Real-time risk early warning system and production scheduling method for in-field flow machines based on ai large model capability

Abstract

The present application relates to the technical field of risk early warning, in particular to a kind of real-time risk early warning system and production scheduling method of in-site mobile machinery based on AI big model capability.It includes: data acquisition and processing unit acquires the operating state parameter of in-site mobile machinery, job environment parameter, and the operating state parameter and job environment parameter are preprocessed;AI big model analysis and risk identification unit extracts the feature vector of operating state parameter and job environment parameter, utilizes BERT model and long short-term memory network model, models the interaction relationship of mobile machinery operating track, behavior mode and environment, generates risk index;Risk early warning and response unit generates alarm information based on risk index.The present application introduces the environment modeling method of perception neural radiation field (F-NeRF), converts discrete sensor data into continuous space-time representation of job environment (such as temperature and humidity, visibility, ground conditions).

Description

Technical Field

[0001] This invention relates to the field of risk warning technology, and more specifically, to a real-time risk warning system and production scheduling method for mobile machinery in a production field based on AI large model capabilities. Background Technology

[0002] In enclosed or semi-enclosed operational scenarios such as ports, industrial parks, and large warehouses, the efficient and safe operation of mobile machinery (such as forklifts, AGVs, and loaders) is crucial to overall production and operation. However, with the expansion of operational scale and the increase in environmental complexity, traditional management systems have revealed many limitations in risk identification and production scheduling. Existing risk warning systems mostly rely on fixed threshold alarms or simple rule judgments, making it difficult to integrate and process multi-source heterogeneous data such as machinery operating status (e.g., location, speed, load, energy consumption) and operating environment (e.g., temperature, humidity, visibility, ground conditions). Furthermore, environmental perception is limited to discrete sensor points, lacking the ability to accurately model continuous spatial changes and future path risks, resulting in problems such as high false alarm rates, serious missed alarms, and delayed response in complex dynamic environments. At the same time, traditional methods generally lack a deep understanding of the interaction between machinery behavior patterns and the environment, especially in the event of sudden environmental changes (e.g., fog, slippery conditions), making it difficult to achieve forward-looking risk prediction. More significantly, current safety monitoring systems and production scheduling systems are often independent, forming information silos: risk warnings cannot effectively drive scheduling adjustments, while task allocation often ignores the risk status of machinery, leading to difficulties in coordinating safety and efficiency, and easily resulting in the dilemma of high-risk machinery operating continuously or sacrificing efficiency for safety. Furthermore, existing models have rigid strategies for multimodal data fusion, lacking the ability to dynamically adjust according to environmental context; when a certain perception mode fails, the system's robustness significantly decreases. Therefore, this paper proposes a real-time risk warning system and production scheduling method for mobile machinery within the production facility based on AI large-scale model capabilities. Summary of the Invention

[0003] The purpose of this invention is to provide a real-time risk warning system and production scheduling method for mobile machinery in a field based on AI large model capabilities, in order to solve the problems mentioned in the background art, such as weak environmental perception capability, lagging risk identification, disconnect between safety and production scheduling, and insufficient multimodal fusion intelligence of traditional mobile machinery management systems in complex dynamic environments.

[0004] To achieve the above objectives, on the one hand, the present invention aims to provide a real-time risk early warning system for mobile machinery in a field based on AI large model capabilities, comprising:

[0005] The data acquisition and processing unit collects the operating status parameters and working environment parameters of the mobile machinery in the field, and preprocesses the operating status parameters and working environment parameters.

[0006] The AI ​​large model analysis and risk identification unit extracts feature vectors of operating status parameters and working environment parameters. It uses the BERT model and the long short-term memory network model to model the interaction between the mobile machinery's operating trajectory, behavior pattern and environment, and generates a risk index. In the process of extracting the feature vectors of working environment parameters, the unit uses the perceptual neural radiation field environment modeling method to enhance the environmental feature vectors of the working environment parameters.

[0007] A risk warning and response unit, which generates alarm information based on a risk index;

[0008] The production scheduling optimization unit integrates operating status parameters, work environment parameters, and risk indices to model the mechanical operation capacity and resource constraints, and generates feasible resource allocation schemes.

[0009] As a further improvement to this technical solution, the data acquisition and processing unit includes a data acquisition module and a data processing module;

[0010] The data acquisition module is used to collect the operating status parameters and working environment parameters of the mobile machinery in the field.

[0011] The data processing module is used to preprocess the collected operating status parameters and working environment parameters.

[0012] As a further improvement to this technical solution, the AI ​​large model analysis and risk identification unit includes a feature extraction module and a risk modeling module;

[0013] The feature extraction module extracts operating state features and environmental features based on preprocessed operating state parameters and operating environment parameters, and forms a high-dimensional temporal feature vector by using a multimodal feature encoding method.

[0014] The risk modeling module is based on high-dimensional time-series feature vectors and uses the BERT model and long short-term memory network model to establish a risk prediction model of the interaction between the operating behavior of mobile machinery and the environment, and generates a risk index.

[0015] As a further improvement to this technical solution, the feature extraction module extracts running state features and environmental features through a multimodal feature encoding method to form a high-dimensional temporal feature vector, including the following steps:

[0016] S1.1. Recurrent neural networks are used to encode the operating state parameters, extract the mechanical motion trajectory and dynamic features, and obtain the temporal operating state feature vector.

[0017] S1.2. The working environment parameters are encoded using the sensory neural radiation field environmental modeling method, and the spatial distribution characteristics and spatiotemporal variation characteristics of the environment are extracted to obtain the environmental feature vector.

[0018] S1.3 Synchronize the running state feature vector with the environment feature vector in time and align the feature dimensions, and obtain the fused feature representation through the attention mechanism;

[0019] S1.4 Based on fusion feature representation, a long short-term memory network is introduced to identify the dynamic dependency between the operating state and environmental changes, forming a temporal context feature representation;

[0020] S1.5 Output high-dimensional temporal feature vectors.

[0021] As a further improvement to this technical solution, in step S1.2, the working environment parameters are encoded using a sensory neural radiation field environment modeling method, including the following steps:

[0022] S1.21 Construct continuous spatiotemporal query coordinates, perform high-frequency position encoding on the spatiotemporal query coordinates, and obtain high-dimensional feature representations;

[0023] S1.22. Input the encoded high-dimensional features into the multilayer perceptron, and use the trained neural radiation field model to infer the query coordinates and output the environmental parameter vector corresponding to the location and time.

[0024] S1.23. Input the continuously collected operating status parameters and working environment parameters as training samples into the neural radiation field model, and fine-tune the parameters of the neural radiation field model online based on the continuous learning strategy.

[0025] S1.24. Using the current position of the machine and the spatiotemporal query coordinates on the predicted path as query points, the predicted values ​​of local environmental parameters and the corresponding uncertainty weights are obtained using the neural radiation field model.

[0026] S1.25. Combine the predicted values ​​of local environmental parameters for each query point with the uncertainty weights to form an environmental feature vector.

[0027] As a further improvement to this technical solution, the risk modeling module establishes a risk prediction model of the interaction between the operating behavior of mobile machinery and the environment, and generates a risk index, including the following steps:

[0028] S2.1. Preprocess the high-dimensional temporal feature vector, and input the preprocessed high-dimensional temporal feature vector into the BERT model to generate semantically enhanced feature representations. ;

[0029] S2.2 Input the high-dimensional temporal feature vector into the Long Short-Term Memory network model to identify the temporal dependency between changes in operating state and environmental dynamics, and output the temporal feature representation. ;

[0030] S2.3. The temporal feature representation output by the Long Short-Term Memory (LSTM) network model and the semantically enhanced feature representation output by the BERT model are fused through a multimodal dynamic adaptive fusion network to generate a joint feature vector. ;

[0031] S2.4, Combine the feature vectors Input the risk prediction model, use a multilayer perceptron for reasoning, and output the risk index.

[0032] As a further improvement to this technical solution, in step S2.3, fusion is performed through a multimodal dynamic adaptive fusion network, including the following steps:

[0033] S2.31, Based on temporal feature representation, semantically enhanced feature representation, and contextual information of the current environment. Construct feature concatenation vector The concatenated feature vector is input into the gating function to calculate the gating vector. ;

[0034] S2.32. Using gating vectors, the temporal feature representation and the semantically enhanced feature representation are weighted and fused to generate the final joint feature vector. .

[0035] As a further improvement to this technical solution, the risk warning and response unit classifies risk levels based on a risk index and a preset threshold, and generates alarm content according to the risk level.

[0036] As a further improvement to this technical solution, the production scheduling optimization unit integrates the current status and task queue of the mobile machinery in the field based on the operating status parameters, working environment parameters, and risk level of the mobile machinery in the field to form real-time scheduling input data. It also calculates the task priority weight by combining the production target, task urgency, and risk level of each machine, models the working capacity, current position, energy consumption level, and remaining working time of the available machines, generates a feasible resource allocation set, and generates a globally optimal scheduling scheme based on the feasible resource allocation set, and outputs scheduling instructions.

[0037] On the other hand, the present invention provides a production scheduling method for mobile machinery in a factory based on AI large model capabilities, which is used in the real-time risk early warning system for mobile machinery in a factory based on AI large model capabilities described in any one of the above, and includes the following steps:

[0038] S3.1 Collect the operating status parameters and working environment parameters of the mobile machinery in the field, and preprocess the operating status parameters and working environment parameters;

[0039] S3.2 Extract the feature vectors of operating status parameters and working environment parameters. Use the BERT model and the long short-term memory network model to model the interaction between the mobile machinery's operating trajectory, behavior pattern and environment, generate a risk index, and enhance the environmental feature vectors of the working environment parameters by using the sensory neural radiation field environmental modeling method during the extraction of the feature vectors of the working environment parameters.

[0040] S3.3 Generate alarm information based on risk index;

[0041] S3.4 Integrate operating status parameters, work environment parameters, and risk indices to model the mechanical operation capabilities and resource constraints, and generate feasible resource allocation schemes.

[0042] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0043] 1. This invention relates to a real-time risk warning system and production scheduling method for mobile machinery in a field based on AI large-scale model capabilities. It introduces a sensory neural radiation field (F-NeRF) environmental modeling method, transforming discrete sensor data into a continuous spatiotemporal representation of the working environment (such as temperature, humidity, visibility, and ground conditions), and outputting an environmental feature vector containing uncertain weights, thus achieving refined modeling of complex dynamic environments. Based on this, a multimodal dynamic adaptive fusion network is used, combining the semantic reasoning capabilities of the BERT large-scale model with the temporal modeling capabilities of LSTM, and introducing an environment context-driven gating mechanism. When harsh environments (such as low visibility) lead to perception degradation, the feature fusion weights are automatically adjusted, enhancing the reliance on high-confidence modalities (such as semantic rules). This mechanism effectively overcomes the problems of rigid feature fusion and high risk misjudgment rates in complex environments of traditional methods, significantly improving the system's risk identification accuracy and robustness in unstructured and dynamically changing scenarios.

[0044] 2. The present invention relates to a real-time risk early warning system and production scheduling method for mobile machinery in a production field based on AI large model capabilities. This system not only achieves real-time risk early warning but also deeply integrates risk indices and levels into the production scheduling decision-making closed loop. By constructing a dynamic model that integrates operating status, environmental parameters, and risk levels, and combining task priority weights and resource constraints, a feasible resource allocation set is generated. Furthermore, a multi-objective optimization algorithm is used to generate a globally optimal scheduling scheme that balances path safety, energy consumption costs, and task urgency. Attached Figure Description

[0045] Figure 1 This is an overall flowchart of the present invention;

[0046] The meanings of the labels in the diagram are as follows:

[0047] 1. Data acquisition and processing unit; 2. AI large-scale model analysis and risk identification unit; 3. Risk warning and response unit; 4. Production scheduling optimization unit. Detailed Implementation

[0048] 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0049] Example 1: Please refer to Figure 1 As shown, a real-time risk early warning system for mobile machinery in a field, based on AI large model capabilities, is provided, including:

[0050] The data acquisition and processing unit 1 collects the operating status parameters and working environment parameters of the mobile machinery in the field, and preprocesses the operating status parameters and working environment parameters.

[0051] In this embodiment, the data acquisition and processing unit 1 includes a data acquisition module and a data processing module;

[0052] The data acquisition module is used to collect the operating status parameters (including position, speed, acceleration, load, and energy consumption) and operating environment parameters (including temperature, humidity, visibility, and ground conditions) of the mobile machinery in the field.

[0053] The data processing module is used to preprocess the collected operating status parameters and working environment parameters, namely, to perform time synchronization, noise reduction and standardization preprocessing, so as to ensure the accuracy of subsequent analysis.

[0054] AI large model analysis and risk identification unit 2 extracts feature vectors of operating status parameters and working environment parameters. Using BERT model and long short-term memory network model, it models the interaction between the operating trajectory, behavior pattern and environment of mobile machinery, generates risk index, and enhances the environmental feature vectors of working environment parameters by using the perceptual neural radiation field environmental modeling method during the extraction of working environment parameter feature vectors.

[0055] In this embodiment, the AI ​​large model analysis and risk identification unit 2 includes a feature extraction module and a risk modeling module;

[0056] The feature extraction module extracts operating state features and environmental features based on preprocessed operating state parameters and operating environment parameters, and forms a high-dimensional temporal feature vector by using a multimodal feature encoding method.

[0057] Furthermore, the feature extraction module extracts runtime state features and environmental features using a multimodal feature encoding method to form a high-dimensional temporal feature vector, including the following steps:

[0058] S1.1 A recurrent neural network (RNN) is used to encode the operating state parameters, extract the mechanical motion trajectory and dynamic features, and obtain a temporal operating state feature vector. Specifically: First, the sequence of operating state parameters (such as position, velocity, acceleration, load, and energy consumption) collected from the flowing machinery in the field is time-synchronized and standardized. Then, the parameter vector of each time step is used as the input sequence of the RNN and input into the network sequentially. The recurrent neural network identifies the continuity and dynamic dependence of mechanical motion by retaining information from the previous time step in the hidden state, and outputs the hidden state vector at each time step. Finally, the hidden states of the entire sequence (which can be the last time step or the concatenation / pooling of all time steps) are used to form a temporal operating state feature vector, which comprehensively reflects the mechanical operation. The trajectory, velocity changes, and dynamic characteristics provide input for subsequent multimodal feature fusion. The recurrent neural network (RNN) consists of an input layer, hidden layers, and an output layer. The input layer receives the running state parameter vector at each time step. The hidden layer contains one or more RNN units (optionally GRU or LSTM units to enhance long sequence dependency capabilities). Each unit recursively updates the hidden state in the time dimension, introducing non-linearity through an activation function (ReLU function) to output the hidden state at the current time step. The output layer can directly output the hidden state vector or, as needed, obtain a feature representation through linear transformation for subsequent feature fusion or risk modeling. The network supports multi-layer stacking to increase modeling capacity and can enhance training stability through methods such as dropout or LayerNorm.

[0059] S1.2. The working environment parameters are encoded using the sensory neural radiation field environmental modeling method, and the spatial distribution characteristics and spatiotemporal variation characteristics of the environment are extracted to obtain the environmental feature vector.

[0060] The NeRF (Neural Radiation Field) environmental modeling method aims to address the core problems of information sparsity, discontinuity, and lack of predictive ability inherent in traditional environmental sensing methods within complex and dynamic industrial environments. Traditional methods rely on limited environmental sensors (such as discrete thermometers, hygrometers, and visibility meters) to provide point-like data, failing to construct continuous, high-precision maps of the entire environmental state. This results in insufficient perception of micro-environmental changes (such as localized fog or oil stains) around machinery and throughout the entire work area. Furthermore, traditional methods typically only provide instantaneous snapshots of the current moment, failing to effectively model the continuous changes in environmental parameters (such as visibility and ground friction coefficient) over time and space. Consequently, it is difficult to predict potential environmental risks encountered by machinery along future movement paths, leading to a lack of proactive early warning. The NeRF environmental modeling method, however, offers the advantage of continuous, high-precision, predictable, and quantifiable uncertainty modeling capabilities. It transforms sparse, discrete sensor data into a continuous digital twin environmental field, enabling the querying of environmental parameters (such as temperature and visibility) at any location and time. This represents a leap from point-to-field perception, providing unprecedented granularity of environmental insight. More importantly, it possesses spatiotemporal prediction capabilities, enabling it to infer environmental conditions along the future trajectory of machinery, transforming risk warning from a passive response to an active anticipation. Simultaneously, its unique uncertainty quantification output (cognitive and accidental uncertainty) provides reliability weights for subsequent risk fusion and decision-making. The system can use this to determine the confidence level of environmental predictions, thereby adopting more conservative strategies when data is unreliable, greatly enhancing the robustness and security of the entire system in a real, uncertain world.

[0061] The working environment parameters are encoded using a sensory neural radiation field environmental modeling method, including the following steps:

[0062] S1.21 Constructing continuous spatiotemporal query coordinates (in The coordinates of the work area are the planar position coordinates. (For time), high-frequency position encoding or multi-resolution hash encoding is performed on the spatiotemporal query coordinates to obtain a high-dimensional feature representation, so that the subsequent neural network can learn the detailed features of the environment;

[0063] S1.22. Input the encoded high-dimensional features into a multilayer perceptron (MLP), and use the trained neural radiation field model to infer the query coordinates, outputting the environmental parameter vector corresponding to the location and time. ,in Indicates temperature. Indicates humidity. The model outputs an uncertainty index for the prediction results, which represents visibility, including cognitive uncertainty and accidental uncertainty, to characterize the reliability of the prediction values.

[0064] S1.23. Input the operating status parameters and working environment parameters continuously collected by the field sensors into the neural radiation field model as training samples (the operating status parameters and working environment parameters are continuously acquired by the field sensors). Based on the continuous learning strategy, fine-tune the parameters of the neural radiation field model online so that the neural radiation field model can evolve over time and adapt to the dynamic changes of the environment.

[0065] In this embodiment, the neural radiation field model is used to map the continuous spatiotemporal query coordinates of the mobile machinery operating area to environmental parameters and output the prediction uncertainty weights corresponding to each parameter. This includes: an input layer receiving query coordinates and optional condition vectors, and generating high-dimensional feature representations through high-frequency position encoding or multi-resolution hash encoding; a core network consisting of a multilayer perceptron (MLP) combined with residual connections and conditional modulation to extract spatiotemporal environmental features; and an output layer regressing environmental parameter values ​​and random uncertainties through at least two heads, while simultaneously estimating cognitive uncertainty through model ensemble, MC-Dropout, or a Bayesian last layer. During training, the model employs negative log-likelihood loss and optional spatial smoothing constraints, supporting online fine-tuning based on on-site sensor data, allowing the model to evolve over time to adapt to dynamic environmental changes. The output environmental parameters and uncertainty weights can be used for mobile machinery risk modeling and gating fusion, achieving high-precision, dynamically adaptive environmental perception.

[0066] The online fine-tuning process is as follows: Operating status parameters and work environment parameters continuously collected by on-site sensors are organized into training samples according to time series, and then combined with existing training data or historical buffer data in small batches and input into the neural radiation field model. Next, the error between the model's predicted values ​​and the measured values ​​is calculated based on a predefined loss function (negative log-likelihood loss), and the model parameters are updated through gradient backpropagation. During the update process, strategies such as low learning rate, small batches, or elastic weight constraints are employed to prevent catastrophic forgetting. Simultaneously, new and old data can be weighted or sample replay can be used to allow the model to gradually adapt to dynamic environmental changes, thereby achieving online fine-tuning of the neural radiation field model so that its predicted environmental parameters and uncertainties can reflect the actual state of the current on-site work environment in real time.

[0067] S1.24. When risk analysis of mobile machinery is required, the spatiotemporal query coordinates of the current position of the machinery and the predicted path are used as query points. The neural radiation field model is used to obtain the predicted values ​​of local environmental parameters and the corresponding uncertainty weights (i.e., when risk analysis of mobile machinery is required, it is necessary to predict the positions and corresponding times that the machinery may pass through in the future, and input the query coordinates composed of these positions and times into the neural radiation field (F-NeRF) model to obtain the predicted values ​​of environmental parameters on the future path).

[0068] S1.25. Combine the predicted values ​​of local environmental parameters for each query point with the uncertainty weights to form an environmental feature vector;

[0069] S1.3. Synchronize the running state feature vector with the environment feature vector in time and align the feature dimensions, and achieve a unified multimodal representation through an attention mechanism to obtain a fused feature representation. Specifically, the encoded running state feature vector and the environment feature vector are time-aligned and dimension-matched. Then, the two are fed into the attention mechanism module as multimodal inputs. By calculating the attention weight of each modality at different time steps or feature dimensions, the importance of the running state and environment features to the overall representation at the current moment is dynamically measured. Then, the features of each modality are weighted and summed or linearly transformed to generate a unified fused feature representation.

[0070] S1.4. Based on fused feature representation, a Long Short-Term Memory (LSTM) network is introduced to identify the dynamic dependency between operating state and environmental changes, forming a temporal context feature representation. The fused feature representation obtained through multimodal feature encoding is used as the input sequence of the LSTM network. The LSTM performs nonlinear transformations on the input features and the previous hidden state at each time step through its gating mechanism (input gate, forget gate, and output gate), dynamically retaining key operating state and environmental change information while discarding irrelevant information, thereby identifying the temporal dependency between mechanical motion and environmental parameters. After sequence processing, the hidden state of the last time step or pooling of the hidden states of all time steps can be used to generate a comprehensive temporal context feature representation. This feature vector reflects both the dynamic evolution of mechanical behavior and integrates environmental change information, providing a high-dimensional temporal context input for subsequent risk modeling.

[0071] S1.5 Output high-dimensional temporal feature vectors, i.e., temporal context feature representations.

[0072] The risk modeling module is based on high-dimensional time-series feature vectors and uses the BERT model and the Long Short-Term Memory Network (LSTM) model to establish a risk prediction model of the interaction between the operating behavior of mobile machinery and the environment, and generates a risk index.

[0073] The risk modeling module establishes a risk prediction model of the interaction between the operating behavior of mobile machinery and the environment, and generates a risk index, including the following steps:

[0074] S2.1. Preprocess the high-dimensional temporal feature vector, and input the preprocessed high-dimensional temporal feature vector into the BERT model to generate semantically enhanced feature representations. The BERT model, based on its pre-trained knowledge and reasoning capabilities, performs semantic-level pattern recognition and causal relationship inference on the input temporal representation, outputting semantically enhanced feature representations. The BERT model architecture includes an input layer, intermediate encoding layers, and an output layer: the input layer receives high-dimensional temporal feature vectors generated by a multimodal feature encoding module, and can be supplemented with contextual information or positional encoding to enhance sequence perception capabilities; the intermediate layers are multi-layer self-attention transformation modules, each containing a multi-head attention mechanism and a feedforward fully connected network, identifying long-range dependencies and semantic associations between features through residual connections and layer normalization, achieving semantic enhancement and causal relationship inference; the output layer generates semantically enhanced feature representations through linear mapping.

[0075] S2.2. In parallel, the high-dimensional temporal feature vector is input into a Long Short-Term Memory (LSTM) network model to identify the temporal dependency between changes in operating state and environmental dynamics, and output the temporal feature representation. The Long Short-Term Memory (LSTM) network model architecture includes an input layer, intermediate LSTM layers, and an output layer. The input layer receives a sequence of fused feature representations generated by a multimodal feature encoding module, with each time step corresponding to a high-dimensional feature vector. The intermediate layer consists of one or more stacked LSTM units, each containing an input gate, a forget gate, and an output gate. The gating mechanism controls the retention and forgetting of information to identify the temporal dependencies between the machine's operating state and environmental changes, and recursively updates the hidden state in the time dimension. The output layer can use the hidden state of the last time step or perform pooling / linear transformation on the hidden states of all time steps to generate a temporal feature representation.

[0076] S2.3. The temporal feature representation output by the Long Short-Term Memory (LSTM) network model and the semantically enhanced feature representation output by the BERT model are fused through a multimodal dynamic adaptive fusion network to generate a joint feature vector. This vector contains both low-level operational state features and high-level semantic reasoning information.

[0077] S2.4, Combine the feature vectors Input risk prediction model (where the gating vector is used) It is also fed as an additional input into the risk prediction model, which uses a multilayer perceptron for reasoning and outputs a risk index. , (This formula learns the interaction between the operating state parameters and working environment parameters of mobile machinery by inputting them together, thereby modeling the risk level of operating behavior under different environmental conditions and outputting a risk index.) In the formula, The risk prediction model architecture is a multilayer perceptron, consisting of an input layer, intermediate multilayer perceptron (MLP) layers, and an output layer. The input layer receives a joint feature vector generated by a multimodal dynamic adaptive fusion network. and gate vector The intermediate layer is a multi-layer fully connected neural network, each layer containing several neurons and using activation functions (such as ReLU or GELU), batch normalization, and optional dropout to identify non-linear relationships in joint features and enhance expressive power; the output layer generates a single risk index through linear transformation.

[0078] Furthermore, by fusing temporal feature representations and semantically enhanced feature representations through a multimodal dynamic adaptive fusion network, the fundamental challenges faced by traditional methods in complex and dynamic industrial environments are addressed. Traditional systems typically use static weights to fuse multi-source information, which cannot cope with the problem of sensor reliability changing drastically with the environment (such as heavy fog or slippery ground), leading to false alarms and missed alarms. Their models lack deep semantic understanding capabilities, making it difficult to distinguish the true intentions behind behavioral patterns (such as emergency obstacle avoidance versus normal operation), resulting in inaccurate warnings. At the same time, they often ignore the inherent uncertainties of data and the environment, making decision-making vulnerable in critical situations. The multimodal dynamic adaptive fusion network, by calculating gating vectors in real time, can dynamically adjust the fusion weights of different modalities (such as sensor temporal data and semantic inference from large AI models) according to the current environmental context (such as visibility). In harsh environments, it relies more on reliable semantic rules rather than malfunctioning sensors, achieving unprecedented environmental adaptability. Simultaneously, it is the first to combine the high-level semantic inference of BERT large models (such as causal judgments and safety rules) with the low-level temporal perception depth of LSTM, enabling the system to not only see behavior but also understand intent, greatly improving the accuracy of risk identification. Furthermore, the system explicitly models and outputs environmental uncertainty through neural radiation field (NeRF), and uses it as a key input for fusion and decision-making, ultimately achieving a more robust and reliable risk warning than any single model.

[0079] Fusion is performed using a multimodal dynamic adaptive fusion network, including the following steps:

[0080] S2.31, Based on temporal feature representation, semantically enhanced feature representation, and contextual information of the current environment. Construct feature concatenation vector The concatenated feature vector is input into the gating function to calculate the gating vector. , ( It is the context information of the current environment (the context information of the current environment) (Key information from the neural radiation field model (F-NeRF), including visibility values ​​and ground friction coefficient, is obtained after the neural radiation field model predicts local environmental parameters and uncertainties. Specifically, it selects the most critical parameters for risk assessment (including visibility and ground friction coefficient) from the environmental feature vector generated in step S1.25 to form the current environmental context information.) The gate function is the Sigmoid function, which maps the weighted result to the interval [0,1] and is used as the gate weight. This is the weight matrix, used to perform linear transformations on the concatenated features. (This is a bias vector used to adjust the offset of the linear transformation result, making the gating computation more flexible.) This step allows the fusion strategy to dynamically adjust with changes in the environment. When environmental conditions are harsh, such as extremely low visibility, the temporal perception information from sensors (such as cameras) may be unreliable; in this case, the gating mechanism will adjust according to the environmental context. Automatic weight adjustment relies more on the semantic reasoning results provided by the BERT model, such as following safety rules like low-speed driving, while relatively reducing the reliance on potentially invalid temporal awareness features, thereby achieving dynamic and intelligent fusion of multimodal information;

[0081] S2.32. Using gating vectors, the temporal feature representation and the semantically enhanced feature representation are weighted and fused to generate the final joint feature vector. , .

[0082] Risk warning and response unit 3 generates alarm information based on the risk index;

[0083] In this embodiment, the risk warning and response unit 3 classifies risk levels based on a risk index and a preset threshold, and generates alarm content according to the risk level, including the risk index, level, possible events, and suggested measures. Specifically, it compares the risk index output by the risk prediction model with preset multi-level safety thresholds, and divides the risk index into four levels: when the risk index... When the risk level is below the safety threshold 'a', the mobile machinery or operation is considered safe; when it is between the safety threshold 'a' and the risk threshold 'b', it is considered low risk; when it is between the risk threshold 'b' and the high threshold 'c', it is considered medium risk; and when it is above the high threshold 'c', it is considered high risk. Subsequently, alarm content is generated based on the risk level, including the risk level, risk index, type of event (such as collision, speeding, fatigue driving, or environmental obstacles), and corresponding suggested measures (such as slowing down, adjusting the route, or suspending the operation). The alarm information is then sent to the operating terminal and the mobile machinery in real time so that the operator or the system can take timely intervention measures to achieve immediate early warning and management of potential risks.

[0084] Production scheduling optimization unit 4 integrates operating status parameters, work environment parameters and risk index to model the mechanical operation capacity and resource constraints, and generate feasible resource allocation schemes.

[0085] In this embodiment, the production scheduling optimization unit 4 integrates the current status and task queue of the mobile machinery in the field based on the operating status parameters, working environment parameters, and risk level of the mobile machinery in the field to form real-time scheduling input data. It also calculates the task priority weight by combining the production target, task urgency and the risk level of each machine. It models the working capacity, current position, energy consumption level and remaining working time of the available machines, generates a feasible resource allocation set, and generates a globally optimal scheduling scheme based on the feasible resource allocation set. It outputs scheduling instructions that include machine operation path planning, task allocation and operation rhythm adjustment.

[0086] Specifically, the process involves: First, integrating the operational status parameters of mobile machinery within the site (such as location, speed, load, and energy consumption), environmental parameters (such as temperature, humidity, visibility, and ground conditions), and the risk level output from the risk prediction module into the task queue information to form real-time scheduling input data. Then, combining production targets and the urgency of each task, a priority weight is assigned to each task (quantifying the importance of each task based on production targets, e.g., key production nodes or high-output tasks receive higher scores; simultaneously, the time sensitivity of the task is assessed based on its urgency, including deadlines, remaining available resources, and current risk level). Next, the production target score and urgency score are weighted and combined according to preset weights to obtain a comprehensive priority index for each task. Finally, the tasks are ranked according to their comprehensive priority index. The ranking results are used as the basis for scheduling optimization and resource allocation. Simultaneously, the availability of each machine is adjusted according to its risk level. Next, a dynamic model is established for the operating capacity, current location, energy consumption level, and remaining operating time of each available machine, generating a set of feasible resource allocations that meet resource constraints and task requirements. Based on this, an optimization calculation engine (mixed integer programming) is invoked to comprehensively consider path length, operating efficiency, energy consumption cost, and risk avoidance requirements, performing multi-objective scoring and ranking of feasible allocation schemes to generate a globally optimal scheduling scheme. Finally, the generated scheduling scheme is converted into specific scheduling instructions, including machine operation path planning, task allocation, and operation rhythm adjustment, and issued to mobile machinery and operating terminals to achieve dynamic collaborative optimization of production efficiency and operational safety.

[0087] The dynamic model includes constraints and states that describe the changes in the operating capacity, current position, energy consumption level, and remaining operating time of each machine over time, including:

[0088] Position update formula (dynamic constraints):

[0089] ;

[0090] Energy consumption update formula:

[0091] ;

[0092] Remaining task time update formula:

[0093] ;

[0094] Feasibility constraints:

[0095] , , ;

[0096] In the formula, For the first The current location of the Taiwanese machinery For time, For velocity vectors, For time step, For cumulative energy consumption, For the first Current load of the Taiwanese machinery This is a load-speed related energy consumption function. , The load linear consumption coefficient ranges from 0.5 to 5 kW / t. For high-efficiency electric machinery (such as new AGVs), the value approaches the lower limit; for traditional hydraulic transmission internal combustion machinery (such as loaders), the value approaches the upper limit. It is calculated by linear regression based on steady-state power consumption test data under no-load (load is 0) and rated load conditions. The velocity square drag coefficient (such as rolling resistance or air resistance) ranges from 0.01 to 0.20 kW / (m / s)². Larger values ​​are assigned to bulky machinery with high wind resistance (such as large container forklifts), while smaller values ​​are assigned to smaller machinery with low tire resistance (such as small electric pallet trucks). It is obtained by fitting a quadratic function to the power consumption test data corresponding to different constant speeds under no-load conditions. The load-speed coupling energy consumption coefficient (representing the additional energy consumption under the combined effect of load and speed) ranges from 0.005 to 0.05 kW / (t·m / s). It is determined by multivariate nonlinear regression using power consumption test data containing different load and speed combinations. For the remaining work time, This is an indicator function that indicates whether the machine is currently performing a task, and if so, the remaining operation time. Decrease over time; if the machine is not in operation, the remaining time remains unchanged. The feasible space for the work area. This represents the maximum permissible energy consumption.

[0097] In this embodiment, the process of generating the feasible resource allocation set is as follows: Based on priority weights and combined with the operating status parameters (location, speed, load, energy consumption), working environment parameters (temperature and humidity, visibility, ground conditions), and risk level of each mobile machine, machines that meet the location constraints, energy consumption constraints, and remaining working time constraints are selected to form feasible machine-task allocation pairs; for example, at a certain moment, the task set includes path clearing task T1 (high priority), material handling task T2 (medium priority), and site inspection task T3 (low priority), and the machine set includes machine M1 (low risk, current location close to area T1), machine M2 (medium risk, sufficient remaining working time), and machine M3 (high risk, poor visibility and energy consumption close to the upper limit); after constraint screening, the feasible resource allocation set is {(M1,T1), (M2,T2)}, while (M3,T3) is eliminated due to its excessively high risk level. Subsequently, the optimization calculation engine performs multi-objective sorting on the feasible set, generates a globally optimal scheduling scheme, and translates it into specific scheduling instructions: M1 executes path clearing task T1, with the path planned as the shortest safe path P1, and automatically decelerates in areas with low visibility; M2 executes material handling task T2, with the path planned as the energy-optimal path P2, and adjusts the work rhythm to match the remaining work time; T3 is automatically postponed due to the lack of available low-risk machinery, awaiting subsequent scheduling. Finally, the scheduling instructions are sent to the operation terminal and machinery in real time, realizing the dynamic, safe, and efficient collaborative execution of production tasks.

[0098] Example 2: The difference between Example 2 and Example 1 is that this example introduces a risk analysis method used in a real-time risk warning system for mobile machinery in a field based on AI large model capabilities.

[0099] A method for scheduling production of mobile machinery within a production facility based on AI large-scale model capabilities, used in any of the above-mentioned real-time risk early warning systems for mobile machinery within a production facility based on AI large-scale model capabilities, includes the following steps:

[0100] S3.1 Collect the operating status parameters and working environment parameters of the mobile machinery in the field, and preprocess the operating status parameters and working environment parameters;

[0101] S3.2 Extract the feature vectors of operating status parameters and working environment parameters. Use the BERT model and the long short-term memory network model to model the interaction between the mobile machinery's operating trajectory, behavior pattern and environment, generate a risk index, and enhance the environmental feature vectors of the working environment parameters by using the sensory neural radiation field environmental modeling method during the extraction of the feature vectors of the working environment parameters.

[0102] S3.3 Generate alarm information based on risk index;

[0103] S3.4 Integrate operating status parameters, work environment parameters, and risk indices to model the mechanical operation capabilities and resource constraints, and generate feasible resource allocation schemes.

[0104] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. An AI large model capability-based real-time risk warning system for in-field flow machines, characterized in that, include: The data acquisition and processing unit (1) collects the operating status parameters and working environment parameters of the mobile machinery in the field, and preprocesses the operating status parameters and working environment parameters. AI large model analysis and risk identification unit (2), the AI ​​large model analysis and risk identification unit (2) extracts the feature vectors of the operating status parameters and the working environment parameters, uses the BERT model and the long short-term memory network model to model the interaction between the mobile machinery's operating trajectory, behavior pattern and environment, generates a risk index, and uses the sensory neural radiation field environment modeling method to enhance the environmental feature vectors of the working environment parameters during the extraction of the feature vectors of the working environment parameters. Risk warning and response unit (3), which generates alarm information based on risk index; Production scheduling optimization unit (4), the production scheduling optimization unit (4) integrates operating status parameters, work environment parameters and risk index, models the mechanical operation capacity and resource constraints, and generates feasible resource allocation schemes; The risk warning and response unit (3) divides risk levels based on the risk index and a preset threshold, and generates alarm content according to the risk level. The production scheduling optimization unit (4) integrates the current status and task queue of the mobile machinery in the field based on the operating status parameters, working environment parameters, and risk level of the mobile machinery in the field, forming real-time scheduling input data. It also calculates the task priority weight by combining the production target, task urgency and the risk level of each machine, models the working capacity, current position, energy consumption level and remaining working time of the available machines, generates a feasible resource allocation set, generates the global optimal scheduling scheme based on the feasible resource allocation set, and outputs scheduling instructions.

2. The AI big model capability-based real-time risk early warning system for in-field flow machinery according to claim 1, characterized in that: The data acquisition and processing unit (1) includes a data acquisition module and a data processing module; The data acquisition module is used to collect the operating status parameters and working environment parameters of the mobile machinery in the field. The data processing module is used to preprocess the collected operating status parameters and working environment parameters.

3. The AI big model capability-based real-time risk early warning system for in-field flow machinery according to claim 1, characterized in that: The AI ​​large model analysis and risk identification unit (2) includes a feature extraction module and a risk modeling module; The feature extraction module extracts operating state features and environmental features based on preprocessed operating state parameters and operating environment parameters, and forms a high-dimensional temporal feature vector by using a multimodal feature encoding method. The risk modeling module is based on high-dimensional time-series feature vectors and uses the BERT model and long short-term memory network model to establish a risk prediction model of the interaction between the operating behavior of mobile machinery and the environment, and generates a risk index.

4. The AI big model capability-based real-time risk early warning system for in-field flow machinery according to claim 3, characterized in that: The feature extraction module extracts runtime features and environmental features using a multimodal feature encoding method to form a high-dimensional temporal feature vector, including the following steps: S1.

1. Recurrent neural networks are used to encode the operating state parameters, extract the mechanical motion trajectory and dynamic features, and obtain the temporal operating state feature vector. S1.2, encode the operation environment parameters by using the perception neural radiation field environment modeling method, extract the spatial distribution characteristics and spatio-temporal variation characteristics of the environment, and obtain an environment feature vector; S1.3, time-synchronize and align the feature dimensions of the running state feature vector and the environment feature vector, and obtain a fused feature representation through an attention mechanism; S1.4, based on the fused feature representation, introduce a long short-term memory network to identify the dynamic dependency relationship between the running state and the environment change, and form a time sequence context feature representation; S1.5, output a high-dimensional time sequence feature vector.

5. The AI big model capability-based real-time risk early warning system for in-field flow machinery according to claim 4, characterized in that: In S1.2, the operation environment parameters are encoded by using the perception neural radiation field environment modeling method, including the following steps: S1.21, construct a continuous spatio-temporal query coordinate, perform high-frequency position encoding on the spatio-temporal query coordinate, and obtain a high-dimensional feature representation; S1.22, input the encoded high-dimensional feature into a multilayer perception machine, and output the environment parameter vector of the corresponding position and time through the inference of the trained neural radiation field model on the query coordinate; S1.23, input the continuously collected running state parameters and operation environment parameters as training samples into the neural radiation field model, and perform online fine-tuning on the parameters of the neural radiation field model based on a continuous learning strategy; S1.24, take the spatio-temporal query coordinates on the current position and the predicted path of the machine as query points, and obtain local environment parameter prediction values and corresponding uncertainty weights by using the neural radiation field model; S1.25, combine the local environment parameter prediction values and the uncertainty weights of each query point to form an environment feature vector. 6.The AI big model capability-based real-time risk early warning system for in-field flow machinery according to claim 3, characterized in that: The risk modeling module establishes a risk prediction model of the interaction between the running behavior of the mobile machine and the environment, and generates a risk index, including the following steps: S2.1, pre-process the high-dimensional time sequence feature vector, input the pre-processed high-dimensional time sequence feature vector into a BERT model to generate a semantic enhanced feature representation ; S2.2, input the high-dimensional time sequence feature vector into a long short-term memory network model, identify the time sequence dependence relationship between the running state change and the environment dynamics, and output a time sequence feature representation ; S2.3, the time sequence feature representation output by the long short-term memory network model and the semantic enhanced feature representation output by the BERT model are fused through a multi-modal dynamic adaptive fusion network to generate a joint feature vector ; S2.4, the joint feature vector The risk prediction model is inputted, inference is performed by using the multilayer perception, and the risk index is outputted.

7. The AI big model capability-based real-time risk early warning system for in-field flow machinery according to claim 6, characterized in that: In S2.3, the fusion is performed through a multi-modal dynamic adaptive fusion network, including the following steps: S2.31, based on the timing feature representation, the semantic enhanced feature representation and the context information of the current environment , construct the feature splicing vector , input the feature splicing vector into the gating function to calculate the gating vector ; S2.32, the timing feature representation and the semantic enhanced feature representation are weighted and fused by using a gating vector to generate a final joint feature vector .

8. An AI large model capability-based in-field mobile mechanical production scheduling method for the AI large model capability-based in-field mobile mechanical real-time risk early warning system according to any one of claims 1-7. including the following steps: S3.1, collect the running state parameters and operation environment parameters of the mobile machine in the field, and pre-process the running state parameters and operation environment parameters; S3.2, extract the feature vectors of the running state parameters and operation environment parameters, model the interaction relationship between the mobile machine running trajectory, behavior mode and environment by using a BERT model and a long short-term memory network model, generate a risk index, and enhance the environment feature vector of the operation environment parameters by using a perception neural radiation field environment modeling method during the extraction of the operation environment parameter feature vector; S3.3, generate an alarm information based on the risk index; S3.4, integrate the running state parameters and operation environment parameters and the risk index, model the operation ability of the machine and the resource constraint conditions, and generate a feasible resource allocation scheme.

Images