threshold. Weight amplification factor, when hour, That is, the weight of the normal temperature sample is 1, when hour, , to reach the maximum weight value.

[0030] The second LSTM model path for the second operating condition includes a two-layer early stopping mechanism, which branches the traffic to different early stopping processing paths based on the judgment result. The specific process is as follows: A special early shutdown judgment condition for the second operating condition is constructed, which is based on the peak error index unique to the second operating condition: ; in: , is the peak error of the model on the validation set after the e-th round of training, which is the maximum value of the prediction error among all samples; , represents the historical peak error; This is the peak error tolerance threshold. For patience value, take 10-15 rounds.

[0031] Furthermore, early stop strategy 1 and early stop strategy 2 are as follows: Early Stop Strategy 1: when Early stopping strategy 1 is triggered, training is terminated directly, and the optimal model is selected from the historical training rounds. The model selection criteria are as follows: ; in, Subtract the number of early stops , To trigger the early stop round, Weighted loss of the validation set in round e. Peak error in round e, The peak error balance coefficient is set to 0.2-0.3. Early Stop Strategy 2: when This triggers early stop strategy 2, allowing for continued training and further adjustments. The basic early stop judgment formula is: ; That is, when the weighted loss of the validation set no longer decreases for several consecutive rounds, the basic early stop is triggered.

[0032] Furthermore, before the basic early stopping condition is met, when the validation set experiences continuous loss... If the learning rate does not improve after each round, then implement learning rate decay: ; in, The learning rate after decay. This is the learning rate decay factor. This is the current learning rate; Training is terminated when the basic early stopping condition is met, and the model with the minimum loss on the validation set is selected. ; When the peak error of the model has not deteriorated in the second operating condition, the model is allowed to continue training to pursue higher overall progress, and the learning rate decay mechanism helps the model break through local optima.

[0033] Furthermore, the predicted component temperatures are physically constrained and validated. The deviation between the predicted temperature rise rate and the physical limit temperature rise rate is calculated; this deviation reflects whether the predicted value violates thermodynamic laws. Based on the comparison between the deviation value and preset first and second confidence thresholds, the confidence level of the predicted value is determined. When the deviation value is less than the first confidence threshold, it is determined to be of high confidence, and the predicted value is directly adopted; when the deviation value is between the first and second confidence thresholds, it is determined to be of medium confidence, and the predicted value is weighted and fused with the current temperature; when the deviation value is greater than the second confidence threshold, it is determined to be of low confidence, downgraded to traditional threshold control, and the predicted value is not involved in subsequent decisions. S5. Obtain the real-time component temperature of the forklift at the current moment, and combine it with the component prediction temperature output in step S3 or the fused temperature processed in step S4. Determine the target control mode through a multi-threshold state machine. The target control mode includes predictive intervention mode, normal following mode, forced full cooling mode and energy-saving sleep mode. Generate corresponding control commands according to the target control mode and send the control commands to the corresponding heat dissipation actuator to adjust the working parameters of the heat dissipation actuator. Example 2

[0034] This embodiment is basically the same as embodiment 1, except that the working condition identification result is the first working condition.

[0035] The forklift is performing routine handling operations with a medium load on a flat road surface. The load factor is 0.45, the slope factor is 0.02, and the current temperature rise rate is 0.5°C / second. It does not meet the conditions for the second working condition, so it is determined to be the first working condition.

[0036] The vehicle controller inputs the preprocessed feature sequence into LSTM model path one. This model uses a three-layer LSTM stacked structure and outputs the predicted component temperature 30 seconds later. In this embodiment, the current motor controller temperature is 75°C, and the predicted temperature after 30 seconds is 78°C. The predicted temperature rise rate deviates little from the theoretical value, and the confidence level is determined to be high.

[0037] The current actual temperature of 75°C and the predicted temperature of 78°C are both within the normal operating range, and the temperature change trend is stable, triggering the normal following mode. The vehicle controller uses a PID control algorithm, with a target temperature of 80°C as the control objective, and linearly adjusts the electric fan speed according to the deviation between the actual temperature and the target temperature. Due to the small temperature rise, the fan operates at a low speed of 20% of its rated speed, maintaining thermal balance and achieving energy saving. Example 3

[0038] This embodiment is basically the same as embodiment 1, except that the confidence level is determined to be low.

[0039] In the second operating condition, if the sensor is subjected to electromagnetic interference, the input current signal will exhibit brief spike noise. Although the signal has been filtered, the noise still affects the input of the LSTM model, causing abnormal jumps in the model's output predicted temperature, with the predicted temperature rise rate far exceeding the theoretical maximum temperature rise rate.

[0040] During the confidence verification phase, the vehicle controller detected that the deviation between the predicted temperature rise rate and the theoretical maximum temperature rise rate exceeded the second confidence threshold (5°C / second), thus classifying it as low confidence. At this point, the controller performs a degradation process, abandoning the use of model predictions and switching to a traditional threshold control mode. This means controlling the cooling actuator based on the comparison between the current actual temperature and a preset threshold. Simultaneously, the controller records this abnormal event for subsequent model monitoring and to trigger retraining. Through the confidence verification and degradation processing mechanisms, malfunctions in cooling operations caused by model misjudgments are avoided, ensuring system reliability.

[0041] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An automatic thermal balance control method for forklifts, characterized in that, Includes the following steps: S1. Collect the operating condition data sequence of the forklift and the corresponding component temperature data sequence during the historical operating period to construct a training sample set; S2. Real-time collection of multi-source data to form a real-time dataset, and the current operating status is divided into the first working condition, the second working condition or the third working condition by threshold judgment. S3. Construct a dual-path LSTM model and execute the corresponding prediction path based on the working condition judgment result: When the first operating condition is identified, the preprocessed sample data is input into LSTM path one, the component prediction temperature at the first preset time in the future is output, and step S5 is executed. When the second working condition is identified, the preprocessed sample data is enhanced to form an extended feature vector. The extended feature vector is then input into LSTM path two, and step S4 is executed. When the third operating condition is identified, skip the prediction step and proceed directly to step S5 and execute the forced full cooling mode. S4. Construct a sample weighted loss function to calculate the gradient of the LSTM model, and construct a two-layer early stop mechanism based on the peak error index in the second working condition. Adaptively balance the LSTM model by improving the learning rate. The LSTM model outputs the component temperature at the second preset time. Execute step S5. S5. Obtain the real-time component temperature of the forklift at the current moment, and combine it with the component predicted temperature output in step S4 or step S5 to determine the target control mode. The target control mode includes predictive intervention mode, normal following mode, forced full cooling mode and energy-saving sleep mode. Generate corresponding control commands according to the target control mode, and send the control commands to the corresponding heat dissipation actuator to adjust the working parameters of the heat dissipation actuator.

2. The automatic thermal balance control method for forklifts according to claim 1, characterized in that: The data acquisition mentioned above is defined Output feature vector at time step for: ； in: These are operation signals, including accelerator pedal opening, hoist handle displacement, and steering angle; , which is a power signal, including traction motor current and voltage, hoisting motor current and power supply, and hydraulic system pressure; These are status signals, including current vehicle speed, cargo weight, and ambient temperature; Differentiated preprocessing paths are adopted based on the different physical properties of the features: Operation signal preprocessing ； ； in, The throttle opening is the result of moving average filtering, and W is the width of the sliding window. The rate of change is used to capture sudden changes in the driver's intentions; Power signal preprocessing ； ； in, The current after first-order low-pass filtering. These are the filter coefficients. For the past The amount of electrical energy consumed within seconds; Status signal preprocessing ； The ambient temperature was filtered using median filtering to remove occasional noise interference.

3. The automatic thermal balance control method for forklifts according to claim 1, characterized in that: The steps involved in classifying the current operating status into different operating condition categories are as follows: Define working condition identification vector ; ， ； in for real-time load rate for Actual power output at any given time This refers to the rated power of the forklift. The load rate threshold, This is the duration threshold; ， ； for The slope factor at any given time. for Slope angle at any time Slope threshold; for The current rate of temperature rise at any given moment. for Real-time temperature of the component. for Real-time component temperature at any given moment. The sampling time interval; Based on the above identification indicators, the operating conditions are divided into three types, and corresponding routes are executed: ； First working case, input to the first path of the LSTM model; The second working condition is input into the second path of the LSTM model; In the third operating condition, the current temperature rise rate has exceeded the critical value, so the forced full cooling mode is activated directly, skipping the prediction stage.

4. The automatic thermal balance control method for forklifts according to claim 3, characterized in that: The calculation of enhanced features and formation of extended feature vectors includes the following steps: The second working condition sample is selected from the original data. For the selected second working condition sample, the following four additional enhancement features are calculated: Cumulative heat generation ； in, for The cumulative heat generated over time, For the current moment, For summation index variables, Total heating power, The sampling period is Corresponding to a cumulative window of 60-120 seconds, This represents the total heat generation power at the k-th sampling point in the past. The sampling period; Heat accumulation index ； in Forgetting factor, To accumulate the number of sampling points within the window, For 60-120 seconds, Here are the weighting coefficients for the k-th step; Temperature rise acceleration ； ； for Component temperature at any given time The component temperature at the previous sampling time. Sampling period, time The rate of temperature rise, The temperature rise rate at the previous sampling time. For a moment The acceleration of temperature rise; Heat dissipation margin ； for Heat dissipation margin at all times This is the system's maximum heat dissipation power. for The actual heat dissipation power at any given time; After calculating the enhanced features, all features are concatenated to form an extended feature vector: ； in, These are operation signals, power signals, and status signals, respectively. These are cumulative heat generation, heat accumulation index, temperature rise acceleration, and heat dissipation margin, respectively.

5. The automatic thermal balance control method for a forklift according to claim 4, characterized in that: The process of constructing the sample-weighted loss function in LSTM path two is as follows: The weighted loss function formula is as follows: ； in, The number of training samples. Let be the true temperature value of the i-th sample. To predict temperature values ​​for the model, This is the L2 regularization term, used to prevent the model from overfitting. is the regularization coefficient, set to 0.001; Sample weights The calculation formula is as follows: ； in This is the upper limit of normal temperature. This is the critical temperature threshold. Weight amplification factor, when hour, That is, the weight of the normal temperature sample is 1, when hour, , to reach the maximum weight value.

6. The automatic thermal balance control method for a forklift according to claim 3, characterized in that: The two-layer early stopping mechanism in the second LSTM path splits traffic to different early stopping processing paths based on the judgment result. The specific process is as follows: A special early shutdown judgment condition for the second operating condition is constructed, which is based on the peak error index unique to the second operating condition: ； in: , is the peak error of the model on the validation set after the e-th round of training, which is the maximum value of the prediction error among all samples; , represents the historical peak error; This is the peak error tolerance threshold. For patience value, take 10-15 rounds.

7. The automatic thermal balance control method for a forklift according to claim 6, characterized in that, The early stopping strategy 1 and early stopping strategy 2 are specifically as follows: Early Stop Strategy 1: when Early stopping strategy 1 is triggered, training is terminated directly, and the optimal model is selected from the historical training rounds. The model selection criteria are as follows: ； in, Subtract the number of early stops , To trigger the early stop round, Weighted loss of the validation set in round e. Peak error in round e, The peak error balance coefficient is set to 0.2-0.

3. Early Stop Strategy 2: when This triggers early stop strategy 2, allowing for continued training and further adjustments. The basic early stop judgment formula is: ； That is, when the weighted loss of the validation set no longer decreases for several consecutive rounds, the basic early stop is triggered.

8. The automatic thermal balance control method for a forklift according to claim 7, characterized in that, Before the basic early stopping condition is met, when the validation set experiences continuous loss If the learning rate does not improve after each round, then implement learning rate decay: ； in, The learning rate after decay. This is the learning rate decay factor. This is the current learning rate; Training is terminated when the basic early stopping condition is met, and the model with the minimum loss on the validation set is selected. ； When the peak error of the model has not deteriorated in the second operating condition, the model is allowed to continue training to pursue higher overall progress, and the learning rate decay mechanism helps the model break through local optima.

Images