An intelligent control method for automatic guiding, conveying, and stacking of container packing equipment.

Abstract

This invention discloses an intelligent control method for automatic guiding, conveying, and stacking of container loading equipment, belonging to the field of automated transportation equipment technology. The method specifically includes the following steps: S1, System initialization and historical big data collection: Inputting container dimensions, cargo box dimensions, and preset parameters; collecting historical big data including AGV operation data, diversion control data, stacking and lifting data, and equipment status data; S2, Historical big data preprocessing: Preprocessing the historical big data by removing outliers, filling in missing values, deduplicating, and standardizing data. By combining dual-laser radar scanning with the "AGV alignment compensation coefficient" output by the intelligent model, the method improves the cargo box collision problem caused by large deviations in traditional mechanical positioning. It compensates for cargo box size fluctuations based on the "lifting height correction value," and combines pressure sensors to detect stacking uniformity, thereby reducing stacking errors, improving layer recognition accuracy, and avoiding the risk of cargo box tipping due to skewed stacking.

Description

Technical Field

[0001] This invention relates to the field of automated transportation equipment technology, and in particular to an intelligent control method for the automatic guiding, conveying and stacking of container packing equipment. Background Technology

[0002] In automated production lines across industries such as logistics, food, and pharmaceuticals, the automated guiding, conveying, and stacking of container packing equipment is a core component ensuring production efficiency and packing quality. While existing container packing equipment and control methods have achieved partial automation, they still suffer from numerous technical shortcomings and struggle to adapt to the complex and ever-changing demands of production scenarios.

[0003] At the data utilization level, existing equipment lacks a comprehensive historical data collection and processing mechanism. Key information generated during the packing process, such as AGV movement parameters (e.g., movement speed, alignment time, deviation value), cargo diversion data (e.g., diversion time, counting error), and lifting and stacking data (e.g., servo motor operating parameters, stacking pressure value), is often not effectively stored, or is only retained in raw data form without in-depth processing and feature extraction. This prevents the equipment from learning and optimizing from historical operation data, leaving the control logic in a state of "passive execution," making it difficult to cope with dynamic adjustment needs under complex operating conditions. Furthermore, when equipment malfunctions, it is impossible to trace the root cause of the problem using historical data, resulting in low maintenance efficiency. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide an intelligent control method for automatic guiding, conveying and stacking of container packing equipment; which can solve the problem that existing equipment does not have a complete historical data collection and processing mechanism.

[0005] Technical Solution: To solve the above-mentioned technical problems, according to one aspect of the present invention, more specifically, an intelligent control method for automatic guiding, conveying, and stacking of container packing equipment, the method specifically includes the following steps:

[0006] S1. System initialization and historical big data collection: Input container size, box size and preset parameters, and collect historical big data including AGV operation data, diversion control data, stacking and lifting data, and equipment status data;

[0007] S2. Historical Big Data Preprocessing: Preprocessing historical big data by removing outliers, filling in missing values, deduplicating, and standardizing data.

[0008] S3. Deep Feature Extraction: The preprocessed historical big data is subjected to basic statistical feature extraction, time-series correlation feature extraction, and sublinear mapping feature extraction in sequence to obtain the deep feature set of the historical big data;

[0009] S4. Intelligent control model construction and training: The intelligent control model is constructed using a BP neural network. The actual control parameters and running results corresponding to the deep feature set in the database are selected to train and test the intelligent control model.

[0010] S5. Intelligent Model-Based Automatic Guided Input and Stacking Control: Real-time data acquisition and feature matching; prediction and parameter adjustment through an intelligent control model; automatic guided transport and stacking based on the output of the intelligent control model.

[0011] S6. Data Archiving and Model Optimization Iteration: Manage and archive all data parameters of this task to the system database, mark them as historical data, and iteratively optimize the model.

[0012] Furthermore, in step S1, the AVG operating data includes: moving speed, alignment time, four-corner coordinate deviation value, and backward distance; diversion control data: diversion time, left and right channel counting error, and guide cylinder action response time; stacking and lifting data: servo motor operating speed, lifting height deviation, stacking pressure distribution value, and layer identification result; and equipment status data: fault codes of each mechanism, operating temperature, and voltage and current parameters.

[0013] Furthermore, in step S2, data exceeding the range of [μ-3σ, μ+3σ] are removed using the 3σ principle; for intermittent missing data with a missing rate ≤5%, linear interpolation is used to complete the missing data; for continuously missing data with a missing rate >5%, the average value of the parameter in similar tasks is replaced; duplicate data caused by repeated sensor triggering is deleted, and valid data collected for the first time is retained; the cleaned data is mapped to the [0, 1] interval using the Min-Max normalization formula.

[0014] Furthermore, step S3 specifically includes the following steps:

[0015] S31. Basic statistical feature extraction: For AGV operation data, calculate the mean, variance, and maximum deviation of the alignment deviation within a single task, the average value and fluctuation range of the moving speed; for diversion control data, calculate the mean absolute value of the counting error within a single task and the median of the guide cylinder action response time; for stacking and lifting data, calculate the standard deviation of the lifting height deviation and the uniformity of the stacking pressure distribution value.

[0016] S32. Temporal correlation feature extraction: Based on timestamps, preprocessed data is segmented according to task cycles. Each completed layer is considered a sub-cycle. The rate of change of the same parameter in adjacent sub-cycles is calculated. The temporal correlation of parameters across modules is analyzed. The sliding window method is used to calculate the sliding mean and sliding variance of temporal features.

[0017] S33. Sublinear mapping feature extraction: The basic statistical features and temporal correlation features are merged into an intermediate feature set. A stacked autoencoder is used for nonlinear mapping to construct a 3-layer neural network. The stacked autoencoder is trained with the stacking error and the layer recognition accuracy as the target labels. The network parameters are optimized through backpropagation so that the model learns the nonlinear correlation features of the data. The features output by the hidden layer of the stacked autoencoder are extracted as deep features to output the deep feature set.

[0018] Furthermore, in step S4, a BP neural network is used to construct an intelligent control model. The number of neurons in the input layer is equal to the depth feature set dimension. The first hidden layer has ten neurons, and the second hidden layer has eight neurons, all using the ReLU activation function. The output layer has three neurons: AGV alignment compensation coefficient, guide diversion response threshold, and lifting height correction value. A certain number of recent complete packing tasks' depth feature sets and corresponding actual control parameters and running results are selected from the database and divided into a training set and a test set in a 7:3 ratio. The stacking accuracy and efficiency are comprehensively optimized. The BP neural network is trained using the gradient descent method, and the weight parameters are iteratively updated until the prediction error of the intelligent control model on the test set meets the standard.

[0019] Furthermore, in step S5, the operating data of the AGV, diversion mechanism, and lifting mechanism are collected in real time. The collected data is preprocessed and deep feature extraction is performed sequentially to obtain the deep feature set of the collected data. The data is then input into the intelligent control model to output the AGV alignment compensation coefficient, guide diversion response threshold, and lifting height correction value for the current task. The PLC corrects the AGV target coordinates, adjusts the guide cylinder action delay time, and corrects the initially calculated lifting height based on the AGV alignment compensation coefficient, guide diversion response threshold, and lifting height correction value.

[0020] Furthermore, in step S5, dual lidars scan the coordinates of the container truck, and the PLC, combined with the corrected target coordinates, controls the AGV's movement, maintaining a stable positioning accuracy of ±0.5mm. Upon reaching its destination, it sends a depalletizing signal to the robot. The stacking robot then delivers the containers to the conveyor mechanism. After detection by photoelectric sensors, the PLC follows an optimized distribution strategy: when the number of containers per layer is even, half the number of containers per layer is first distributed to the left, then half the number of containers per layer is distributed to the right; when the number of containers per layer is odd, half the number of containers per layer minus one is first distributed to the left, then half the number of containers per layer minus one is distributed to the right. The system controls the cylinders. When the current number of cartons being distributed equals the number of cartons per layer, the conveying mechanism and depalletizing robot stop. The dual-screw servo motor pushes the cartons into the lifting mechanism. During the pushing process, the PLC monitors the pushing position in real time through the encoder of the servo motor. If the position deviation exceeds the expected deviation, an abnormal alarm is immediately triggered, the pushing stops, and a fault code is recorded. After the cartons are pushed into place, the PLC controls the servo motor to run in reverse, the pushing plate returns to the initial position, and the current number of cartons is reset to zero. The PLC inputs the collected data depth feature set into the intelligent control model to obtain the lifting height correction value, calculates the corrected lifting height, and controls the three servo motors of the lifting mechanism. The motor drives the support platform to rise to the corrected lifting height. If the uniformity does not exceed the uniformity threshold, the stacking is considered stable, and the PLC controls the ejected plate to return, incrementing the current layer number. If the uniformity exceeds the uniformity threshold, the stacking is considered unstable, the PLC triggers an alarm, stops subsequent actions, and awaits manual inspection. If the current layer number matches the preset total layer number: the PLC controls the lifting mechanism to descend to the initial position, continues destacking and conveying, and begins stacking the next layer. If the current layer number matches the preset total layer number: the PLC sends a reversal permission signal to the AGV control cabinet and simultaneously stops the current actions of all mechanisms. After receiving the reversal permission signal, the AGV will proceed according to the preset... The AGV slowly reverses, with dual LiDAR scanners scanning the distance between the AGV and the packing equipment in real time. When the distance between the AGV and the packing equipment equals the length of the container, the PLC sends a stop reversal signal. The AGV stops and sends a reversal positioning signal. After receiving the reversal positioning signal, the PLC compares the current column number with the preset total column number. If the current column number is less than the preset total column number, the PLC returns the current column number to the AGV for dynamic alignment adjustment, controlling the AGV to move to the next target column position and start the next column stacking. If the current column number equals the preset total column number, the PLC sends an exit permission signal to the AGV, and the AGV exits the container packing area at a preset speed.

[0021] Furthermore, in S6, after the AGV completely exits the container, it sends an exit signal to the PLC. The PLC stops all mechanisms from running, clears the current count, and generates a task report. After each packing task is completed, the depth feature set, actual control parameters, and running results (stacking error and efficiency) of the current task are added to the training set as new samples. For every preset number of new samples accumulated, the intelligent control model is fine-tuned and trained to continuously improve the model's control accuracy.

[0022] Beneficial effects: By combining dual-LiDAR scanning with the "AGV alignment compensation coefficient" output by the intelligent model, the problem of cargo box collisions caused by large positioning deviations in traditional mechanical systems is improved. Based on the "lifting height correction value," cargo box size fluctuations are compensated for. Combined with pressure sensors to detect stacking uniformity, stacking errors are reduced, layer count recognition accuracy is improved, and the risk of cargo box tipping due to misaligned stacking is avoided. The "guided diversion response threshold" optimizes the timing of cylinder actions, solving the counting error problem in traditional fixed diversion modes. Combined with the coordinated control of AGVs, depalletizing robots, and conveying mechanisms, overall packing efficiency is improved. The entire process requires no manual parameter adjustment; intervention is only needed when abnormal alarms occur, reducing reliance on manual labor and avoiding problems such as layer count errors and low efficiency caused by manual intervention.

[0023] During the pushing process, positional deviations are monitored in real time via servo encoders, and stability is detected through pressure distribution during stacking. This allows for early interception of faults such as "cargo box tilting" and "push misalignment," reducing equipment downtime for maintenance. All operational data is archived using "task number + timestamp," enabling rapid root cause identification through historical data when equipment malfunctions, improving maintenance efficiency. After each task is completed, "deep features + actual control parameters + operational results" are added as new samples. The BP model is fine-tuned after accumulating a preset number of samples, allowing the model to continuously adapt to complex operating conditions such as "cargo box material changes," "ground flatness fluctuations," and "equipment aging." There is no need to manually rewrite the control program. By parametrically inputting container and cargo box dimensions, combined with the dynamic adjustment capabilities of the intelligent model, it can quickly switch to different packing scenarios without requiring large-scale modifications to the equipment hardware. By predicting equipment wear and tear using historical data, "preventive maintenance" can be implemented, reducing high repair costs caused by sudden failures. At the same time, iterative model optimization reduces ineffective movements of mechanical parts, extending equipment lifespan. Based on intelligent computing, the "diversion strategy for the number of boxes per layer" combined with precise stacking control improves container space utilization, reduces the number of container transports, and indirectly reduces logistics costs. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the method. Detailed Implementation

[0025] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0026] Example 1

[0027] A scene of loading large containers (12m long × 2.4m wide × 2.6m high) in the logistics industry.

[0028] The first step is system initialization and historical big data collection:

[0029] Parameter input: Operators input the container dimensions (length 12m, width 2.4m, height 2.6m), box dimensions (length 0.6m, width 0.4m, height 0.3m) and preset parameters (number of boxes per layer N=10, total number of layers Ktotal=8, total number of columns Mtotal=20) through the HMI, and the system loads the parameters into the cache.

[0030] Historical data acquisition: Activate the data acquisition unit (acquisition frequency 10Hz) to collect historical big data, specifically including:

[0031] AGV operating data: moving speed (0.2-0.5m / s), alignment time (15-30s), four-corner coordinate deviation (±0.3-±1.2mm), backward distance (0.6-1.2m);

[0032] Diversion control data: Diversion time (0.8-1.5s / unit), left and right channel counting error (0-2 units), guide cylinder action response time (0.1-0.3s);

[0033] Stacking and lifting data: servo motor operating speed (800-1200rpm), lifting height deviation (±0.2-±0.8mm), stacking pressure distribution value (50-120N), layer recognition result (1-8 layers, recognition accuracy 98.5%-99.8%);

[0034] Equipment status data: Fault codes for each mechanism (such as servo motor overload code E01, radar fault code F03), operating temperature (35-45℃), voltage and current (AC380V±5%, 5-15A).

[0035] The collected data is stored in a MySQL database in real time, and an index is created by "task number + timestamp". A total of 1,000 historical data of similar container loading are collected as the basic dataset.

[0036] The second step is historical big data preprocessing:

[0037] Outlier removal: Calculate the mean μ and standard deviation σ of each parameter. For example, if the mean AGV alignment deviation is μ = 0.6 mm and σ = 0.2 mm, remove outlier data that exceeds the range of [0, 1.2 mm] using the 3σ principle (such as an alignment deviation of 1.5 mm in a certain instance). At the same time, remove cross-parameter logical outlier data (such as stacking pressure value = 80 N when the lifting height is 0).

[0038] Missing value completion: For intermittent missing data with a missing rate ≤ 5% (such as a missing stack pressure value), linear interpolation is used for completion. For example, in a time series, the pressure value is 60N at t1 = 10s, 70N at t3 = 12s, and the missing value at t2 = 11s is completed as 65N. For continuous missing data with a missing rate > 5% (such as missing radar coordinate data within a 5s period), it is replaced with the average value of the parameter in the last 10 similar tasks (such as the average AGV alignment time of 22s).

[0039] Deduplication: Delete duplicate trigger data from sensors. For example, if a photoelectric sensor outputs a "cargo box present" signal twice in a row due to vibration, only retain the valid data from the first time (t=15.2s).

[0040] Data standardization: The Min-Max standardization formula X is used. std =(XX) min ) / (X max -X min The cleaned data is mapped to the [0, 1] interval. Taking the AGV's moving speed as an example, X... min =0.2m / s, X max =0.5m / s, when a certain data point X = 0.35m / s, the standardized X std = (0.35-0.2) / (0.5-0.2) = 0.5, ensuring that the data volume is consistent across different dimensions.

[0041] The third step is deep feature extraction:

[0042] 1. Basic statistical feature extraction:

[0043] For AGV operation data: Calculate the mean (e.g., 0.6mm) and variance (0.04mm) of the alignment deviation within a single task. 2 The maximum deviation (1.2mm), average moving speed (0.35m / s), and fluctuation amplitude (0.3m / s) are calculated. For the diversion control data: the mean absolute value of the counting error within a single task (0.8 counts) and the median of the guide cylinder action response time (0.2s) are calculated. For the stacking and lifting data: the standard deviation of the lifting height deviation (0.2mm) and the uniformity of the stacking pressure distribution value are calculated, such as (|55-60|+|62-60|+…) / 8=2.5N.

[0044] The output basic statistical feature set has a total of 12 feature dimensions.

[0045] 2. Temporal correlation feature extraction:

[0046] The process is divided into sub-cycles based on "completing one layer of stacking" (each sub-cycle in this scenario is approximately 60 seconds). After data segmentation, the rate of change of parameters between adjacent sub-cycles is calculated. For example, the AGV alignment deviation is 0.6 mm in the first sub-cycle and 0.5 mm in the second sub-cycle, with a rate of change of (0.5-0.6) / 0.6 × 100% ≈ -16.7%. Cross-module temporal correlations are analyzed, such as the difference between "the guide cylinder's action completion time (0.2 s) and the subsequent cargo box pushing time (1.2 s) = 1.0 s" and "the Pearson correlation coefficient r between radar scanning frequency (10 Hz) and AGV alignment accuracy (0.6 mm) = -0.7 (negative correlation; higher frequency means higher accuracy)". A sliding window method (window size = 5 sub-cycles) is used to calculate the sliding mean and sliding variance, such as the AGV moving speed sliding mean (0.35 m / s) and sliding variance (0.02 m / s). 2 The output time-series correlation feature set provides 18 dimensions, which, together with the basic statistical feature set, form an intermediate feature set.

[0047] 3. Nonlinear Mapping Feature Extraction: A 3-layer stacked autoencoder (SAE) is constructed: an input layer with 30 neurons (corresponding to F1+F2 dimensions), a hidden layer with 15 neurons (ReLU activation function), and an output layer with 30 neurons (Sigmoid activation function). Using "stacking error ≤ 1mm" and "layer recognition accuracy ≥ 99.9%" as target labels, the SAE is trained with 500 high-quality historical task data sets. Backpropagation is used to optimize network parameters (learning rate 0.001, 1000 iterations) to enable the model to learn nonlinear correlation features (such as the implicit correlation between "radar scanning frequency - AGV alignment deviation - stacking error"). The 15 features from the hidden layer output are extracted as a deep feature set, such as "the coupling feature between the AGV alignment deviation change rate and radar frequency" and "the correlation feature between stacking pressure uniformity and lifting height deviation".

[0048] Step 4: Intelligent control model construction and training:

[0049] Model structure design: The model is constructed using a BP neural network as described in claim 5, with 15 neurons in the input layer (corresponding to the depth feature set dimension), 10 neurons in the first hidden layer and 8 neurons in the second hidden layer (both with ReLU activation function), and 3 neurons in the output layer (AGV alignment compensation coefficient α: 0.98-1.02, guidance shunting response threshold β: 0.3-0.7s, lifting height correction value Δh: -0.5-0.5mm).

[0050] Dataset partitioning and training: Select 1000 complete task data from the historical database and divide them into a training set (700 samples) and a test set (300 samples) in a 7:3 ratio. Each sample contains "deep feature set + actual control parameters + running results (placing error, efficiency)".

[0051] With the goal of "maximizing both placement accuracy and efficiency", a multi-objective loss function (Loss = 0.6 × MSE (output, true parameters) + 0.4 × (1 - efficiency improvement rate / 40%)) is adopted, and the model is trained using gradient descent (AdamW optimizer, learning rate 0.0005) and the weight parameters are iteratively updated.

[0052] Training stops when the test set prediction error (RMSE≤0.05mm) meets the standard, and the model parameters (.pth format) are saved to the PLC local storage. The trained model can output control parameters adapted to this scenario (such as α=1.01, β=0.45s, Δh=0.2mm).

[0053] Step 5: Automatic guiding, conveying, and stacking control based on intelligent models:

[0054] Real-time data acquisition and feature matching: Real-time acquisition of operating data of AGV (moving speed 0.38m / s, alignment deviation 0.55mm), diversion mechanism (diversion time 1.1s / unit, counting error 1 unit), and lifting mechanism (servo speed 1000rpm, lifting deviation 0.3mm), preprocessing and feature extraction to generate real-time depth feature set.

[0055] Model prediction and parameter tuning: Input the deep feature set _current into the trained BP model, and output the current task control parameters: α = 1.005, β = 0.42s, Δh = 0.15mm; PLC correction parameters:

[0056] Corrected AGV target coordinates: Initial X0 = 12m / 2 - 0.6m / 2 = 5.7m, corrected X0' = 5.7m × 1.005 = 5.7285m. Adjusted guide cylinder action delay time: changed from default 0.5s to 0.42s. Corrected lifting height: Initial H = K × 0.3m (H = 0 when K = 0), corrected H' = 0 + 0.15mm = 0.15mm (due to slight fluctuations in cargo box height).

[0057] Specific control execution: AGV positioning: Dual laser radar scans the coordinates of the four corners of the container, and the PLC controls the movement of the AGV in combination with X0' = 5.7285m. The positioning accuracy is stable at ±0.5mm. After reaching the position, a depalletizing signal is sent to the depalletizing robot.

[0058] Depalletizing and Diverting: The robot grabs the cartons and sends them to the conveyor (speed 0.5m / s). After the photoelectric sensor detects the cartons, the PLC controls the guide cylinder according to the optimized strategy (N=10 is an even number, 5 on the left and 5 on the right). The action is delayed by 0.42s. First, 5 cartons are diverted to the left (counted by the left sensor), and then 5 are diverted to the right. When the number of diverted cartons C=10, the conveying and the robot stop.

[0059] Pushing and Lifting Stacking: A dual-screw servo motor (1000rpm) pushes 10 boxes into the lifting mechanism. The PLC monitors the pushing position via an encoder (accuracy ±0.1mm). If the deviation exceeds ±0.3mm, an alarm is immediately triggered. After the boxes are pushed into place, the servo reverses its direction, and C is zeroed. The PLC controls the lifting mechanism to rise according to H' = 0.15mm, pushing the boxes into the container. The pressure sensor detects that the pressure uniformity is 2.1N (≤ threshold 3N), indicating stability. The ejector plate returns, and the number of layers K = 1.

[0060] Layer and column number determination: K = 1 < 8, the lifting mechanism descends to the initial position, and the above process is repeated until K = 8; the PLC sends a signal to the AGV to allow retreat, the AGV retreats at 50mm / s, and stops when the radar detection distance = 0.6m (cargo box length), and sends a retreat completion signal; the current column number M = 1 < 20, M is updated to 2, the AGV moves to the next column target position (X0” = 5.7285m + 2 × 0.6m = 6.9285m), and continues stacking until M = 20; the AGV exits the container at 100mm / s and sends an exit completion signal.

[0061] Step 6: Data archiving and model optimization iteration:

[0062] Data archiving: After the AGV exits, the PLC stops all mechanisms from running, clears the counters (C=0, K=0, M=0), and generates a task report (task number: WL20240501, time: 120min, number of boxes packed: 20×8×10=1600, stacking error 0.3mm, efficiency improvement 42%, no faults); the "F3_current + actual control parameters + task report" of this task are archived to the historical database.

[0063] Model iteration: According to claim 8, every 50 new task samples (including the data in this embodiment) are accumulated, the new samples are added to the training set, and step S4 is repeated to fine-tune the model and update the model parameters in the PLC so that the model can adapt to the working conditions such as batch differences of cargo boxes (such as changes in cardboard box hardness) and fluctuations in ground flatness in this scenario, and continuously improve the control accuracy (such as reducing the stacking error to 0.25mm).

[0064] Example 2

[0065] Packing scenario for small containers (2m long × 1.2m wide × 1.5m high) in the pharmaceutical industry.

[0066] The first step is system initialization and historical big data collection:

[0067] Parameter input: Enter the container dimensions (length 2m, width 1.2m, height 1.5m), carton dimensions (length 0.2m, width 0.15m, height 0.1m) and preset parameters (N=7, Ktotal=12, Mtotal=10) through the HMI.

[0068] Historical data collection:

[0069] AGV operating data: moving speed (0.1-0.3m / s), alignment time (8-15s), four-corner coordinate deviation (±0.2-±0.8mm), and backward distance (0.2-0.4m).

[0070] Diversion control data: Diversion time (0.3-0.8s / each), left and right channel counting error (0-1), cylinder response time (0.08-0.2s).

[0071] Stacking and lifting data: servo speed (600-900rpm), lifting deviation (±0.1-±0.5mm), stacking pressure (20-50N), layer recognition result (1-12 layers, accuracy 99.0%-99.7%).

[0072] Equipment status data: fault codes (e.g., photoelectric sensor fault E05), temperature (25-30℃, constant temperature in pharmaceutical workshops), voltage and current (AC380V, 3-8A).

[0073] A total of 800 historical data entries for similar tasks have been collected and stored in the database.

[0074] The second step is historical big data preprocessing:

[0075] Outlier removal: If the average AGV alignment deviation is μ = 0.4 mm and σ = 0.15 mm, remove outlier data that exceeds [0.05, 0.75 mm] (e.g., a deviation of 0.9 mm in a certain instance).

[0076] Missing value completion: A certain servo speed data is missing (missing rate 3%), and is completed using linear interpolation; a certain 3s radar data is missing (missing rate 6%), and is replaced with the average value of the last 10 times (0.25m / s).

[0077] Deduplication: Delete repetitive cylinder action signals (such as two consecutive "extend" signals, retaining only the first one).

[0078] Standardization: such as stacking pressure X min =20N, X max=50N, a certain data X =35N, after standardization X_std = (35-20) / (50-20) = 0.5.

[0079] The third step is deep feature extraction:

[0080] 1. Basic statistical characteristics: such as the mean AGV alignment deviation of 0.4mm and variance of 0.02mm. 2 The mean absolute value of the shunt counting error is 0.3, the standard deviation of the lifting height deviation is 0.12mm, and the basic statistical feature set is output.

[0081] 2. Temporal correlation features: Taking "one layer stacking (about 30s)" as the sub-cycle, calculate the change rate of the rise and fall deviation between adjacent sub-cycles (e.g., -10%), analyze "the difference between "cylinder response time (0.15s) and push time (0.6s) = 0.45s", calculate the moving average using a sliding window (5 sub-cycles), output the temporal correlation feature set, and combine it with the basic statistical feature set to form an intermediate feature set.

[0082] 3. Nonlinear mapping: The SAE was trained with 400 high-quality data sets to extract 15 deep features from the hidden layer (such as "the correlation features between pressure uniformity and rise / fall deviation under constant temperature in a pharmaceutical workshop").

[0083] Step 4: Intelligent control model construction and training:

[0084] Model structure: BP neural network with 15-dimensional input layer, 10+8 neurons in hidden layer, and output layer α (0.98-1.02), β (0.3-0.7s), Δh (-0.5-0.5mm).

[0085] Training: Divide 800 data points into a training set (560 data points) and a test set (240 data points) in a 7:3 ratio. Use gradient descent to train the model. Stop training when the RMSE of the test set is less than or equal to 0.04 mm. The model outputs the adaptation parameters (e.g., α = 0.995, β = 0.38s, Δh = -0.1 mm).

[0086] Fifth step, control execution based on intelligent model:

[0087] Real-time data processing: Collect data from AGV (speed 0.22m / s, deviation 0.38mm), diversion mechanism (diversion time 0.5s / unit), and lifting mechanism (speed 800rpm, deviation 0.2mm), generate a depth feature set, input the model and output α=0.998, β=0.35s, Δh=-0.08mm.

[0088] Parameter correction: AGV target coordinate X0=2m / 2-0.2m / 2=0.9m, after correction X0'=0.9×0.998=0.8982m; cylinder delay 0.35s; lifting height H'=K×0.1m-0.08mm.

[0089] Control execution:

[0090] AGV alignment: After radar scanning, move by pressing X0' with an accuracy of ±0.5mm, and send a destacking signal.

[0091] Diversion: N=7 (odd number), PLC controls the cylinder to act with a delay of 0.35s, first divide 3 to the left (7 / 2-1=2.5→rounded to 3), then divide 3 to the right, and the remaining 1 is centered. When C=7, the conveying stops.

[0092] Push and stack: Servo pushes the box, monitors position deviation (alarm if ±0.3mm), lifts the mechanism by pressing H', pressure uniformity = 1.2N (≤ threshold 2N), K = 1.

[0093] Layer and column number determination: When K=12, the AGV moves backward and stops when the distance is 0.2m (cargo box length); when M=10, the AGV exits the container.

[0094] Step 6: Data archiving and model optimization:

[0095] Archiving: Generate a task report (ID YY20240502, time 45min, number of boxes packed 10×12×7=840, stacking error 0.2mm, efficiency improvement 45%), and archive the data.

[0096] Iteration: Every 50 new samples (including this embodiment) are accumulated, the model parameters are fine-tuned to adapt the model to the soft and easily deformable characteristics of the pharmaceutical aluminum-plastic box material, and the layer recognition accuracy is improved to 99.95%.

[0097] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. An intelligent control method for automatic guiding, conveying and stacking of a container loading apparatus, characterized in that, The method specifically comprises the following steps: S1, system initialization and historical big data collection: enter the size of the container, the size of the box and the preset parameters, collect historical big data including AGV running data, shunt control data, stacking lifting data and equipment state data; S2, historical big data preprocessing: preprocessing of historical big data including outlier rejection, missing value completion, deduplication and data standardization; S3, deep feature extraction: sequentially extracting basic statistical features, time series correlation features and linear mapping features from the preprocessed historical big data to obtain a deep feature set of the historical big data; S4, intelligent control model construction and training: constructing an intelligent control model using a BP neural network, selecting the deep feature set from the database corresponding to the actual control parameters and running results to train and test the intelligent control model; S5, automatic guiding input stacking control based on the intelligent model: real-time data collection and feature matching, parameter adjustment through the intelligent control model, and automatic guiding and stacking based on the output of the intelligent control model S6, data archiving and model optimization iteration: archiving the data parameters of the task to the system database, marking them as historical data, and iterating and optimizing the model; The step S3 specifically comprises the following steps: S31, basic statistical feature extraction: calculating the mean, variance and maximum deviation of the alignment deviation, the average and fluctuation amplitude of the moving speed in a single task, the absolute value mean of the counting error, the median of the guiding cylinder action response time in the shunt control data, the standard deviation of the lifting height deviation and the uniformity of the stacking pressure distribution value in the stacking lifting data; S32, time series correlation feature extraction: segmenting the preprocessed data by task period based on the timestamp, calculating the change rate of the same parameter in adjacent sub-periods, analyzing the time series correlation of cross-module parameters, and calculating the sliding mean and sliding variance of the time series features using the sliding window method; S33, linear mapping feature extraction: combining the basic statistical features and time series correlation features into an intermediate feature set, performing nonlinear mapping using a stacked autoencoder, constructing a 3-layer neural network, training the stacked autoencoder with the stacking error and layer recognition accuracy as the target label, optimizing the network parameters through back propagation, making the model learn the nonlinear correlation features of the data, and extracting the features output by the hidden layer of the stacked autoencoder as the deep feature set.

2. The intelligent control method for the automatic guiding, conveying and stacking of a container loading device according to claim 1, characterized in that: In the step S1, the AVG running data includes: moving speed, alignment time, four-corner coordinate deviation value, and backward distance; the shunt control data includes: shunt time, left and right channel counting error, and guiding cylinder action response time; the stacking lifting data includes: servo motor running speed, lifting height deviation, stacking pressure distribution value, and layer recognition result; and the equipment state data includes: each mechanism fault code, running temperature, and voltage and current parameters.

3. The intelligent control method for the automatic guided conveying and stacking of the container packing equipment according to claim 1, characterized in that: In the step S2, data beyond the range of [μ-3σ, μ+3σ] is removed by the 3σ principle; for intermittent missing data with a missing rate of ≤5%, linear interpolation is used to complete the data, and for continuous missing data with a missing rate of >5%, the average value of the parameter in the same task is used to replace the data; repeated data generated by repeated triggering of the sensor is deleted, and the effective data collected for the first time is retained; the cleaned data is mapped to the interval [0, 1] by using the Min-Max standardization formula.

4. The intelligent control method for the automatic guided conveying and stacking of the container packing equipment according to claim 1, characterized in that: In the step S4, a BP neural network is used to construct an intelligent control model, the number of neurons in the input layer is the dimension of the deep feature set, there are ten neurons in the first hidden layer and eight neurons in the second hidden layer, both of which use the ReLU activation function, and there are three neurons in the output layer: AGV alignment compensation coefficient, guiding shunt response threshold, and lifting height correction value; the deep feature set of a certain number of complete container loading tasks and the corresponding actual control parameters and running results are selected from the database, and are divided into a training set and a test set according to a ratio of 7:3, the stacking accuracy and efficiency are comprehensively optimized, the BP neural network is trained by using the gradient descent method, the weight parameters are iteratively updated, and the prediction error of the intelligent control model on the test set is up to standard.

5. The intelligent control method for the automatic guided conveying and stacking of the container packing equipment according to claim 1, characterized in that: In the step S5, the running data of the AGV, the shunt mechanism and the lifting mechanism are collected in real time, the collected data is pretreated and deep features are extracted to obtain a deep feature set of the collected data, and the deep feature set is input into the intelligent control model to output the AGV alignment compensation coefficient, the guiding shunt response threshold and the lifting height correction value of the current task, the PLC corrects the AGV target coordinates, adjusts the action delay time of the guiding cylinder and corrects the initially calculated lifting height according to the AGV alignment compensation coefficient, the guiding shunt response threshold and the lifting height correction value.

6. The intelligent control method for the automatic guided conveying and stacking of a container loading apparatus according to claim 1, characterized in that: In the step S5, the double laser radar scans the container truck box coordinates, and the PLC controls the AGV to move in combination with the corrected target coordinates, the positioning accuracy is stable at ±0.5 mm, and a destacking signal is sent to the robot after reaching the position, the stacking robot sends the container to the conveying mechanism, and after detection by the photoelectric sensor, the PLC controls the guiding cylinder according to the optimized shunting strategy: when the number of boxes per layer is even, the left half of the number of boxes per layer is first divided, and then the right half of the number of boxes per layer is divided, when the number of boxes per layer is odd, the left half of the number of boxes per layer minus one is first divided, and then the right half of the number of boxes per layer minus one is divided, the conveying mechanism and the destacking robot are stopped when the current shunting quantity is equal to the number of boxes per layer, the double-screw rod servo motor pushes the container into the lifting mechanism, in the pushing process, the PLC monitors the pushing position in real time through the encoder of the servo motor, if the position deviation exceeds the deviation expectation, an abnormal alarm is triggered immediately, the pushing is stopped and the fault code is recorded, after the pushing is in place, the PLC controls the servo motor to operate in reverse, the pushing plate returns to the initial position, and the current shunting quantity is cleared at the same time, the PLC inputs the collected data depth feature set into the intelligent control model to obtain a lifting height correction value, calculates a corrected lifting height, and controls the three servo motors of the lifting mechanism to drive the bearing platform to rise to the corrected lifting height, if the uniformity does not exceed the uniformity threshold, it is determined that the stacking is stable, the PLC controls the pushing plate to return, and the current layer number is increased by one, if the uniformity exceeds the uniformity threshold, it is determined that the stacking is unstable, the PLC triggers an alarm and stops the subsequent actions, and waits for manual inspection, if the current layer number is equal to the preset total layer number: the PLC controls the lifting mechanism to descend to the initial position, continues to destack and convey, and starts to stack the next layer, if the current layer number is equal to the preset total layer number: the PLC sends a permission to retreat signal to the AGV control cabinet, and stops all mechanisms at the current action, after receiving the permission to retreat signal, the AGV retreats slowly at a preset speed, the double laser radar scans the distance between the AGV and the loading device in real time, when the distance between the AGV and the loading device is equal to the length of the container, the PLC sends a stop retreat signal, the AGV stops and sends a retreat in place signal, after receiving the retreat in place signal, the PLC compares the current column number with the preset total column number: if the current column number is less than the preset total column number: the PLC returns the current column number to the AGV alignment dynamic adjustment, controls the AGV to move to the next column target position, and starts to stack the next column, if the current column number is equal to the preset total column number: the PLC sends a permission to exit signal to the AGV, and the AGV exits the container loading area at a preset speed.

7. The intelligent control method for the automatic guided conveying and stacking of a container loading apparatus according to claim 1, characterized in that: In the step S6, after the AGV completely exits the container, a retreat in place signal is sent to the PLC, the PLC stops the operation of all mechanisms, clears the current count, generates a task report, and after each container loading task is completed, the depth feature set, the actual control parameter and the running result: stacking error and efficiency of the current task are supplemented to the training set as new samples, and the intelligent control model is fine-tuned once every time a preset number of new samples are accumulated, so as to continuously improve the control accuracy of the model.

Images