Quay crane automatic control system and method based on deep learning

Abstract

The invention provides a quay crane automatic control system based on deep learning and a method thereof, and aims to realize fine automatic control of quay crane operation through a deep learning technology. The system comprises a data collection module, a data preprocessing module, a model training module and a deployment reasoning module. According to the method, the performance and robustness of automatic control of the quay crane are remarkably improved, the manpower and equipment cost is reduced, and the operation continuity and stability are improved. By introducing the deep learning technology, the driving behavior of a quay crane operator can be learned and simulated, and an accurate control instruction is generated. Meanwhile, the invention provides an efficient model training process and optimization method, and ensures that the trained deep learning model has excellent performance and generalization ability.

Description

Technical Field

[0001] The present invention relates to the field of automation control technology, and in particular to a deep learning-based quay crane automation control system and method, which utilizes deep learning technology to achieve automated control of the quay crane's container grabbing and releasing functions. Background Art

[0002] In the port logistics system, quay cranes are core hub equipment, and their automation and intelligent transformation have immeasurable value in promoting a qualitative leap in port operation efficiency and a significant improvement in cost-effectiveness. In particular, the automatic grab and release function of quay cranes, as a key technical link to achieve efficient and precise operations, is directly related to the efficiency and cost structure of the overall port operation.

[0003] At present, despite the increasingly obvious trend of port automation, most quay cranes still rely heavily on manual operation for container grabbing and placing operations, which not only limits the operation speed, but also increases the dependence on the operator's skill level, which is not conducive to the optimal allocation of resources. In addition, the long-term, high-intensity continuous operation environment, such as the continuous influence of natural factors such as sea breeze and tide, poses a severe challenge to the operator's physical and mental state, which may affect the safety and continuity of the operation and increase potential safety risks. For example, patent CN114955856A proposes an innovative anti-sway control method for quay crane trolleys based on trajectory planning, which focuses on optimizing the motion stability of the trolley, but has not yet achieved full automation of container grabbing and placing operations. Moreover, the scheme is based on trajectory planning, and the scheme effect depends on the modeling accuracy of the quay crane dynamic system and the response of the quay crane motor to the desired speed. In addition, patent CN107720556B proposes an automated quay crane control method and system. Although the system automates the movement of the trolley and the lifting movement, the intervention of the quay crane operator is still required in the container grabbing and placing process, and it has not completely gotten rid of the dependence on manual operation. Summary of the invention

[0004] To solve the above problems, the present invention proposes a deep learning-based quay crane automation control system and method, which can clone and imitate the demonstration of a human operator driving a quay crane to realize the automation of the quay crane's container grabbing and releasing functions.

[0005] The present invention is conceived as follows:

[0006] Using deep learning technology, the neural network model is trained to clone the demonstration of human operators driving the quay crane, realizing the automation of the container grabbing and releasing function of the quay crane. Specifically, building the control system of the quay crane involves two key stages: model training and deployment reasoning. The model training stage requires the collection of demonstration data generated by the quay crane operator in actual operation. This includes various observation signals, such as images captured by multi-angle cameras, the speed and position of the trolley and lifting mechanism recorded by the PLC, and control signals, namely the lifting speed signal and trolley speed signal expected in the quay crane operation. These collected data will undergo a series of preprocessing steps, including data cleaning and normalization, to ensure the quality and consistency of the data. Subsequently, according to the training algorithm, the demonstration data can be used to systematically train the deep neural network so that it learns how to extract valuable information from the observation signal and convert it into precise control instructions. The deployment reasoning stage is to apply the trained model to real-time operations. In each control cycle, the system will capture the observation signal in real time and pass it as input to the model for reasoning analysis. The model will synthesize this information and infer a series of control signals, which will be sent to the PLC through the communication system to perform precise motion control on the quay crane. This process forms a closed-loop control system, ensuring the continuity, stability and responsiveness of quay crane operations.

[0007] The deep neural network is designed with reference to the ACT (Action Chunking with Transformers) algorithm. Compared with the traditional imitation learning algorithm, this algorithm has more embodied design and has made great innovations in two aspects. First, the algorithm adopts action chunking. The design of action chunking is inspired by neuroscience, which means combining independent actions into a unit to execute, thereby improving the smoothness of action execution and thus improving the success rate of the task. In the ACT algorithm, every k steps, the model obtains input once, predicts the next k steps of action, and then executes these actions in sequence. Compared with traditional control algorithms and reinforcement learning algorithms, this design effectively prevents the accumulation of single-step errors and improves the success rate of fine-grained task execution. Second, when training the model, the algorithm uses the conditional automatic variation encoder as the overall framework structure, and only uses the decoder part to generate a k-step action sequence during reasoning. There is also the design idea of ​​embodied intelligence behind this. Using the generative model as the policy network can better handle and reproduce the noise and randomness in human demonstration data. The following will specifically introduce the implementation details of the present invention in terms of model structure, loss function and model training process.

[0008] In terms of model structure, the conditional variational autoencoder is used as the overall network framework. The encoder part is implemented using a Transformer encoder. Unlike the standard conditional variational autoencoder, the encoder part of the present invention does not introduce all conditional variables. The encoder only uses the vectorized partial observation data plus the k-step action sequence a in the demonstration data. t：t+k As input, the output representation is compressed into a potential representation vector z through average pooling. The decoder part is implemented using a complete Transformer model. Among them, the input of the encoder part of the Transformer is the current vectorized shore bridge state and the potential expression vector z finally output by the encoder part; the input of the decoder part of the Transformer is a fixed position embedding (positional embedding) of length k, and its output is the total generated k-step action sequence after being projected to the appropriate dimension through a linear layer. In addition, the images involved in the decoder part can be converted into feature vectors through a convolutional neural network and input into the decoder. For example, if the input image size is 80×80×3, the convolutional neural network ResNet18 can be used to extract it into a feature vector with a shape of 1×128, and then input into the Transformer encoder for subsequent calculations.

[0009] In terms of loss function, the loss function of this deep learning algorithm consists of two parts: reconstruction loss and KL divergence loss. For the reconstruction error part, let the observation of the system at time t be o t , the corresponding action sequence demonstrated by the expert is a t：t+k , the demonstration data set is D, the model weight parameter is θ, and the decision model is π θ In imitation learning, the method of maximizing the log-likelihood of the demonstration strategy is usually used for training, that is, minimizing the loss function

[0010]

[0011] In the crane anti-sway problem designed by the present invention, the action input of the system is set as the expected speed of the quay crane trolley and the lifting, and its action space is a continuous action space. Therefore, the mean square error loss is directly used to characterize the reconstruction loss. The main function of the KL divergence part loss function is to make the posterior probability distribution learned by the encoder Reduce to a hypothesized prior probability distribution Since both are assumed to be normally distributed, is a standard normal distribution with a mean of 0 and a variance of 1. If the model parameter is assumed to be φ, the KL divergence partial loss function is

[0012]

[0013] The above two loss functions are weighed by weight β and added together to obtain the final loss function.

[0014] The technical solution of the present invention is as follows:

[0015] On the one hand, the present invention provides a deep learning-based quay crane automation control system, which is characterized by comprising:

[0016] A data collection module is configured to install cameras at key locations of the quay crane and record demonstration data generated during manual operations of the quay crane, including multi-angle images, speed and position information of the quay crane trolley and hoisting mechanism recorded by the PLC, and control signals generated by the operator when driving the quay crane;

[0017] A data preprocessing module is configured to align the observation data with the control instruction data according to the timestamp to form a data pair, and clean and normalize the data;

[0018] a model training module, configured to train the deep neural network model using an imitation learning algorithm, using the processed data pairs as training data, the observation data as model input, the control sequence as a target value, and updating the model weights through a back-propagation algorithm;

[0019] The deployment inference module is configured to deploy the trained model in the server at the quay crane, collect images from various perspectives and PLC perception data in each control cycle, input the observation signal into the deep learning model for inference, obtain the control sequence signal, and send the control sequence signal to the PLC, which performs the underlying drive of the quay crane motor.

[0020] Furthermore, the data collection module includes multiple cameras installed at key positions of the quay crane to capture multi-angle images.

[0021] Furthermore, the data preprocessing module includes a data cleaning submodule for removing abnormal noise points, and a data normalization submodule for ensuring data consistency.

[0022] Furthermore, the imitation learning algorithm used by the model training module is capable of learning complex mappings from observation signals to control instructions.

[0023] Furthermore, the deployment reasoning module can maintain high performance under changing operating conditions and ensure the continuity, stability and responsiveness of the quay crane operation through a closed-loop control system.

[0024] On the other hand, the present invention also provides a deep learning-based quay crane automation control method, which is characterized in that it includes the following steps:

[0025] S1. Install cameras at key locations of the quay crane and collect demonstration data generated during manual operations of the quay crane, including multi-angle images, observation signals of the speed and position information of the quay crane trolley and lifting mechanism recorded by the PLC, and control signals generated by the operator when driving the quay crane;

[0026] S2. Align the observation data and the control instruction data according to the timestamp to form a data pair, and clean and normalize the data pair;

[0027] S3. train the deep neural network model using an imitation learning algorithm, using the data pair as training data, the observed data as model input, the control sequence as the target value, and updating the model weights through a back propagation algorithm;

[0028] S3. Deploy the trained model on the server at the quay crane, collect images from various perspectives and PLC perception data in each control cycle, input the observation signal into the deep learning model for reasoning, obtain the control sequence signal, and send the control sequence signal to the PLC. The PLC performs the underlying drive of the quay crane motor to realize the automated control of the quay crane.

[0029] Further, the step S3. using the deep neural network system trained by the ACT algorithm specifically includes:

[0030] a. Model structure, using conditional variational autoencoder as the overall framework, where:

[0031] The encoder part is implemented using the Transformer encoder, which receives the vector representation of part of the observation data and the step action sequence in the demonstration data as input, and compresses the output into a potential representation vector through average pooling;

[0032] The decoder part is implemented using a complete Transformer model. The input of the Transformer encoder includes the current vectorized state information and the potential representation vector output by the encoder. The input of the Transformer decoder is a fixed-position embedding, and the output is processed by a linear layer to generate a step action sequence.

[0033] If image data is involved, it is converted into a feature vector through a convolutional neural network such as ResNet18 and then input into the Transformer encoder;

[0034] b. The loss function consists of reconstruction loss and KL divergence loss, specifically:

[0035] Reconstruction loss, used to evaluate the difference between the generated action sequence and the expert demonstration action sequence, using the mean square error loss function;

[0036] KL divergence loss is used to reduce the posterior probability distribution learned by the encoder to the preset prior probability distribution. Both are assumed to be normal distributions, where the prior distribution is a standard normal distribution, and the KL divergence between the two is calculated as the loss;

[0037] The reconstruction loss and the KL divergence loss are added by preset weights to obtain a final loss function;

[0038] c. Action chunking mechanism, which allows the model to predict and sequentially execute a series of actions for each input.

[0039] Furthermore, the model training process includes: using a demonstration data set to train the conditional variational autoencoder, and optimizing the model weight parameters by minimizing the final loss function until a preset convergence condition is met.

[0040] Compared with the prior art, the technical effects of the present invention are as follows:

[0041] 1) Significantly improve automation control performance:

[0042] By introducing deep learning technology, precise automated control of quay crane operations is achieved. The system can learn and imitate the driving behavior of quay crane operators and generate precise control instructions, thus significantly improving the automated control performance of quay crane operations.

[0043] 2) Enhance system robustness and adaptability:

[0044] By collecting and processing multi-angle images, speed and position information recorded by PLC, and control signals from operators, a rich training data set was constructed. This enables the trained deep learning model to maintain high performance under changing operating conditions, enhancing the robustness and adaptability of the system.

[0045] 3) End-to-end automated control process:

[0046] Combining the deep learning model with the PLC underlying driver realizes the end-to-end automated control process from observation data to control instructions. It simplifies the perception and control part of the traditional quay crane automation system and improves the overall operation efficiency.

[0047] 4) Improve operation continuity and stability:

[0048] The closed-loop control system can ensure the continuity, stability and responsiveness of the quay crane operation. The system collects real-time data in each control cycle and inputs it into the deep learning model for reasoning to obtain accurate control sequence signals, thereby achieving stable control of the quay crane.

[0049] 5) Reduce labor and equipment costs:

[0050] The automatic control system of the present invention can significantly reduce labor costs and reduce dependence on quay crane operators. At the same time, by simplifying the perception and control parts, the equipment costs required for some deployment algorithms are saved, and the deployment difficulty and maintenance cost of the overall system are reduced.

[0051] 6) Innovative model training and optimization methods:

[0052] The imitation learning algorithm is used to train the deep neural network, and combined with the structural characteristics of the conditional variational autoencoder, the accurate imitation of the quay crane operation behavior is achieved. At the same time, by introducing the action block mechanism, the model is allowed to predict and sequentially execute a series of actions at each input, further improving the control effect and operation success rate. The model weight parameters are optimized by minimizing the final loss function until the preset convergence conditions are met. This ensures that the trained deep learning model has excellent performance and generalization ability, and can perform well in actual operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 Schematic diagram of a quay crane automation control system based on deep learning of the present invention;

[0054] Figure 2 Schematic diagram of a deep learning neural network based on the ACT algorithm in the present invention, wherein a is the encoder part and b is the decoder part;

[0055] Figure 3 Schematic diagram of deep learning neural network training based on ACT algorithm in the present invention DETAILED DESCRIPTION

[0056] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. The technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the invention and the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0057] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. The experimental methods in the following embodiments without specifying specific conditions are usually carried out under conventional conditions or under conditions recommended by the manufacturer.

[0058] In the face of complex interference factors at the quay crane operation site, as well as the dynamic modeling difficulties caused by the lifting height of the quay crane and the length of the steel rope connecting the trolley and the spreader, this invention adopts cutting-edge deep learning technology. By using deep neural networks, this system can accurately imitate the quay crane driving skills of experienced operators and learn the complex mapping of feedback signals to control commands from operation demonstrations. This imitation learning-based method gives the control system excellent generalization capabilities, and is expected to significantly enhance the robustness and stability of the automated catch and release function, maintaining high performance even under variable operating conditions.

[0059] Specifically, building a quay crane control system includes two key stages: model training and deployment inference.

[0060] The model training phase requires the collection of demonstration data generated by the quay crane operators in actual operations. This includes various observation signals, such as images captured by multi-angle cameras, the speed and position of the trolley and lifting mechanism recorded by the PLC, and control signals, i.e. the lifting speed signal and trolley speed signal expected in quay crane operations. These collected data will undergo a series of preprocessing steps, including data cleaning and normalization, to ensure data quality and consistency. Subsequently, according to the training algorithm, the demonstration data can be used to systematically train the deep neural network so that it learns how to extract valuable information from the observation signals and convert them into precise control instructions.

[0061] The deployment inference phase is to apply the trained model to real-time operations. In each control cycle, the system will capture the observation signal in real time and pass it as input to the model for inference analysis. The model will synthesize this information and infer and output a series of control signals, which will be sent to the PLC through the communication system to perform precise motion control on the quay crane. This process forms a closed-loop control system to ensure the continuity, stability and responsiveness of the quay crane operation.

[0062] See also Figure 1 , Figure 1 This is a schematic diagram of the quay crane automation control system based on deep learning of the present invention. As shown in the figure, sensors such as cameras collect observation data of the current quay crane; then, the model decoder part generates a k-step control action sequence according to the observation data; then, the control instruction is sent to the PLC through the communication system; finally, the PLC drives the motor at the bottom according to the control instruction to form a closed loop of the control system.

[0063] Figure 2 This is a schematic diagram of a deep learning neural network based on the ACT algorithm of the present invention. As shown in the figure, a is the encoder part of the model, and its core function is to efficiently compress and encode the input action sequence to generate a potential representation vector z of a fixed length, which is specifically achieved by the following steps:

[0064] 1) Input layer: receives the original action sequence data.

[0065] 2) Linear layer: Preliminary adjustment of the dimension of the input data in preparation for subsequent processing.

[0066] 3) Transformer encoder: It adopts a stacked architecture of multiple encoders, each of which integrates a self-attention mechanism and a feedforward neural network. This design enables the model to capture complex dependencies in the input sequence and encode this information into the potential representation vector z.

[0067] 4) Average pooling layer: Average pooling is performed on the output of the Transformer encoder to further refine the information and generate the final potential representation vector z.

[0068] (b) is the decoder part of the model, whose task is to generate a latent representation vector z and observation information o according to the encoder. t , generating a reconstructed control action sequence. The decoding process also relies on the powerful capabilities of Transformer. The specific steps are as follows:

[0069] 1) Input fusion: The potential representation vector z and the observation information o t As input, it is initially processed through a linear layer so that the decoder can use both parts of information at the same time;

[0070] 2) Transformer decoder: It adopts a stacked architecture of multiple decoders. Each decoder not only contains a self-attention mechanism, but also introduces a cross-attention mechanism to focus on the output of the encoder. This enables the decoder to be based on the potential representation vector z and the observation information o t , and gradually generate the reconstructed control action sequence.

[0071] 3) Output layer: The output of the decoder is mapped to the space of action sequences through a linear layer to generate the final reconstructed control action sequence.

[0072] This embodiment adopts a conditional variational autoencoder (CVAE) as the overall network framework, which includes two parts: an encoder and a decoder.

[0073] Encoder part:

[0074] Input: Receive the vectorized partial observation data and the k-step action sequence a_(t:t+k) in the demonstration data as input.

[0075] Output: Compressed by average pooling to form a latent representation vector z.

[0076] Decoder part:

[0077] Input: The current vectorized quay crane state and the potential representation vector z output by the encoder, and a fixed position embedding (positional embedding) of length k. For image data, feature extraction can be performed through a convolutional neural network (such as ResNet18). For example, an input image of size 80×80×3 is extracted into a feature vector of shape 1×128 through ResNet18, and then input into the Transformer encoder for subsequent calculations.

[0078] Output: After being projected to the appropriate dimension by the linear layer, the final k-step action sequence is generated.

[0079] This embodiment uses the conditional variational autoencoder framework, combined with the Transformer encoder and decoder, and convolutional neural network for image processing, to achieve effective use of partial observation data and demonstration data and generate a reasonable k-step action sequence. The designed deep learning neural network architecture based on the ACT algorithm is particularly suitable for processing sequence-to-sequence tasks, such as compression and reconstruction of action sequences. Through the encoder's efficient encoding of the input sequence and the decoder's accurate reconstruction based on the latent representation vector and observation information, the model can achieve efficient sequence processing and generation in complex scenarios.

[0080] Figure 3 FIG. 1 is a schematic diagram of deep learning neural network training based on ACT algorithm in the present invention. In each iteration cycle, data pairs are first sampled from the demonstration data set. t 、a t：t+k , input the model encoder part, calculate the mean and variance μ, σ of the potential expression distribution; then, sample the potential expression vector z in the distribution; then, input the corresponding content into the model decoder part to predict the action sequence Finally, the loss function is calculated according to the formula, and the model parameters are updated through back propagation.

[0081] During the training phase, the model follows the following steps for iterative optimization:

[0082] 1. Data sampling:

[0083] Randomly select data pairs from the demonstration dataset D, including the system observation o_t at time t and the corresponding expert demonstration action sequence a_(t:t+k).

[0084] Encoder Processing:

[0085] The sampled data pairs are input into the encoder, which calculates the mean μ and variance σ of the potential expression distribution.

[0086] According to this distribution N(μ,σ), the potential expression vector z is sampled.

[0087] Decoder prediction:

[0088] The potential expression vector z is input into the decoder together with the system's current observation o_t.

[0089] The decoder predicts the action sequence a_(t:t+k) for the next k steps.

[0090] 2. Loss calculation and parameter update:

[0091] Calculate the loss function L, which consists of two parts: reconstruction loss L_reconst and KL divergence loss L_reg, and weigh them by weight β.

[0092] The back-propagation algorithm is used to update the parameters of the encoder and decoder to minimize the loss function.

[0093] Algorithm innovation:

[0094] Action Chunking:

[0095] Borrowing concepts from neuroscience, independent actions are combined into units to perform, thereby improving the smoothness of action execution and the success rate of tasks.

[0096] The model obtains input once every k steps, predicts and sequentially executes the actions of the next k steps, effectively preventing the accumulation of single-step errors.

[0097] Conditional Automatic Variational Encoder Structure:

[0098] CVAE is used as the overall framework during training, and only the decoder is used to generate action sequences during inference.

[0099] This generative model, as a policy network, can better handle and reproduce the noise and randomness in human demonstration data.

[0100] Loss function design:

[0101] Reconstruction loss: The mean squared error loss is used to characterize the difference between the action sequence predicted by the model and the action sequence demonstrated by the expert.

[0102] KL divergence loss: It forces the posterior probability distribution learned by the encoder to be reduced to the assumed prior probability distribution, i.e., the standard normal distribution.

[0103] This paper provides an innovative solution to the crane anti-sway problem by deeply integrating deep learning technology and embodied intelligent design ideas. This solution not only improves the success rate of fine-grained task execution, but also enhances the model's ability to handle noise and randomness in human demonstration data.

[0104] Example:

[0105] S1. Data collection and preprocessing

[0106] Data Collection:

[0107] Use multi-angle cameras installed around the quay crane to capture real-time images during operations.

[0108] The speed, position and other status information of the quay crane trolley and the lifting mechanism are recorded through the PLC system.

[0109] The control signals input by the operator through the control system are recorded simultaneously, including the desired lifting speed and trolley speed.

[0110] Data preprocessing:

[0111] Clean the image data to remove blurry, occluded and other low-quality images.

[0112] Normalize the PLC data to ensure that all input data have consistent dimensions and improve model training efficiency.

[0113] Convert the control signal (desired speed) into a format suitable for model input.

[0114] S2. Model construction

[0115] Encoder build:

[0116] A Transformer encoder is used as the encoder body, and the input includes the vectorized partial observation data and the step action sequence in the demonstration data.

[0117] Features are extracted through the Transformer encoder and the output is compressed into a latent representation vector through average pooling.

[0118] Decoder construction:

[0119] The decoder part adopts the complete Transformer model, including encoder and decoder.

[0120] The input of the Transformer encoder is the current vectorized quay crane state and the potential representation vector output by the encoder.

[0121] The decoder input is a fixed-length position embedding, and the output is projected into a sequence of step actions through a linear layer.

[0122] If an image is input, its features are first extracted through a convolutional neural network such as ResNet18, and then used as an additional input to the Transformer encoder.

[0123] S3. Loss function design

[0124] Reconstruction loss:

[0125] The mean square error loss function is used to calculate the difference between the predicted action sequence and the expert demonstration action sequence.

[0126] The calculation formula is: Lrecon = N1∑t = 1N(at-a^t)2, where at is the action demonstrated by the expert, a^t is the action predicted by the model, and N is the sequence length.

[0127] KL divergence loss:

[0128] Compute the KL divergence between the encoder posterior probability distribution and the standard normal distribution.

[0129] Assuming that both the posterior distribution and the prior distribution are normal distributions, the calculation formula is:

[0130] LKL=21(tr(Σ-1Σ^)+(μ-μ^)TΣ-1(μ-μ^)-log∣Σ∣∣Σ^∣-d)

[0131] where μ, Σ are the mean and covariance of the prior distribution, μ^, Σ^ are the mean and covariance of the posterior distribution, and d is the dimension.

[0132] Total loss function:

[0133] The reconstruction loss and KL divergence loss are weighed and added by weight λ to obtain the total loss function:

[0134] Ltotal=Lrecon+λLKL

[0135] S4. Model training process

[0136] S4.1 Initialization parameters:

[0137] Initialize encoder parameters θe and decoder parameters θd.

[0138] S4.2 Iterative training:

[0139] For each iteration, a data pair (ot, at) is randomly sampled from the dataset, where ot is the observed data and at is the corresponding expert demonstration action.

[0140] The observation data ot is input into the encoder to obtain the potential representation vector z.

[0141] Sample a latent expression vector z′ from the latent distribution q(z|ot,at).

[0142] Input z′ and the current state st into the decoder and predict the action sequence a^t+1:t+T.

[0143] Calculate the total loss function Ltotal.

[0144] The encoder parameters θe and decoder parameters θd are updated using the back-propagation algorithm.

[0145] S4.3 Model evaluation and tuning:

[0146] The model performance was evaluated on an independent validation set, including metrics such as reconstruction error and KL divergence.

[0147] Adjust the model structure, parameters or training strategy based on the evaluation results.

[0148] S4.4 Model deployment:

[0149] The trained model is deployed to the quay crane control system for real-time operation control.

[0150] Observation data is captured in real time, input into the model for reasoning, and control signals are generated and sent to the PLC for execution.

Claims

1. A quay crane automation control system based on deep learning, characterized in that: include: A data collection module is configured to install cameras at key locations of the quay crane and record demonstration data generated during manual operations of the quay crane, including multi-angle images, speed and position information of the quay crane trolley and hoisting mechanism recorded by the PLC, and control signals generated by the operator when driving the quay crane; A data preprocessing module is configured to align the observation data with the control instruction data according to the timestamp to form a data pair, and clean and normalize the data; a model training module, configured to train the deep neural network model using an imitation learning algorithm, using the processed data pairs as training data, the observation data as model input, the control sequence as a target value, and updating the model weights through a back-propagation algorithm; The deployment inference module is configured to deploy the trained model in the server at the quay crane, collect images from various perspectives and PLC perception data in each control cycle, input the observation signal into the deep learning model for inference, obtain the control sequence signal, and send the control sequence signal to the PLC, which performs the underlying drive of the quay crane motor.

2. The deep learning-based quay crane automation control system according to claim 1 is characterized in that: The data collection module includes multiple cameras installed at key positions of the quay crane to capture multi-angle images.

3. The deep learning-based quay crane automation control system according to claim 1 or 2, characterized in that: The data preprocessing module includes a data cleaning submodule for removing abnormal noise points and a data normalization submodule for ensuring data consistency.

4. The deep learning-based quay crane automation control system according to any one of claims 1 to 3, characterized in that: The imitation learning algorithm used by the model training module is capable of learning complex mappings from observation signals to control commands.

5. The deep learning-based quay crane automation control system according to any one of claims 1 to 4, characterized in that: The deployment reasoning module can maintain high performance under changing operating conditions and ensure the continuity, stability and responsiveness of the quay crane operation through a closed-loop control system.

6. A deep learning-based quay crane automation control method, characterized in that: The following steps are involved: S1. Install cameras at key locations of the quay crane and collect demonstration data generated during manual operations of the quay crane, including multi-angle images, observation signals of the speed and position information of the quay crane trolley and lifting mechanism recorded by the PLC, and control signals generated by the operator when driving the quay crane; S2. Align the observation data and the control instruction data according to the timestamp to form a data pair, and clean and normalize the data pair; S3. train the deep neural network model using an imitation learning algorithm, using the data pair as training data, the observed data as model input, the control sequence as the target value, and updating the model weights through a back propagation algorithm; S3. Deploy the trained model on the server at the quay crane, collect images from various perspectives and PLC perception data in each control cycle, input the observation signal into the deep learning model for reasoning, obtain the control sequence signal, and send the control sequence signal to the PLC. The PLC performs the underlying drive of the quay crane motor to realize the automated control of the quay crane.

7. The deep learning-based quay crane automation control method according to claim 1, characterized in that: The step S3. using the deep neural network system trained by the ACT algorithm specifically includes: a. Model structure, using conditional variational autoencoder as the overall framework, where: The encoder part is implemented using the Transformer encoder, which receives the vector representation of part of the observation data and the step action sequence in the demonstration data as input, and compresses the output into a potential representation vector through average pooling; The decoder part is implemented using a complete Transformer model. The input of the Transformer encoder includes the current vectorized state information and the potential representation vector output by the encoder. The input of the Transformer decoder is a fixed-position embedding, and the output is processed by a linear layer to generate a step action sequence. If image data is involved, it is converted into a feature vector through a convolutional neural network such as ResNet18 and then input into the Transformer encoder; b. The loss function consists of reconstruction loss and KL divergence loss, specifically: Reconstruction loss, used to evaluate the difference between the generated action sequence and the expert demonstration action sequence, using the mean square error loss function; KL divergence loss is used to reduce the posterior probability distribution learned by the encoder to the preset prior probability distribution. Both are assumed to be normal distributions, where the prior distribution is a standard normal distribution, and the KL divergence between the two is calculated as the loss; The reconstruction loss and the KL divergence loss are added by preset weights to obtain a final loss function; c. Action chunking mechanism, which allows the model to predict and sequentially execute a series of actions for each input.

8. The deep learning-based quay crane automation control method according to claim 6, characterized in that: The model training process includes: using a demonstration data set to train the conditional variational autoencoder, and optimizing the model weight parameters by minimizing the final loss function until the preset convergence conditions are met.

Images