Logistics dispatching system and dispatching method based on forklift inspection data collection within the factory

Abstract

The present invention discloses a logistics scheduling system and a scheduling method based on the inspection data collection of in-plant logistics forklifts. In the system, a camera and an AI vision host are deployed on each in-plant logistics forklift, a data acquisition module and an abnormal event recognition model are deployed on the AI ​​vision host, and a scheduling strategy reasoning model is deployed on the scheduling server; the camera monitors the operating environment of the in-plant logistics forklift on which it is deployed to obtain an environmental monitoring video stream, the data acquisition module samples the environmental monitoring video stream to obtain video sampling data, the abnormal event recognition model extracts inspection key features from the video sampling data, and identifies in-field logistics abnormal events based on the extracted inspection key features; the scheduling strategy reasoning model processes the in-field logistics abnormal events identified by the abnormal event recognition model on all in-plant logistics forklifts to generate a scheduling strategy for all in-plant logistics forklifts.

Description

Technical Field

[0001] The present invention relates to the technical field of inspection and dispatching, and in particular to a logistics dispatching system and a dispatching method based on the collection of in-plant logistics forklift inspection data. Background Art

[0002] In the field of industrial automation and intelligent logistics, forklifts are key equipment for in-plant logistics, and their efficiency and safety are crucial to the entire production process. With the development of technology, more and more automation and intelligent technologies are being applied to the logistics field to improve operational efficiency and safety. For the inspection and monitoring of in-plant conditions, existing technologies mainly rely on fixed surveillance cameras and sensor networks to collect data, and perform logistics scheduling through centralized control systems. These systems usually have certain limitations, such as blind spots in monitoring, poor real-time data, and untimely response to emergencies. For example, fixed cameras may not cover all areas where forklifts operate, resulting in blind spots in monitoring; centralized control systems may encounter delays when processing large amounts of data, affecting scheduling efficiency.

[0003] Specifically, the existing technologies for monitoring and dispatching forklifts in factory logistics still have the following defects and deficiencies: (1) Limitations of data collection: In the existing technologies, data collection mainly relies on fixed monitoring equipment, which limits the monitoring scope and makes it difficult to achieve comprehensive coverage, especially in the dynamically changing factory environment. This limitation makes it impossible to monitor all situations around the forklift in real time, such as emergency events such as people falling and open flames. (2) Poor real-time performance: The centralized logistics dispatching system has delays in processing and responding to emergencies, because the data needs to be transmitted from the site to the central server, and then the instructions are issued after processing. This process may lead to longer response time and affect the efficiency of emergency handling. Summary of the invention

[0004] The present invention provides a logistics dispatching system based on the inspection data collection of in-plant logistics forklifts, which comprises: a plurality of in-plant logistics forklifts and a dispatching server, each of which is equipped with a camera and an AI vision host, the AI ​​vision host is equipped with a data collection module and an abnormal event recognition model, and the dispatching server is equipped with a dispatching strategy reasoning model;

[0005] The camera is used to monitor the operating environment of the in-plant logistics forklift where it is deployed to obtain an environmental monitoring video stream, the data acquisition module is used to sample the environmental monitoring video stream to obtain video sampling data, and the abnormal event recognition model is used to extract inspection key features from the video sampling data, and identify in-plant logistics abnormal events based on the extracted inspection key features;

[0006] The scheduling strategy inference model is used to perform the following steps on the abnormal events of in-plant logistics identified by the abnormal event recognition model on all in-plant logistics forklifts to generate a scheduling strategy for all in-plant logistics forklifts:

[0007] Perform cluster analysis on all abnormal logistics events within the site to generate an abnormal classification tree;

[0008] Obtaining a description of in-plant logistics resources and matching it with the abnormal classification tree to generate an impact assessment value and an impact range assessment value of all in-plant logistics abnormal events on the overall operation of in-plant logistics;

[0009] Reasoning the forklift task adjustment target and the forklift path planning target according to the impact degree assessment value and the impact range assessment value;

[0010] Based on the forklift task adjustment target and forklift path planning target obtained by reasoning, a scheduling strategy for all in-plant logistics forklifts is generated.

[0011] The present application also provides a scheduling method for a logistics scheduling system based on in-plant logistics forklift inspection data collection, which includes:

[0012] The camera monitors the operating environment of the logistics forklift in the factory where it is deployed to obtain an environmental monitoring video stream;

[0013] The data acquisition module samples the environment monitoring video stream to obtain video sampling data;

[0014] The abnormal event recognition model extracts inspection key features from the video sampling data, and identifies abnormal logistics events within the field based on the extracted inspection key features;

[0015] The scheduling strategy inference model performs the following steps on the abnormal events in the field identified by the abnormal event recognition model on all the in-plant logistics forklifts to generate a scheduling strategy for all the in-plant logistics forklifts:

[0016] Perform cluster analysis on all abnormal logistics events within the site to generate an abnormal classification tree;

[0017] Obtaining a description of in-plant logistics resources and matching it with the abnormal classification tree to generate an impact assessment value and an impact range assessment value of all in-plant logistics abnormal events on the overall operation of in-plant logistics;

[0018] Reasoning the forklift task adjustment target and the forklift path planning target according to the impact degree assessment value and the impact range assessment value;

[0019] Based on the forklift task adjustment target and forklift path planning target obtained by reasoning, a scheduling strategy for all in-plant logistics forklifts is generated. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic diagram of a logistics scheduling system based on the collection of in-plant logistics forklift inspection data according to an embodiment of the present invention.

[0021] Figure 2 It is a schematic diagram of a logistics scheduling method based on the collection of in-plant logistics forklift inspection data according to an embodiment of the present invention. DETAILED DESCRIPTION

[0022] Figure 1 Schematic diagram of a logistics dispatching system based on in-plant logistics forklift inspection data collection according to an embodiment of the present invention. Figure 1 The logistics scheduling system based on the inspection data collection of in-plant logistics forklifts shown in the figure includes: multiple in-plant logistics forklifts 1, a scheduling server 2, each in-plant logistics forklift 1 is deployed with a camera 3 and an AI visual host 4, the AI ​​visual host 4 is deployed with a data acquisition module and an abnormal event recognition model 5, and the scheduling server 2 is deployed with a scheduling strategy reasoning model 6; the camera 3 is used to monitor the operating environment of the in-plant logistics forklift 1 deployed with it to obtain an environmental monitoring video stream, the data acquisition module is used to sample the environmental monitoring video stream to obtain video sampling data, and the abnormal event recognition model is used to extract inspection key features of the video sampling data, and identify in-field logistics abnormalities based on the extracted inspection key features. Event; the scheduling strategy reasoning model 6 is used to execute the following steps for the in-plant logistics abnormal events identified by the abnormal event recognition model on all in-plant logistics forklifts 1 to generate a scheduling strategy for all in-plant logistics forklifts 1: cluster analysis is performed on all in-plant logistics abnormal events to generate an abnormal classification tree; the in-plant logistics resource description is obtained and matched with the abnormal classification tree to generate an impact degree assessment value and an impact range assessment value of all in-plant logistics abnormal events on the overall operation of the in-plant logistics; according to the impact degree assessment value and the impact range assessment value, the forklift task adjustment target and the forklift path planning target are inferred; based on the forklift task adjustment target and the forklift path planning target obtained by reasoning, a scheduling strategy for all in-plant logistics forklifts 1 is generated.

[0023] Optionally, the scheduling strategy reasoning model 6 includes an abnormal event input network and an event feature extraction network, and the event feature extraction network includes a convolution layer, a pooling layer, and a fully connected layer; the clustering analysis of all in-field logistics abnormal events to generate an abnormal classification tree includes: based on the abnormal event input network, all in-field logistics abnormal events are vectorized to generate a feature sequence representation of in-field logistics abnormal events; based on the convolution layer, local feature extraction is performed on the feature sequence representation of in-field logistics abnormal events to generate an abnormal characterization local feature vector, and the abnormal characterization local feature vector includes abnormal characterization edge features, abnormal characterization corner features, abnormal characterization texture features, abnormal characterization color features, and abnormal characterization motion features; based on the pooling layer, the abnormal characterization local feature vector is downsampled to generate an abnormal characterization key feature vector; based on the fully connected layer, the abnormal characterization key feature vector is feature mapped to obtain a feature representation vector of abnormal events to generate the abnormal classification tree. Specifically, there is Abnormal logistics events within the field, each abnormal event Contains multiple attributes, such as the time when the event occurred 、 Location 、Forklift No. , Exception type description . All the abnormal logistics events in the field are vectorized to generate the characteristic sequence representation of the abnormal logistics events in the field, including the following process: , convert it into the number of seconds from a fixed time point (such as midnight of the day) to get the value For location attributes The warehouse is divided into Regions are encoded using a one-hot encoding method. Mapped to a Dimensional vector , where only the elements corresponding to the region number positions are , and the remaining elements are For forklift number , if there are Forklifts, also using one-hot encoding Dimensional vector For exception type description (in text form), use a word vector model (such as Word2Vec) to convert it into a fixed-length Vector . All attribute vectors of each abnormal event are concatenated to obtain a vector with dimension Vector , that is, for each abnormal event All The vectors obtained after vectorization of abnormal events constitute the characteristic sequence representation of abnormal events in the field logistics. ,for The matrix of .

[0024] To this end, the technical benefits of vectorizing the abnormal logistics events within the field based on the scheduling strategy reasoning model 6 to generate feature sequence representation are as follows:

[0025] 1. Data standardization and unified format. Easy to process and analyze: Convert different types of attributes (time, location, forklift number, abnormal type description) into numerical vectors, so that various information is unified in mathematical form. In computer processing, a unified data format makes it easier to perform various operations, such as storage, calculation, and model input. Different algorithms and models usually require input data to have a specific format. Vectorization operations ensure that all abnormal event data can be processed in a consistent manner, greatly improving the efficiency and accuracy of data processing. Eliminate data heterogeneity: The original abnormal event attributes have different types and magnitudes. For example, time is represented by a timestamp, location is an area number, forklift number is a discrete identifier, and abnormal type description is text. These heterogeneous data are difficult to compare and analyze directly. Through vectorization, all attributes are converted into numerical vectors, eliminating the differences in data types, so that different attributes can be operated and analyzed in the same vector space, laying the foundation for subsequent machine learning and data analysis tasks.

[0026] 2. Retain key information. Time information quantification: The time of an event is converted into the number of seconds from a fixed time point, retaining the order and time interval information of the events. This quantification method is very useful in analyzing the time series characteristics of abnormal events. For example, potential patterns can be discovered by analyzing the time intervals of different abnormal events, or the cumulative impact of events on the logistics system can be evaluated based on the time sequence of events. Location and forklift number encoding: The location and forklift numbers are processed using one-hot encoding, which can accurately represent the uniqueness of each location area and forklift. In the one-hot encoding vector, only the elements at the corresponding position are 1, and the rest are 0. This encoding method not only retains the identification information of the location and forklift, but also facilitates distinction and processing in vector operations. For example, when analyzing whether a certain area or a certain forklift frequently has abnormal events, the one-hot encoding vector can be directly used for related calculations and judgments. Abnormal type text vectorization: Using a word vector model (such as Word2Vec) to convert the abnormal type description into a fixed-length vector can capture the semantic information in the text. Text data is usually highly complex and diverse, and it is difficult to directly perform mathematical processing. The word vector model maps each word into a low-dimensional vector space by learning from a large amount of text data, so that words with similar semantics are closer in the vector space. Therefore, after converting the abnormal type description into a word vector, not only the semantic information of the text is retained, but also the similarity between different abnormal types can be measured by calculating the distance between vectors, which is helpful for the classification and cluster analysis of abnormal events.

[0027] 3. Support machine learning and data analysis. Model input adaptation: The generated feature sequence representation of abnormal logistics events in the field is a , which is very suitable as the input of machine learning models. Many common machine learning algorithms, such as neural networks, support vector machines, decision trees, etc., can directly process this structured numerical data. By inputting the vectorized abnormal event data into these models, tasks such as classification, prediction, and clustering of abnormal events can be performed, such as predicting the type of abnormal events that may occur in the future, or clustering analysis of similar abnormal events to better formulate response strategies. Feature extraction and dimensionality reduction: In the subsequent machine learning and data analysis process, based on the vectorized feature sequence representation, feature extraction and dimensionality reduction operations can be further performed. For example, principal component analysis (PCA), singular value decomposition (SVD) and other methods can be used to reduce the dimension of the feature matrix, remove redundant information, and extract the most critical features. These reduced-dimensional features can not only reduce the dimension of the data and improve computational efficiency, but also avoid overfitting problems and improve the generalization ability of the model. At the same time, through the analysis of the features, the degree of influence of different attributes on abnormal events can be found, providing valuable information for in-depth understanding of the operation status of the logistics system.

[0028] 4. System integration and scalability. Easy to integrate into existing systems: The vectorized abnormal event data exists in a unified vector form, which is easy to integrate with other logistics system modules. Whether it is integrated with the logistics scheduling system, monitoring system or management decision-making system, it can interact through standard data interfaces and formats. This integration enables the entire logistics system to work together more efficiently, such as adjusting the logistics scheduling strategy in real time according to the feature sequence representation of abnormal events, or providing managers with more accurate decision support information. Strong scalability: The solution has good scalability, and it is easy to adjust and expand the vectorization process when the scale or demand of the logistics system changes. For example, if the number of warehouse areas increases or the number of forklifts changes, just adjust the dimension of the unique hot encoding accordingly; if new abnormal event attributes need to be added, they can also be converted into vectors and added to the existing feature sequence representation. This scalability ensures that the solution can maintain effectiveness and adaptability in different logistics scenarios, and can be continuously optimized and upgraded as the logistics system develops.

[0029] To realize the above-mentioned processing vectorization process, the abnormal event input network includes a data receiving module, a data parser, a time attribute processing module, a location attribute encoding module, a forklift number encoding module, an abnormal type description processing module, and a vector splicing module, wherein:

[0030] The data receiving module is used to receive abnormal logistics events within the field; the data parser parses the received abnormal logistics events within the field and converts them into structured data. The exemplary processing code is as follows:

[0031] importjson

[0032] data='{"time":"2023100110:00:00","location":3,"forklift_id":5,"description":"Forkliftcollision"}'

[0033] parsed_data = json.loads(data)

[0034] The time attribute processing module converts the event occurrence time into the number of seconds from a fixed time point (such as midnight of the day). The exemplary processing code is as follows:

[0035] fromdatetimeimportdatetime

[0036] defconvert_time_to_seconds(time_str):

[0037] dt=datetime.strptime(time_str,'%Y%m%d%H:%M:%S')

[0038] zero_time=datetime(dt.year,dt.month,dt.day)

[0039] returnint((dtzero_time).total_seconds())

[0040] time_str="2023100110:00:00"

[0041] seconds=convert_time_to_seconds(time_str)

[0042] The location attribute encoding module maps the warehouse area number to a vector of a set dimension. The warehouse is divided into m areas, and the area number is l. The exemplary code is as follows:

[0043] importnumpyasnp

[0044] defone_hot_encode_location(location,num_regions):

[0045] one_hot = np.zeros(num_regions)

[0046] one_hot[location]=1

[0047] returnone_hot

[0048] num_regions=10

[0049] location=3

[0050] one_hot_location=one_hot_encode_location(location,num_regions)

[0051] The forklift number encoding module maps the forklift number to a vector of a specified dimension. There are k forklifts in total, and the forklift number is f. The exemplary code is as follows:

[0052] defone_hot_encode_forklift(forklift_id,num_forklifts):

[0053] one_hot=np.zeros(num_forklifts)

[0054] one_hot[forklift_id]=1

[0055] returnone_hot

[0056] num_forklifts=20

[0057] forklift_id=5

[0058] one_hot_forklift=one_hot_encode_forklift(forklift_id,num_forklifts)

[0059] The exception type description processing module uses a word vector model (such as Word2Vec) to convert the text-based exception type description into a vector of fixed length p. The sample code is as follows:

[0060] fromgensim.modelsimportWord2Vec

[0061] importnumpyasnp

[0062] # A text list containing descriptions of the exception types

[0063] texts=[["forklift","collision"],["cargo","drop"]]

[0064] model=Word2Vec(texts,min_count=1)

[0065] defget_description_vector(description,model,vector_size):

[0066] words = description.split()

[0067] vectors=[model.wv[word]forwordinwordsifwordinmodel.wv]

[0068] ifnotvectors:

[0069] returnnp.zeros(vector_size)

[0070] return np.mean(vectors,axis=0)

[0071] description="forkliftcollision"

[0072] vector_size=100

[0073] description_vector=get_description_vector(description,model,vector_size)

[0074] The vector concatenation module concatenates the attribute vectors processed by the above modules to obtain a vectorized representation of each abnormal event, and obtains the time attribute vector t', the location attribute vector l, the forklift number attribute vector f and the abnormal type description vector d by way of example. The implementation code is as follows:

[0075] defconcatenate_vectors(time_vector,location_vector,forklift_vector,description_vector):

[0076] returnnp.concatenate((time_vector,location_vector,forklift_vector,description_vector))

[0077] time_vector=np.array([seconds])

[0078] concatenated_vector=concatenate_vectors(time_vector,one_hot_location,one_hot_forklift,description_vector)

[0079] For the feature sequence representation generation module, the vectorized vectors of all n abnormal events are combined into a matrix, that is, the feature sequence representation of the abnormal events in the field logistics. The n vectorized vectors are recorded as :

[0080] defgenerate_feature_sequence(vectors):

[0081] returnnp.vstack(vectors)

[0082] vectors=[concatenated_vector]*n

[0083] feature_sequence=generate_feature_sequence(vectors)

[0084] Specifically, the convolution layer consists of multiple convolution kernels. convolution kernels, the size of each convolution kernel is ( represents the window size of the convolution kernel in the vector dimension). Each convolution kernel has a set of weight parameters ( ) and a bias term .

[0085] The local feature extraction process is as follows:

[0086] The characteristic sequence of abnormal logistics events in the field is expressed as As the input of the convolutional layer.

[0087] For each convolution kernel , in the input matrix According to the step length Slide. At each slide, the convolution kernel is element-wise multiplied with the corresponding area of ​​the input matrix and the sum is added, plus the bias term , and get a new value.

[0088] First Taking a convolution kernel as an example, when sliding at a certain position, the input matrix area covered by the convolution kernel is ( represents the position index), then the convolution output at that position for: The convolution kernel is applied to the entire input matrix After sliding up, a new feature map will be generated .all The feature maps generated by the convolution kernels are combined together to form a local feature vector representing the abnormality. The size of the feature map is (here and The local feature vector of the abnormal representation is determined by the input matrix size, convolution kernel size and step size. is a Specifically, when downsampling based on the pooling layer, the local feature vector representing the abnormality is Perform downsampling operation. In each feature map On the pooling window size Slide. Select the maximum value in each pooling window as the output of the window. Preferably, for the feature map The pooling window at a certain position in is ( and represents the position index within the window), then the pooled output of the window is for: .in and Represents the position index of the pooling window on the feature map. After the pooling operation, the height and width of each feature map become the original Assume that after pooling, the new height of the feature map is h' and the new width is w', then the key feature vector of abnormal characterization is obtained is a Tensor of .

[0089] To this end, the technical benefits of the above convolutional layer and pooling layer operations in processing the characteristics of abnormal logistics events within the field are as follows:

[0090] 1. Local feature extraction. Focus on detail information: The convolution layer slides on the input matrix through the convolution kernel, focusing only on the information of the local area each time. This local perception feature enables the convolution layer to capture the detailed information in the feature sequence of abnormal events, such as a specific time segment, location area, or a combination of local features related to the abnormal type. For abnormal events in on-site logistics, some specific local feature combinations may accurately identify the type or severity of the abnormal event, and the convolution layer can effectively extract this key information. Parameter sharing and reduced computation: Each convolution kernel has a fixed set of weight parameters, which are shared when sliding on the entire input matrix. This means that no matter where the convolution kernel is in the input matrix, it extracts local features in the same way. This parameter sharing mechanism greatly reduces the number of parameters of the model and reduces the computational complexity. Compared with the fully connected layer, the convolution layer is more efficient in processing large-scale data. It can reduce training time and memory consumption while ensuring the effect of feature extraction, which is particularly important for processing a large amount of on-site logistics abnormal event data.

[0091] 2. Feature combination and abstraction. Generate feature maps: Each convolution kernel generates a new feature map after sliding on the input matrix. Different convolution kernels can capture different types of local features. The feature maps generated by all convolution kernels are combined together to form a local feature vector that represents the abnormality, which contains rich feature information. This feature information is an abstraction and combination of the original input data, which can more comprehensively reflect the characteristics of abnormal events. For example, one convolution kernel may be sensitive to the changing pattern of the time when the abnormal event occurs, and another convolution kernel may be sensitive to the specific regional characteristics of the location where the abnormal event occurs. Through the combination of feature maps, these different features can be integrated together to provide more valuable information for subsequent analysis. Mining potential patterns: The convolution layer can mine potential patterns and regularities in the data by combining and abstracting local features. These patterns may not be obvious in the original data, but they can be effectively extracted through the processing of the convolution layer. For abnormal events in on-site logistics, the convolution layer may find specific association patterns of certain abnormal events in time and location, or the potential connection between certain abnormal types and specific forklift operations. This information is of great significance for in-depth understanding of the occurrence mechanism of abnormal events and formulating corresponding prevention strategies.

[0092] 3. Downsampling and data dimension reduction. Reducing the amount of data: The pooling layer downsamples the feature map by sliding on the feature map with a certain window size and selecting the maximum value in the window as the output. After the pooling operation, the height and width of each feature map become 1 / u of the original, which greatly reduces the dimension and quantity of the data. When processing large-scale data of abnormal logistics events in the field, reducing the amount of data can significantly reduce the complexity of subsequent calculations and improve the operating efficiency of the model. At the same time, reducing the data dimension also helps to reduce the risk of overfitting, because the number of parameters that the model needs to learn is reduced, making it easier to generalize to new data. Retaining key features: Although the pooling layer reduces the amount of data, it retains the key information in the feature map by selecting the maximum value. In each pooling window, the maximum value usually represents the most significant feature in the area. By retaining these key features, the pooling layer will not lose too much important information while reducing the data dimension. For the feature representation of abnormal logistics events in the field, this means that even after the pooling operation, the key features related to the abnormal event can still be retained, thereby ensuring the accuracy of the subsequent analysis and processing of the abnormal event.

[0093] 4. Improve the robustness of features. Anti-noise and small changes: In the process of downsampling the feature map, the pooling layer can smooth out noise and small changes to a certain extent. Since the pooling operation selects the maximum value within a window, some local noise or small changes may not affect the final output. This makes the feature representation robust to small changes in the input data, and can accurately extract features related to abnormal events even when there is some noise or inaccuracy in the data. For example, there may be some small interference caused by illumination changes or sensor noise in the video sampling data. The pooling layer can effectively reduce the impact of these interferences on feature extraction and improve the accuracy of abnormal event recognition. Enhance the representativeness of features: Through the pooling operation, the features in the feature map are further compressed and integrated, making each feature more representative. While reducing the data dimension, the pooling layer can highlight the most important features, making these features easier to identify and use in subsequent processing. For the classification and analysis of abnormal events in on-site logistics, this enhanced feature representativeness helps to improve the classification accuracy and analysis efficiency of the model, and can more accurately judge the type and severity of abnormal events. Specifically, the input of the fully connected layer is the key feature vector of abnormal characterization after processing by the pooling layer , the output is the feature representation vector of the abnormal event. The fully connected layer contains multiple neurons, each neuron is connected to all elements of the input vector, and each connection has a weight parameter. The fully connected layer has Neuron, The weight vector of a neuron is , the bias term is , the mathematical processing process of realizing feature mapping is as follows:

[0094] First, the key feature vector of the abnormal characterization Flatten it and convert it into a one-dimensional vector . is a Tensor, then the flattened vector The length is For the fully connected layer neurons, whose output for: . After activation function (such as ReLU function: ) is processed to obtain the final feature representation vector of the abnormal event This feature representation vector contains the comprehensive feature information of abnormal events extracted after multi-layer processing, providing more representative data for subsequent clustering analysis.

[0095] To this end, the fully connected layer has the following technical benefits in the above abnormal event feature processing flow:

[0096] 1. Feature fusion and global information integration. Integrate all features: Each neuron in the fully connected layer is connected to all elements of the key feature vector of abnormal characterization after processing by the pooling layer, which means that it can comprehensively integrate various local and key features extracted by the previous convolutional layer and pooling layer. In the previous convolution and pooling operations, the data was processed into a tensor form with a specific structure, retaining feature information at different levels and aspects. The fully connected layer integrates these scattered features together in this fully connected way, so that the final output can comprehensively reflect the various features of the abnormal event, thereby describing the abnormal event more comprehensively. Capture global feature relationships: By weighted summation of all features, the fully connected layer can capture the complex relationship between different features. Unlike the convolutional layer and the pooling layer, which mainly focus on local features, the fully connected layer can analyze how these features interact to affect the nature of the abnormal event from a global perspective. For example, it can discover the potential connection between different attribute features such as time, location, forklift number, and abnormal type, thereby digging out deeper information, which is very important for accurately classifying abnormal events and understanding their nature.

[0097] 2. Flexible feature mapping. Nonlinear transformation capability: After the output of the fully connected layer is processed by an activation function (such as the ReLU function), nonlinear transformation can be introduced. This allows the fully connected layer to learn the complex nonlinear relationship between the input features and the output, greatly enhancing the model's expressive power. Without an activation function, the fully connected layer is just a linear transformation and can only learn the linear combination of the input features, making it difficult to accurately model complex abnormal event feature relationships. Through the activation function, the fully connected layer can perform nonlinear mapping on the results of the linear transformation, thereby better fitting the complex patterns in the data and improving the ability to represent abnormal event features. Adjust feature weights: Each connection has a weight parameter, which is constantly adjusted during the training process. This allows the fully connected layer to automatically learn the importance of each feature to the final output based on the characteristics of the data, and then assign different weights to different features. For features that are closely related to abnormal events, the model will learn larger weights, while for irrelevant or interfering features, the weights will be smaller. This ability to automatically adjust feature weights enables the fully connected layer to more effectively extract and utilize key feature information, improving the ability to identify and analyze abnormal events.

[0098] 3. Provide effective data for subsequent tasks. Suitable for cluster analysis: The feature representation vector of abnormal events output by the fully connected layer contains comprehensive feature information extracted after multi-layer processing, providing more representative data for subsequent cluster analysis. The purpose of cluster analysis is to classify similar abnormal events into one category, and this feature representation vector can accurately reflect the essential characteristics of abnormal events, so that the similarity between different abnormal events can be more accurately measured in the clustering process, thereby improving the accuracy and effectiveness of clustering. For example, when clustering a large number of abnormal logistics events in the field, the feature vector based on the output of the fully connected layer can cluster abnormal events with similar characteristics together, helping analysts to discover different types and patterns of abnormal events. Support multiple machine learning tasks: The feature representation vector output by the fully connected layer is not only suitable for cluster analysis, but also can be used as input for other machine learning tasks, such as classification and prediction. Because the vector integrates multiple features of abnormal events, it can provide rich information for these tasks, so that the model can make more accurate decisions when performing tasks such as abnormal event classification and prediction of the probability of abnormal event occurrence. This provides strong support for building a complete on-site logistics abnormal event analysis system, enabling the system to perform diversified data analysis and processing according to different needs.

[0099] Optionally, the scheduling strategy reasoning model 6 also includes an inter-feature attention analysis network and a clustering network; the clustering analysis of all in-field logistics abnormal events to generate an abnormal classification tree also includes: based on the inter-feature attention analysis network, a multi-head cross-attention analysis is performed on the feature representation vectors of different abnormal events to generate a comprehensive feature vector of abnormal events; based on the clustering network, a cluster analysis is performed on the comprehensive feature vector of abnormal events to generate the abnormal classification tree. Specifically, the inter-feature attention analysis network is mainly composed of a multi-head attention mechanism module. The multi-head attention mechanism calculates attention in parallel through multiple different "heads", which can capture the relationship between features from different subspaces. Each "head" contains a linear transformation layer, an attention calculation layer, and an output merging layer. Specifically, the multi-head cross-attention analysis processing process is as follows: the set of feature representation vectors of different input abnormal events is ,in It is The feature representation vector of an abnormal event has a dimension of .

[0100] 1. Linear transformation: For each "head", the input feature representation vector is first linearly transformed. The first "head" "heads", through the weight matrix , and Perform linear transformation on the input vectors to obtain the query vectors , key vector Sum value vector For the The feature representation vector of an abnormal event , whose corresponding query vector , key vector Sum value vector The calculation of is as follows: , , ,in, , and The dimension is , and The weight matrix ( ).

[0101] 2. Attention calculation: Calculate the attention weight of each "head". The attention weight represents the degree of correlation between the feature representation vector of an abnormal event and the feature representation vectors of all other abnormal events. The first The attention weight between an abnormal event and all other abnormal events The calculation is as follows: .in, The function normalizes the input vector so that the sum of its elements is , represents the probability distribution.

[0102] 3. Weighted summation: According to the attention weight, the value vector is weighted summed to obtain the output of each "head" : .

[0103] 4. Output merging: The outputs of the two “heads” are concatenated, and then a linear transformation is performed to obtain the final abnormal event comprehensive feature vector First, The output of the "heads" is concatenated into a matrix by column Then through the weight matrix Perform a linear transformation: ,in, The dimension is Specifically, the clustering network adopts a hierarchical clustering model, which includes a distance calculation module and a cluster merging module. To this end, the technical processing process for implementing clustering is as follows:

[0104] 1. Distance calculation: Calculate the distance between the comprehensive feature vectors of abnormal events. Common distance measurement methods include Euclidean distance and cosine distance. Taking Euclidean distance as an example, for two comprehensive feature vectors of abnormal events and , the Euclidean distance between them The calculation is as follows: .in, and They are vectors and No. elements.

[0105] 2. Cluster merging: Initially, each abnormal event comprehensive feature vector is regarded as a separate cluster. Then, based on the distance calculation results, the two clusters with the closest distance are merged into a new cluster.

[0106] Repeat the above merging process until all clusters are merged into a large cluster or the set termination condition is met (such as the number of clusters reaches a certain threshold). During the merging process, the information of each merge is recorded to construct an anomaly classification tree. The nodes of the anomaly classification tree represent clusters, and the leaf nodes represent single anomalies. The hierarchical structure of the tree reflects the similarity and clustering relationship between anomalies.

[0107] To this end, the above multi-head cross-attention analysis process has the following technical benefits:

[0108] 1. Capture multi-dimensional feature relationships. Enhance feature interaction: Multi-head cross-attention analysis calculates attention in parallel through multiple "heads", and each "head" uses a different weight matrix to perform a linear transformation on the input feature representation vector, so that the relationship between features can be captured from different subspaces. This means that it can focus on multiple aspects of abnormal event features at the same time and explore the complex interactions between different features. For example, when analyzing abnormal logistics events in the field, different "heads" can focus on the relationship between the time and location of the event, the association between the abnormal type and the forklift number, etc., so that the model can understand the feature structure of the abnormal event more comprehensively. Improve feature representation capabilities: The attention weights and outputs calculated by each "head" reflect different perspectives of the input features. The outputs of these "heads" are spliced ​​and linearly transformed to obtain the final comprehensive feature vector of the abnormal event, which can fuse information from multiple dimensions to generate richer and more representative feature representations. This multi-dimensional feature fusion can improve the model's ability to describe abnormal events, enable it to better distinguish different types of abnormal events, and provide stronger support for subsequent analysis and decision-making.

[0109] 2. Adaptively focus on important information. Dynamically allocate attention: The attention weight represents the degree of association between the feature representation vector of an abnormal event and the feature representation vectors of all other abnormal events. After normalization by the Softmax function, these weights can be regarded as probability distributions. This means that the model can adaptively allocate different attention to different abnormal events and features according to the characteristics of the input data. For abnormal events and features that are more relevant to the current analysis task, the model will give higher attention weights, so as to more effectively extract and utilize these important information. For example, when dealing with complex on-site logistics abnormal event scenarios, the model can automatically focus on those abnormal events that have a greater impact on the overall logistics operation and ignore some minor interference factors. Improve model robustness: The adaptive characteristics of the multi-head cross-attention mechanism make the model more robust to data changes and noise. Since the model can dynamically adjust the attention allocation according to the input, it can maintain good performance by focusing on relatively important information even when there is some noise in the data or the features are incomplete. This helps to improve the stability and reliability of the model in practical applications and reduce analysis errors caused by data quality issues.

[0110] The technical processing of the above hierarchical clustering model has the following technical benefits:

[0111] 1. Reveal the hierarchy and similarity of abnormal events. Intuitively display the relationship: The abnormal classification tree constructed by the hierarchical clustering model shows the similarity and clustering relationship between abnormal events in an intuitive tree structure. The leaf node represents a single abnormal event, the node represents a cluster, and the hierarchical structure of the tree reflects the classification process of abnormal events from fine-grained to coarse-grained. This visual representation helps analysts quickly understand the inherent connection between different abnormal events and discover the potential patterns and classification rules of abnormal events. For example, by observing the abnormal classification tree, analysts can clearly see which abnormal events are most similar in features and which abnormal events can be classified into the same category, thereby providing guidance for further analysis and processing. Mining potential classification patterns: In the clustering process, the model continuously merges clusters according to the distance between the comprehensive feature vectors of abnormal events. This process can automatically mine potential classification patterns in the data. Even without pre-defined classification criteria, the hierarchical clustering model can classify abnormal events into different categories based on their feature similarities, which is of great significance for discovering new types of abnormal events or classifying unknown abnormal situations. For example, when faced with some new types of abnormal logistics events within the field, the hierarchical clustering model can perform cluster analysis on them together with existing abnormal events to help analysts determine the category to which they belong or discover their unique characteristics.

[0112] 2. Flexibility and adaptability. No need to pre-set the number of categories: Unlike some clustering algorithms that require the number of categories to be specified in advance, the hierarchical clustering model does not need to determine the final number of clusters in advance during the clustering process. It starts with each abnormal event as a separate cluster and gradually merges them until the set termination condition is met. This flexibility enables the model to adapt to different data distributions and the complexity of abnormal events, and automatically determine the appropriate number of clusters based on the characteristics of the data itself. For example, when dealing with data sets of abnormal events in on-site logistics of different sizes and complexities, the hierarchical clustering model can effectively perform cluster analysis without manual intervention to adjust the number of categories. Applicable to different types of data: The distance measurement method used by the hierarchical clustering model (such as Euclidean distance, cosine distance, etc.) can be applied to various types of feature vector data, whether continuous or discrete. This makes it well adaptable to different forms of comprehensive feature vectors of abnormal events and can handle abnormal event data from various sources and formats. For example, whether it is the representation of abnormal events based on numerical features or the representation of encoded text or image features, the hierarchical clustering model can perform clustering analysis through appropriate distance measurement methods, thereby expanding the scope of application of the model.

[0113] Optionally, the scheduling strategy reasoning model 6 also includes a strategy generation and evaluation network, which includes: a resource description input layer, an abnormal classification tree encoding layer, and a matching and evaluation layer; the acquisition of the on-site logistics resource description and matching it with the abnormal classification tree to generate an impact degree assessment value and an impact range assessment value of all on-site logistics abnormal events on the overall operation of the factory logistics, including: based on the resource description input layer, vectorizing the on-site logistics resource description to generate an on-site logistics resource vector representation; based on the abnormal classification tree encoding layer, encoding the structure and node information of the abnormal classification tree to generate an abnormal classification code; based on the matching and evaluation layer, matching the on-site logistics resource vector representation with the abnormal classification code to obtain a matching result, and linearly transforming and activating the matching result to generate an impact degree assessment value and an impact range assessment value. Specifically, the resource description input layer is responsible for receiving the on-site logistics resource description information. This layer contains a data parsing module and a feature extraction and encoding module. The data parsing module is used to process resource description data in different formats and convert it into structured data; the feature extraction and encoding module extracts key features according to the content of the resource description and encodes them to generate a vector representation.

[0114] The description of on-site logistics resources includes multiple aspects of information, such as the number of forklifts , forklift type collection , Number of warehouse areas , Collection of goods types stored in each area and the number of available handling equipment wait.

[0115] The data parsing module processes resource description data in different formats and converts it into structured data as follows:

[0116] 1. Forklift quantity feature coding: the number of forklifts Directly used as a numerical feature, converted into a scalar ,Right now .

[0117] 2. Forklift type encoding: Use one-hot encoding to encode the forklift type. Forklift type, for a certain forklift type , whose one-hot encoded vector is a dimensional vector, where only the positional elements of the corresponding type are , and the remaining elements are Preferably, if Indicates electric forklift, and electric forklift is the first Type, then . Concatenate the one-hot encoded vectors of all forklift types into one vector in order , .

[0118] 3. Warehouse area quantity feature coding: Similar to the number of forklifts, the number of warehouse areas Convert to a scalar ,Right now .

[0119] 4. Goods type coding: The same one-hot coding is used to encode the types of goods stored in each area. Types of goods, for the type of goods , whose one-hot encoded vector is a dimensional vector, only the corresponding position elements are , and the rest are . Concatenate the unique hot encoding vectors of all cargo types into one vector in order , .

[0120] 5. Characteristic coding of the number of handling equipment: the number of available handling equipment Convert to a scalar ,Right now .

[0121] Finally, all the above feature vectors and scalars are concatenated to obtain the in-field logistics resource vector representation: : Specifically, the anomaly classification tree contains the node set , each node With attributes The feature extraction and encoding module includes an abnormal classification tree encoding layer, which includes a tree structure parsing module and a node feature encoding module. The tree structure parsing module is used to analyze the hierarchical structure, node relationship and other information of the abnormal classification tree. The node feature encoding module encodes the node attributes (such as abnormal type, occurrence frequency, etc.) to generate a vector representation. The specific process includes the following:

[0122] 1. Node attribute encoding: For each node Properties If it is a discrete attribute (such as anomaly type), it can be encoded using one-hot encoding. Types of exceptions, for nodes Exception type , whose one-hot encoded vector is a dimensional vector, only the element corresponding to the exception type position is , and the rest are .

[0123] If it is a continuous attribute (such as frequency of occurrence), normalization can be performed. The value range of frequency of occurrence is , for the node Frequency of occurrence , the normalized value is .

[0124] Concatenate the attribute encoding vectors of each node in order to obtain a vector containing the attribute encodings of all nodes : .

[0125] 2. Tree structure coding: Use the tree adjacency matrix to represent the tree structure. Let the adjacency matrix be is a The matrix of Representation Node and nodes If the node and nodes There are edges connecting them, then ,otherwise . The adjacency matrix Flatten it to get a one-dimensional vector Finally, the node attribute encoding vector and tree structure encoding vector Splice to get the abnormal classification code : Specifically, the matching and evaluation layer is used to represent the logistics resource vector in the field. and abnormal classification code Matching is performed, and the matching results are processed to generate impact degree evaluation values ​​and impact range evaluation values. This layer contains matching module, linear transformation module and activation module. The matching module calculates the matching degree between two vectors; the linear transformation module performs linear transformation on the matching results; the activation module converts the linear transformation results into the final evaluation value through the activation function. The specific processing process is as follows:

[0126] 1. Matching calculation: Dot product operation is used to calculate the vector representation of logistics resources within the field and abnormal classification code The degree of match between . Let the matching result vector be ,but: .

[0127] 2. Linear transformation: matching result vector Perform linear transformation. Suppose the weight matrix of linear transformation is , the bias vector is , then the vector after linear transformation for: .

[0128] 3. Activation operation: using activation function After linear transformation, the vector Activate to generate impact assessment values and impact range assessment Activation function Is a nonlinear function, such as the sigmoid function .

[0129] Will The first element of the activation function obtains the impact evaluation value : ;Will The second element of the activation function obtains the impact range evaluation value : Specifically, the impact assessment value is generated I and impact range assessment S The activation function used can be as follows:

[0130] If the degree and scope of impact need to be divided into multiple categories for evaluation, the Softmax function can be used. dimensional real vector into a The probability distribution vector of dimension is expressed as: . is a dimensional vector, for the impact evaluation value and impact range assessment , you can Before Elements are used to calculate the probability distribution of the impact degree, and then Elements ( ) is used to calculate the probability distribution of the impact range. , have Category, The probability of each category is: , For the impact range assessment value , have Category, The probability of each category is: , .

[0131] To this end, the entire processing process from the resource description input layer to the matching and evaluation layer has the following technical benefits:

[0132] 1. Data standardization and unification. Convenient data processing: The resource description input layer converts the in-field logistics resource description data in different formats into structured data, and further encodes it into a unified vector representation. This enables subsequent processing flows to operate on the data in a consistent manner, avoiding the complexity caused by inconsistent data formats. Whether it is storage, transmission or various calculations, a unified data format can improve efficiency and reduce the occurrence of errors. Easy model integration: In machine learning and data analysis, most models require input data to be in a specific format. After converting the resource description into a vector representation, it can be easily integrated with various models. Whether it is a model for prediction, classification or optimization, these vectors can be directly used as input, thereby enhancing the versatility and scalability of the system.

[0133] 2. Retain key features. Accurately reflect resource information: By encoding key features such as the number of forklifts, forklift types, number of warehouse areas, cargo types, and number of handling equipment, the resource description input layer can accurately retain the core information of on-site logistics resources. These features are crucial for understanding the current status of the logistics system and making subsequent analysis and decisions. For example, information on the type and number of forklifts can help the scheduling system to arrange tasks reasonably, and information on cargo types and warehouse areas can help optimize storage and transportation strategies. Support multi-dimensional analysis: Different types of features are encoded as vectors and spliced, and the resulting on-site logistics resource vector representation contains information from multiple dimensions. This allows logistics resources to be evaluated and decided from multiple perspectives in subsequent analysis. For example, the matching of cargo storage and forklift resources in different areas can be analyzed, or the impact of the number of handling equipment on overall logistics efficiency can be studied.

[0134] 3. Structured information representation. Clearly present the tree structure: The adjacency matrix is ​​used to encode the structure of the abnormal classification tree and flatten it into a one-dimensional vector, which can clearly and accurately represent the tree's hierarchy and node relationships. This representation method not only retains all the information of the tree structure, but also facilitates storage and processing in the computer. In subsequent analysis, the relevant information of the tree, such as the parent-child relationship of the node, the path length, etc., can be quickly obtained by operating on this vector. Conducive to model understanding: The structure and node attributes of the abnormal classification tree are encoded as vectors, making it easier for machine learning models to understand and process this information. The model can explore the intrinsic connections and patterns between abnormal events by learning these vectors. For example, by analyzing the node attribute encoding vector, the frequency distribution pattern of different types of abnormalities, as well as the similarities and correlations between them, can be discovered.

[0135] 4. Feature extraction and encoding. Highlight key attributes: Use one-hot encoding for discrete attributes of nodes (such as anomaly type) and normalize continuous attributes (such as frequency of occurrence), which can highlight the key attributes of each node and make these attributes numerically comparable. One-hot encoding can clearly represent the uniqueness of different anomaly types, while normalization maps continuous attributes to a unified range, making it easier for the model to learn and compare. Enhance data representation: The node attribute encoding vector and the tree structure encoding vector are concatenated, and the resulting anomaly classification encoding contains rich information and can more comprehensively describe the characteristics of the anomaly classification tree. This enhanced data representation helps to improve the model's ability to classify and analyze abnormal events, allowing the model to more accurately identify different types of abnormal events and evaluate their impact on the logistics system.

[0136] 5. Quantify the degree of matching. Measuring correlation: The degree of matching between the on-site logistics resource vector representation and the abnormal classification code is calculated through dot product operations, which can quantify the correlation between the two. The results of this quantification can serve as an important basis for evaluating the impact of abnormal events on logistics resources. For example, if the value of the matching result vector is high, it means that the current abnormal event has a strong correlation with certain characteristics of logistics resources, which may have a greater impact on the logistics system. Provide decision support: The matching results provide important reference information for subsequent decisions. By analyzing the degree of matching, decision makers can determine which abnormal events need to be focused on and which logistics resources are more vulnerable in the face of abnormal events. This helps to formulate targeted response strategies and improve the stability and reliability of the logistics system.

[0137] 6. Flexible evaluation methods. Linear transformation and nonlinear activation: Linear transformation and nonlinear activation of matching results can further process and optimize data. Linear transformation can adjust the weight and bias of data so that the model can better fit the data; nonlinear activation function introduces nonlinear characteristics, enhances the expressiveness of the model, and can more accurately capture the complex relationships in the data. For example, using the sigmoid function as the activation function, the result after linear transformation can be mapped to a probability interval, which is convenient for evaluating the degree and scope of influence. Multi-category evaluation support: When multi-category evaluation of the degree and scope of influence is required, the Softmax function provides an effective solution. It converts an n-dimensional real number vector into an n-dimensional probability distribution vector, so that the degree and scope of influence of different categories can be estimated probabilistically. This is very useful in practical applications. For example, different countermeasures can be formulated according to the probabilities of different categories to improve the accuracy and pertinence of decision-making.

[0138] Optionally, the scheduling strategy reasoning model 6 also includes a reasoning layer; the reasoning of the forklift task adjustment target and the forklift path planning target according to the impact degree evaluation value and the impact range evaluation value includes: based on the reasoning layer, the impact degree evaluation value and the impact range evaluation value are linearly transformed to generate a logistics impact reasoning latent vector, and the logistics impact reasoning latent vector is nonlinearly activated to generate a logistics scheduling decision representation driven by abnormal impact, and the logistics scheduling decision representation driven by abnormal impact includes the reasoned forklift task adjustment target and forklift path planning target. Specifically, the reasoning layer includes a linear transformation module and a nonlinear activation module. The linear transformation module is responsible for linearly combining and transforming the input impact degree evaluation value and the impact range evaluation value to generate a logistics impact reasoning latent vector. The nonlinear activation module performs a nonlinear transformation on the generated latent vector to introduce nonlinear features and generate the final abnormal impact driven logistics scheduling decision representation, which includes the forklift task adjustment target and the forklift path planning target.

[0139] To this end, the technical process of linear transformation to generate logistics impact inference latent vector is as follows: Let the impact evaluation value be The impact range is evaluated as , forming a two-dimensional vector The linear transformation module uses a weight matrix and the bias vector For the input vector Perform a linear transformation. Weight matrix is a The matrix ( represents the dimension of the latent vector of logistics impact reasoning), bias vector is a dimensional vector. Then the mathematical expression of linear transformation is: .in It is the generated logistics impact reasoning latent vector, The specific expansion calculation is: ,in is the weight matrix No. Line The elements of the column, is the input vector No. elements, is the bias vector No. elements.

[0140] Specifically, the technical process of generating abnormal impact driven logistics scheduling decision representation by nonlinear activation is as follows: the logistics impact reasoning latent vector Perform nonlinear activation and use a nonlinear activation function This activation function can be in various forms, preferably ReLU (rectified linear unit) function, Sigmoid function or Tanh function. Taking ReLU function as an example, its expression is . Reasoning about the latent vector of logistics impact Each element of is activated nonlinearly: .in It is the logistics scheduling decision representation driven by abnormal impact generated after nonlinear activation.

[0141] The vector It includes the inferred forklift task adjustment target and forklift path planning target. Specifically, the vector Different dimensions of correspond to different decision information, among which: some dimensions represent the priority adjustment information of forklift tasks. No. Dimensions Indicates the adjustment value of the forklift task priority. When it is greater than a certain threshold, it may mean that the corresponding forklift task needs to increase its priority. Other dimensions represent the parameters related to forklift path planning. For example, the vector No. Dimensions It can indicate the adjustment value of the forklift's driving speed in a certain area. The interpretation of the data can determine whether the forklift should speed up or slow down in this area.

[0142] To this end, the above reasoning layer processes the impact degree evaluation value and the impact range evaluation value through the linear transformation module and the nonlinear activation module to generate a logistics scheduling decision representation driven by abnormal impacts including the forklift task adjustment target and the forklift path planning target, which has the following technical benefits:

[0143] 1. Feature fusion and weighted combination. Comprehensive consideration of multiple factors: The impact degree assessment value and the impact range assessment value are combined into a two-dimensional vector, and linear transformation is performed through the weight matrix, which can comprehensively consider these two key factors. Different elements of the weight matrix can assign different weights to the impact degree and impact range, so as to flexibly adjust their relative importance in generating the logistics impact reasoning latent vector according to actual needs. For example, in some cases, the impact degree may have a greater impact on the logistics scheduling decision, and this factor can be highlighted by setting the corresponding elements of the weight matrix. Mining potential relationships: Linear transformation is not just a simple weighted summation, but can also mine the potential relationship between the impact degree and the impact range. By learning the parameters of the weight matrix, the model can discover how these two factors interact to affect the logistics system, thereby generating a more representative logistics impact reasoning latent vector. The mining of this potential relationship helps to more deeply understand the impact mechanism of abnormal events on the logistics system and provide a more comprehensive basis for subsequent decision-making.

[0144] 2. Data dimension conversion and adaptation. Meeting different task requirements: The dimension of the generated logistics impact reasoning latent vector is determined by the number of rows in the weight matrix, which allows the dimension of the output vector to be adjusted according to specific task requirements. For example, if subsequent analysis requires more dimensional information to describe the state of the logistics system, the number of rows in the weight matrix can be increased to generate a higher dimensional latent vector. This flexibility enables the reasoning layer to adapt to different logistics scheduling scenarios and decision-making requirements. Facilitating subsequent processing: Converting the two-dimensional input vector into a logistics impact reasoning latent vector of a specified dimension provides a unified format for subsequent processing. Whether it is further feature extraction, model training or decision making, operations can be performed based on this unified dimensional latent vector, which improves the operability and compatibility of the system.

[0145] 3. Introduce nonlinear features. Enhance the expressiveness of the model: Using nonlinear activation functions to transform the logistics impact reasoning latent vector can introduce nonlinear features. Linear transformation can only learn the linear combination relationship of input features, but in actual logistics systems, there are often complex nonlinear relationships between abnormal events and forklift task adjustment targets and path planning targets. Nonlinear activation functions can break this linear limitation, allowing the model to learn more complex functional relationships, thereby enhancing the expressiveness of the model and more accurately capturing the impact of abnormal events on logistics scheduling. Improve the adaptability of the model: Different nonlinear activation functions have different characteristics and are suitable for different scenarios. For example, the ReLU function can effectively alleviate the gradient vanishing problem and is suitable for use in deep neural networks; the Sigmoid function can map the output value to the (0, 1) interval and is often used to represent probability; the output range of the Tanh function is symmetric about the origin and performs well in some scenarios where positive and negative effects need to be considered. By selecting a suitable nonlinear activation function, the adaptability of the model to different logistics scenarios can be improved and the performance of the model can be optimized.

[0146] 4. Generate decision-related representations. Clarify decision information: The logistics scheduling decision representation vector driven by abnormal impact generated after nonlinear activation has different dimensions corresponding to different decision information, such as the priority adjustment information of forklift tasks and the relevant parameters of path planning. This clear correspondence makes it possible to directly interpret specific decision suggestions from the vector, providing clear guidance for logistics scheduling. For example, by judging whether the dimension representing the task priority adjustment value in the vector is greater than a certain threshold, it can be quickly determined whether the priority of the corresponding forklift task needs to be increased. Support decision making: Integrating the decision information obtained by reasoning into a vector facilitates the formulation of logistics scheduling decisions. Decision makers can formulate more reasonable and optimized logistics scheduling strategies based on the information of each dimension in this vector, comprehensively considering the needs of forklift task adjustment and path planning. This data-driven decision-making method can improve the efficiency and accuracy of logistics scheduling and reduce the subjectivity and blindness of human decision-making.

[0147] Optionally, the strategy generation and evaluation network also includes: a strategy generation layer, the strategy generation layer includes: a scheduling target encoding module and a strategy generation module; the forklift task adjustment target and the forklift path planning target obtained based on reasoning are used to generate a scheduling strategy for all in-plant logistics forklifts 1, including: based on the scheduling target encoding module, the forklift task adjustment target and the forklift path planning target are encoded to generate a reasoning result code; based on the strategy generation module, the reasoning result code is mapped to the constructed scheduling strategy space to generate a forklift task allocation scheduling strategy vector and a driving path scheduling strategy vector, and accordingly generate a scheduling strategy for all in-plant logistics forklifts 1. Specifically, assuming that the forklift task adjustment target includes a task type , Task Priority The forklift path planning target includes the path node sequence , Path time limit The scheduling target encoding module is responsible for converting the forklift task adjustment target and the forklift path planning target into a coding form suitable for subsequent processing. This module consists of a target parsing unit and an encoding unit. The target parsing unit is used to analyze and understand the specific content of the forklift task adjustment target and the forklift path planning target; the encoding unit performs encoding operations based on the parsed content. The specific implementation process is as follows:

[0148] 1. Task type encoding: Use one-hot encoding to encode the task type. Task types, for task types , whose one-hot encoded vector is a dimensional vector, only the position element corresponding to the task type is , and the remaining elements are . Concatenate the one-hot encoded vectors of all task types into one vector in order ; .

[0149] 2. Task priority encoding: Encode the task priority Normalize the task priority. , for task priority , the normalized value is . The normalized values ​​of all task priorities form a vector , .

[0150] 3. Path node sequence encoding: Use integer encoding to encode the path nodes. Use integers Indicates that the encoding of all path nodes is composed into a vector , .

[0151] 4. Path time limit encoding: Normalize the path time limit L. The value range of the time limit is , the normalized value is Finally, all the above encoding vectors are concatenated to obtain the inference result encoding : .

[0152] The strategy generation module is responsible for mapping the reasoning result encoding to the constructed scheduling strategy space, generating specific forklift task allocation scheduling strategy vectors and driving path scheduling strategy vectors, and then generating scheduling strategies for all in-plant logistics forklifts 1. This module is mainly composed of a strategy space mapping unit and a strategy integration unit. The strategy space mapping unit maps the encoded reasoning result to the scheduling strategy space; the strategy integration unit generates the final scheduling strategy based on the mapping result. The specific process is as follows:

[0153] 1. Construct the scheduling strategy space: define the scheduling strategy space. The dimension of the forklift task allocation scheduling strategy vector space is , the dimension of the driving route scheduling strategy vector space is The scheduling strategy space is considered as a dimensional vector space.

[0154] 2. Mapping to the scheduling strategy space: through a mapping function Encode the inference results Mapped to the scheduling strategy space. Mapping function is a linear transformation function consisting of a The weight matrix ($m$ is the inference result encoding dimensions) and a The bias vector of The mapped vector for: .Will Split into two parts, the front The forklift task allocation scheduling strategy vector ,back The driving path scheduling strategy vector : , .

[0155] 3. Generate a scheduling strategy for all in-plant logistics forklifts 1: According to the forklift task allocation scheduling strategy vector and the driving route scheduling strategy vector , combined with the number and specific information of forklifts, generate a scheduling strategy for all in-plant logistics forklifts 1. Forklift, for the Forklift, its task allocation information can be obtained from The path planning information can be obtained from Obtained from.

[0156] Preferably, from According to certain rules (such as index mapping), the The task type, task priority and other information of each forklift; Determine the Specifically, the specific process of generating the scheduling strategy for all in-plant logistics forklifts 1 is as follows: The forklift task allocation scheduling strategy vector and the driving route scheduling strategy vector , and the total number of in-plant logistics forklifts 1 is For task allocation information acquisition, Forklift ( ), from the forklift task allocation scheduling strategy vector The specific steps to obtain task allocation information are as follows:

[0157] 1. Task type determination: During the encoding process, the encoding information of the task type is The front The dimensions are reflected (corresponding to the dimensions of the previous task type encoding). Through an index mapping function To determine the The forklift task type is coded here Preferably, by defining ( is the length of the encoding vector for each task type).

[0158] from Take out the encoding vector of the corresponding position , the median value of this vector is The position corresponds to The type of mission for the forklift. Preferably, if is a dimensional vector, and the The location is , then it means The mission type of the forklift is the predefined Type of tasks.

[0159] 2. Task priority determination: In the example, the encoding information of the task priority follows the task type encoding, and its length is dimensions (corresponding to the dimensions of the previous task priority encoding). Also through the index mapping function To determine the The forklift task priority code is here The position in the dimension, preferably . Get the value at the corresponding position , this value is The normalized priority of the forklift task. To get the actual priority value, you can multiply it by the preset maximum priority value , that is, Actual mission priorities for each forklift .

[0160] In addition, in the path planning information acquisition, for the Forklift, from the driving path scheduling strategy vector The specific steps to obtain path planning information are as follows:

[0161] 1. Path node sequence determination: During the encoding process, the encoding information of the path node sequence is The front The dimensions are reflected (corresponding to the dimensions of the previous path node sequence encoding). To determine the The node sequence of the forklift path is encoded here The position in the dimension, preferably .

[0162] from Take out the encoding vector of the corresponding position , each element in this vector Represents a node in the path. Connect these nodes in order and you get The node sequence of the driving path of a forklift.

[0163] 2. Path time limit determination: In the example, the encoding information of the path time limit is after the path node sequence encoding, and the length is dimensions (corresponding to the dimensions of the previous path time constraint encoding). Through the index mapping function To determine the The time constraints of the forklift path are encoded in this dimension, such as (Because there is only one dimension representing the time limit). Take out the value at the corresponding position This is the first The normalized value of the forklift path time limit. To get the actual time limit value, multiply it by the preset maximum time limit value. , that is, Actual path time limit for forklifts .

[0164] To generate a complete scheduling strategy, task allocation information and path planning information are obtained for each forklift. By integrating this information, a scheduling strategy for all in-plant logistics forklifts 1 can be generated. For example, the sample code is as follows:

[0165] For each forklift, the scheduling strategy can be represented as a structure containing information such as task type, task priority, driving path node sequence and path time limit. Preferably, the following structure can be defined in Python:

[0166] classForkliftSchedule:

[0167] def__init__(self,task_type,task_priority,path_nodes,path_time_limit):

[0168] self.task_type=task_type

[0169] self.task_priority=task_priority

[0170] self.path_nodes = path_nodes

[0171] self.path_time_limit=path_time_limit

[0172] #Store a list of all forklift scheduling strategies

[0173] forklift_schedules=[]

[0174] foriinrange(1,n+1):

[0175] #Get task allocation information

[0176] task_type=get_task_type_from_Y1(Y1,i)

[0177] task_priority=get_task_priority_from_Y1(Y1,i)

[0178] #Get path planning information

[0179] path_nodes=get_path_nodes_from_Y2(Y2,i)

[0180] path_time_limit=get_path_time_limit_from_Y2(Y2,i)

[0181] #Create a scheduling strategy structure

[0182] schedule=ForkliftSchedule(task_type,task_priority,path_nodes,path_time_limit)

[0183] forklift_schedules.append(schedule)

[0184] To this end, the above series of technical processes from scheduling target encoding to generating a complete forklift scheduling strategy have the following technical benefits:

[0185] 1. Data standardization and unified format. Facilitate data processing and model input: Different types of forklift task adjustment targets and path planning target information are converted into a unified vector form through a specific encoding method. This standardized data format makes subsequent processing and analysis more convenient, and various machine learning algorithms or data analysis tools can operate on these encoded data in a unified manner. Whether it is storage, calculation or model training, it can improve efficiency and accuracy and reduce errors caused by inconsistent data formats. Eliminate data heterogeneity: The original task adjustment targets and path planning targets contain many different types of data, such as task type is discrete classification data, task priority is numerical data, path node sequence is sequence data, and path time limit is also numerical data. These heterogeneous data are difficult to directly analyze and process comprehensively. By encoding, they are converted into a unified vector representation, eliminating the difference in data types, so that different information can be operated and compared in the same vector space, providing a solid data foundation for subsequent strategy generation.

[0186] 2. Retain key information. Accurately characterize task and path features: Using unique hot encoding to encode task types can accurately represent the uniqueness of each task type and retain all information about the task type. For the encoding of task priority, path node sequence, and path time limit, reasonable normalization and integer encoding methods are used to compress data while retaining the key features and relative relationships of this information to the greatest extent. For example, the normalization of task priority not only retains the relative size relationship between different task priorities, but also maps them to a unified numerical range, which is convenient for subsequent analysis and calculation. Support multi-dimensional analysis: All encoding vectors are spliced, and the resulting inference result encoding contains multi-dimensional information of tasks and paths. This multi-dimensional encoding representation allows the forklift scheduling objectives to be analyzed and considered from multiple perspectives in the subsequent strategy generation process. For example, the relationship between factors such as task type, priority, and path planning can be comprehensively considered to develop a more comprehensive and reasonable scheduling strategy.

[0187] 3. Flexible mapping and strategy space construction. Customized strategy generation: By defining a flexible scheduling strategy space and mapping the reasoning results to the space, customized forklift scheduling strategies can be generated according to specific logistics needs and business rules. The settings of different forklift task allocation scheduling strategy vector space dimensions a and driving path scheduling strategy vector space dimensions b can be adjusted according to actual conditions to adapt to in-plant logistics systems of different scales and complexities. This flexibility enables the technology to be widely used in various logistics scenarios, improving the versatility and adaptability of the system. Mining potential strategy relationships: The mapping function adopts a linear transformation, which consists of a weight matrix and a bias vector. This linear transformation can not only map the encoded reasoning results to the scheduling strategy space, but also mine the potential relationship between the task adjustment target and the path planning target and the final scheduling strategy by learning the parameters of the weight matrix. For example, by adjusting the weight matrix, it is possible to find the degree of influence of different task types and priorities on forklift task allocation, as well as the relationship between path planning parameters and driving path scheduling strategies, thereby providing a basis for generating more optimized scheduling strategies.

[0188] 4. Accurate information extraction and strategy integration. Accurately obtain task and path information: Through a carefully designed index mapping function, the task allocation information and path planning information of each forklift can be accurately extracted from the forklift task allocation scheduling strategy vector and the driving path scheduling strategy vector. This precise information extraction method ensures that the scheduling strategy of each forklift can accurately reflect its corresponding task and path requirements, avoiding information confusion and errors. Efficient integration to form a complete strategy: Integrate the information obtained from the two vectors to generate a complete scheduling strategy for all in-plant logistics forklifts 1. This integration method organically combines task allocation and path planning, takes into account multiple key factors in the forklift scheduling process, and makes the generated scheduling strategy more comprehensive, reasonable and efficient. In this way, the operational efficiency of in-plant logistics can be effectively improved, resource waste can be reduced, and logistics processes can be optimized.

[0189] Optionally, the abnormal logistics event in the field includes an emergency event of a person falling; the abnormal event recognition model is used to extract key inspection features of the video sampling data, and identify abnormal logistics events in the field according to the extracted key inspection features, including: performing target detection on the human body in the video sampling data to obtain the changes in the height of the center of gravity of the human body and the position of key joints of the human body over time; the key joints of the human body include the head, shoulders, elbows, knees and ankles; based on the human posture stability index SI Determine whether there are motion features of a human body falling, and if the motion features of a human body falling are identified, generate the emergency event of the human body falling; the human body posture stability index SI Calculated based on the following formula: .in, is the first human body in two adjacent frames of video sampling data i The position change of key joints, is the time interval between two adjacent frames of video sampling data, For the j The change in the height of the human body's center of gravity in the frame video sampling data relative to the previous frame, is the height of the center of gravity of the human body in the previous frame, N To calculate the number of frames.

[0190] Optionally, the abnormal logistics event in the field includes an emergency event of cargo falling, and the abnormal event recognition model is used to extract inspection key features of the video sampling data, and identify the abnormal logistics event in the field according to the extracted inspection key features, including: performing target detection on the cargo in the video sampling data to obtain the load center distance; obtaining the fork spacing and the fork height from the ground from the operating status data; based on the cargo placement stability index SSCDetermine whether there is a cargo falling action feature, and if the cargo falling action feature is identified, generate the cargo falling emergency event; the cargo placement stability index SSC Calculated based on the following formula: .in, ; is the load center distance of the current cargo, is the standard load center distance; ; is the relative standard deviation of the current fork spacing, It is the standard fork spacing; ; is the relative standard deviation of the current cargo height above the ground, is the standard height above the ground; W 1 , W 2 , W 3 They are the corresponding weights of the three factors that affect cargo stability: load center distance, fork spacing, and fork height from the ground.

[0191] Optionally, the abnormal logistics events within the field also include open flame emergency events, and the abnormal event recognition model is used to extract inspection key features of the video sampling data, and identify abnormal logistics events within the field based on the extracted inspection key features, including: performing color feature analysis on the video sampling data to obtain video sampling data containing flame color features; performing texture feature analysis on the video sampling data containing flame color features to obtain video sampling data containing flame texture features; performing flame dynamic feature analysis on the video sampling data before and after the video sampling data containing flame texture features to identify open flames; if an open flame is identified, the open flame emergency event is generated. Specifically, the specific implementation process of each step in identifying open flame emergency events is as follows: In order to extract flame color features, a color threshold range can be defined to filter out pixels that may contain flames. Assume that each frame of the video sampling data is ,in and Respectively represent the horizontal and vertical coordinates of the image. For each pixel , whose RGB value is A series of color threshold conditions are set to determine whether the pixel may belong to a flame. For example, common flame color features may meet the following conditions in RGB space: .in , , are the thresholds for the red, green, and blue channels, and is the color difference threshold. Pixels that meet the above conditions are considered to be likely to belong to flames, and the video sampling data where these pixels are located are marked as video sampling data containing flame color features. In the texture feature analysis, it is considered that flames have unique texture features, such as irregular shapes and dynamically changing textures. In order to extract flame texture features, the gray level co-occurrence matrix (GLCM) is used to implement it. For the video sampling data containing flame color features , calculate its GLCM. GLCM describes the probability distribution of different gray level pairs in the image. Indicates the distance and direction The gray value is and The probability of the occurrence of a pixel pair. Multiple texture feature parameters such as contrast, correlation, energy and entropy can be calculated through GLCM. For flame texture, the threshold range of these feature parameters is set to further filter out video sampling data containing flame texture features.

[0192] Taking contrast as an example, the calculation formula is: ;in is the gray level of the image. Assume that the contrast threshold is set to , when the calculated contrast is greater than When , the area is considered to have flame texture features. After analyzing the entire video sampling data containing flame color features, the video sampling data containing flame texture features is obtained. When analyzing the dynamic characteristics of flames, it is considered that flames have dynamic characteristics, such as flickering and fluttering. In order to identify these dynamic characteristics, the video sampling data before and after the video sampling data containing flame texture characteristics are analyzed. The process is as follows: express The video sampling data frame containing the flame texture feature at each moment. The change between two adjacent frames is calculated to analyze the dynamic characteristics of the flame. Preferably, the pixel difference between two adjacent frames is calculated: ;in and are the width and height of the image respectively.

[0193] At the same time, the dynamic characteristics of the flame, such as the area change and shape change, are analyzed. Assuming that the area of ​​the flame is Time is ,exist Time is , then the area change rate is: A series of thresholds are set for dynamic features such as pixel difference and area change rate. When these dynamic features meet the set threshold conditions, it is considered that an open fire has been identified. For example, when and hour( and When an open fire emergency event is generated, if an open fire is finally identified after the above color feature analysis, texture feature analysis, and flame dynamic feature analysis, an open fire emergency event is generated.

[0194] To this end, the above specific implementation process of identifying open fire emergency events has many technical benefits:

[0195] 1. Rapid screening of potential flame areas By defining a color threshold range to screen pixels that may contain flames, potential flame areas can be quickly located from a large amount of video sampling data. This greatly reduces the amount of data to be processed later and improves the efficiency of the entire recognition process. Compared with a comprehensive analysis of the entire image, this preliminary screening based on color features can quickly focus on the part where flames may exist, saving computing resources and time. Color is a significant feature of flames. Using the characteristics of common flame colors in RGB space to set threshold conditions can effectively capture the color information of flames. This method has high accuracy for most common flame scenes and can accurately distinguish flames from other objects, providing a reliable basis for subsequent more accurate feature analysis.

[0196] 2. Strong adaptability The thresholds of the red, green, and blue channels and the color difference threshold can be flexibly adjusted according to different application scenarios and actual needs. For example, under different lighting conditions, or for different types of flames (such as small flames indoors and large fires outdoors), these thresholds can be adjusted to optimize the recognition effect of flame color features, making the system more adaptable and robust.

[0197] 3. Deeply explore flame features Flames have unique texture features. The gray-level co-occurrence matrix (GLCM) can effectively describe the probability distribution of different gray-level pairs in the image, thereby extracting the texture information of the flame. By calculating texture feature parameters such as contrast, correlation, energy, and entropy, the texture features of the flame can be deeply explored from multiple angles. These features further enrich the description of the flame and improve the accuracy of recognition. Compared with simple color features, texture features can provide more information about the shape and structure of the flame. For example, the irregular shape and dynamically changing texture of the flame can be reflected through texture feature parameters, which helps to distinguish flames from other objects with similar colors but different textures, reducing the possibility of misjudgment.

[0198] 4. Customizable screening criteria Set the threshold range of each characteristic parameter for flame texture, and customize the screening criteria according to actual conditions. Different application scenarios may have different requirements for flame texture features. By adjusting the threshold, flame textures that meet specific conditions can be screened more accurately, further improving the reliability of recognition. For example, in some scenarios that require high accuracy in flame recognition, a stricter threshold can be set to ensure that only real flames are recognized.

[0199] 5. Accurately identify dynamic flames The dynamic features of flames, such as flickering and fluttering, are important signs that distinguish flames from other static or stable objects. By analyzing the previous and next frames of the video sampling data containing flame texture features and calculating dynamic features such as pixel difference and area change rate, the dynamic change information of the flame can be accurately captured. This analysis of dynamic features can effectively eliminate some static interference factors and improve the ability to identify real flames. By comprehensively considering multiple dynamic features, such as pixel difference and area change rate, and setting corresponding threshold conditions, it is possible to determine whether it is an open flame from multiple dimensions. This multi-dimensional judgment method is more comprehensive and accurate, avoiding the possible misjudgment based on a single feature, and improving the accuracy and reliability of recognition.

[0200] 6. Real-time monitoring and response This video frame-based dynamic feature analysis method is suitable for real-time monitoring systems. In practical applications, the video sampling data can be analyzed in real time. Once the dynamic features that meet the threshold conditions are detected, the existence of open fire can be quickly determined and an open fire emergency event can be generated in time. This is crucial to ensure the safety of scenes such as in-plant logistics. Alarms can be issued in time at the early stage of a fire, and corresponding measures can be taken to reduce losses.

[0201] 7. Comprehensive flame recognition The entire recognition process uses three steps: color feature analysis, texture feature analysis, and flame dynamic feature analysis to comprehensively describe and analyze the flame from different angles. This multi-step, multi-feature recognition method can make full use of the various characteristics of the flame, improve the accuracy of the recognition of the open flame, and reduce the cases of missed and misjudgment. Each step is interrelated and complementary. Color feature analysis provides potential flame areas for texture feature analysis, texture feature analysis further refines the characteristics of the flame, and flame dynamic feature analysis ultimately determines whether it is a real open flame. This step-by-step in-depth analysis process makes the entire recognition system more robust and reliable.

[0202] Figure 2 Schematic diagram of a logistics scheduling method based on the collection of inspection data of a forklift 1 in a factory according to an embodiment of the present invention. Figure 2As shown, the scheduling method of the logistics scheduling system based on the inspection data collection of the in-plant logistics forklift 1 includes: the camera 3 monitors the operating environment of the in-plant logistics forklift 1 deployed thereon to obtain an environmental monitoring video stream; the data acquisition module samples the environmental monitoring video stream to obtain video sampling data; the abnormal event recognition model extracts the inspection key features of the video sampling data, and identifies the in-field logistics abnormal events based on the extracted inspection key features; the scheduling strategy reasoning model 6 performs the following steps on the in-field logistics abnormal events identified by the abnormal event recognition model on all in-plant logistics forklifts 1 , in order to generate a scheduling strategy for all in-plant logistics forklifts 1: cluster analysis is performed on all in-plant logistics abnormal events to generate an abnormal classification tree; the in-plant logistics resource description is obtained and matched with the abnormal classification tree to generate an impact degree assessment value and an impact range assessment value of all in-plant logistics abnormal events on the overall operation of in-plant logistics; according to the impact degree assessment value and the impact range assessment value, the forklift task adjustment target and the forklift path planning target are inferred; based on the inferred forklift task adjustment target and the forklift path planning target, a scheduling strategy for all in-plant logistics forklifts 1 is generated. Figure 2 In the embodiment, the exemplary explanation of each step can be similar to the above Figure 1 Records of.

Claims

1. A logistics dispatching system based on the collection of in-plant logistics forklift inspection data, characterized in that: include: Multiple in-plant logistics forklifts and dispatch servers. Each in-plant logistics forklift is equipped with a camera and an AI vision host. The AI ​​vision host is equipped with a data acquisition module and an abnormal event recognition model. The dispatch server is equipped with a dispatch strategy reasoning model. The camera is used to monitor the operating environment of the in-plant logistics forklift where it is deployed to obtain an environmental monitoring video stream, the data acquisition module is used to sample the environmental monitoring video stream to obtain video sampling data, and the abnormal event recognition model is used to extract inspection key features from the video sampling data, and identify in-plant logistics abnormal events based on the extracted inspection key features; The scheduling strategy inference model is used to perform the following steps on the abnormal events of in-plant logistics identified by the abnormal event recognition model on all in-plant logistics forklifts to generate a scheduling strategy for all in-plant logistics forklifts: Perform cluster analysis on all abnormal logistics events within the site to generate an abnormal classification tree; Obtaining a description of in-plant logistics resources and matching it with the abnormal classification tree to generate an impact assessment value and an impact range assessment value of all in-plant logistics abnormal events on the overall operation of in-plant logistics; Reasoning the forklift task adjustment target and the forklift path planning target according to the impact degree assessment value and the impact range assessment value; Based on the forklift task adjustment target and forklift path planning target obtained by reasoning, a scheduling strategy for all in-plant logistics forklifts is generated; Among them, the abnormal logistics events within the field include emergency events of personnel falling; The abnormal event recognition model is used to extract inspection key features from the video sampling data, and identify abnormal logistics events within the field according to the extracted inspection key features, including: Performing target detection on the human body in the video sampling data to obtain changes in the height of the human body's center of gravity and the positions of key joints of the human body over time; the key joints of the human body include the head, shoulders, elbows, knees and ankles; Judging whether there are motion features of a human body falling based on the human body posture stability index, and generating the human body falling emergency event if the motion features of a human body falling are identified; The human body posture stability index is calculated based on the following formula: in, SI represents the human body posture stability index, is the first human body in two adjacent frames of video sampling data i The position change of key joints, is the time interval between two adjacent frames of video sampling data, For the j The change in the height of the human body's center of gravity in the frame video sampling data relative to the previous frame, is the height of the center of gravity of the human body in the previous frame, N To calculate the number of frames.

2. A logistics dispatching system based on in-plant logistics forklift inspection data collection according to claim 1, characterized in that: The scheduling strategy reasoning model includes an abnormal event input network and an event feature extraction network, and the event feature extraction network includes a convolution layer, a pooling layer, and a fully connected layer; The cluster analysis of all abnormal logistics events within the field is performed to generate an abnormal classification tree, including: Based on the abnormal event input network, all the abnormal events of logistics within the field are vectorized to generate a characteristic sequence representation of the abnormal events of logistics within the field; Based on the convolution layer, local feature extraction is performed on the feature sequence representation of the abnormal logistics event in the field to generate an abnormal characterization local feature vector, wherein the abnormal characterization local feature vector includes an abnormal characterization edge feature, an abnormal characterization corner feature, an abnormal characterization texture feature, an abnormal characterization color feature, and an abnormal characterization motion feature; Based on the pooling layer, down-sampling the abnormal characterization local feature vector to generate an abnormal characterization key feature vector; Based on the fully connected layer, feature mapping is performed on the key feature vector of the abnormal characterization to obtain a feature representation vector of the abnormal event to generate the abnormal classification tree.

3. A logistics dispatching system based on in-plant logistics forklift inspection data collection according to claim 2, characterized in that: The scheduling strategy reasoning model also includes an inter-feature attention analysis network and a clustering network; The cluster analysis of all the abnormal logistics events within the field to generate an abnormal classification tree also includes: Based on the inter-feature attention analysis network, a multi-head cross-attention analysis is performed on the feature representation vectors of different abnormal events to generate a comprehensive feature vector of the abnormal event; Based on the clustering network, cluster analysis is performed on the comprehensive feature vector of the abnormal event to generate the abnormal classification tree.

4. A logistics dispatching system based on in-plant logistics forklift inspection data collection according to claim 3, characterized in that: The scheduling strategy reasoning model also includes a strategy generation and evaluation network, which includes: a resource description input layer, an abnormal classification tree encoding layer, and a matching and evaluation layer; The acquisition of the on-site logistics resource description and matching with the abnormal classification tree to generate the impact assessment value and impact range assessment value of all on-site logistics abnormal events on the overall operation of the on-site logistics include: Based on the resource description input layer, vectorizing the in-field logistics resource description to generate an in-field logistics resource vector representation; Based on the abnormal classification tree coding layer, encoding the structure and node information of the abnormal classification tree to generate an abnormal classification code; Based on the matching and evaluation layer, the on-site logistics resource vector representation is matched with the abnormal classification code to obtain a matching result, and the matching result is linearly transformed and activated to generate an impact degree evaluation value and an impact range evaluation value.

5. A logistics dispatching system based on in-plant logistics forklift inspection data collection according to claim 4, characterized in that: The scheduling strategy reasoning model also includes a reasoning layer; The reasoning of the forklift task adjustment target and the forklift path planning target according to the impact degree evaluation value and the impact range evaluation value includes: Based on the reasoning layer, the impact degree evaluation value and the impact range evaluation value are linearly transformed to generate a logistics impact reasoning latent vector, and the logistics impact reasoning latent vector is nonlinearly activated to generate an abnormal impact driven logistics scheduling decision representation, and the abnormal impact driven logistics scheduling decision representation includes the inferred forklift task adjustment target and forklift path planning target.

6. A logistics dispatching system based on in-plant logistics forklift inspection data collection according to claim 5, characterized in that: The strategy generation and evaluation network further includes: a strategy generation layer, the strategy generation layer including: a scheduling target encoding module and a strategy generation module; The forklift task adjustment target and forklift path planning target obtained based on the reasoning generate a scheduling strategy for all in-plant logistics forklifts, including: Based on the scheduling target encoding module, the forklift task adjustment target and the forklift path planning target are encoded to generate an inference result code; Based on the strategy generation module, the reasoning result encoding is mapped to the constructed scheduling strategy space to generate a forklift task allocation scheduling strategy vector and a driving path scheduling strategy vector, and accordingly generate a scheduling strategy for all in-plant logistics forklifts.

7. A logistics dispatching system based on in-plant logistics forklift inspection data collection according to any one of claims 1 to 6, characterized in that: The abnormal logistics events within the field also include emergency events of cargo falling; The abnormal event recognition model is used to extract inspection key features from the video sampling data, and identify abnormal logistics events within the field according to the extracted inspection key features, including: Performing target detection on the cargo in the video sampling data to obtain the load center distance; Obtain fork spacing and fork height from the ground from the operating status data; judging whether there is a cargo falling action feature based on the cargo placement stability index, and generating the cargo falling emergency event if the cargo falling action feature is identified; The cargo placement stability index is calculated based on the following formula: in, SSC Indicates the stability index of the cargo placement, ; is the load center distance of the current cargo, is the standard load center distance; ; is the relative standard deviation of the current fork spacing, It is the standard fork spacing; ; is the relative standard deviation of the current cargo height above the ground, is the standard height above the ground; W 1. W 2. W 3 are the corresponding weights of the three factors affecting cargo stability: load center distance, fork spacing, and fork height from the ground.

8. The logistics dispatching system based on the collection of in-plant logistics forklift inspection data according to claim 7 is characterized in that: The abnormal logistics events within the site also include open fire emergencies; The abnormal event recognition model is used to extract inspection key features from the video sampling data, and identify abnormal logistics events within the field according to the extracted inspection key features, including: Performing color feature analysis on the video sampling data to obtain video sampling data containing flame color features; Performing texture feature analysis on the video sampling data containing flame color features to obtain video sampling data containing flame texture features; Performing flame dynamic feature analysis on the video sampling data before and after the video sampling data containing the flame texture feature to identify an open flame; If an open fire is identified, the open fire emergency event is generated.

9. A scheduling method for a logistics scheduling system based on in-plant logistics forklift inspection data collection according to any one of claims 1 to 8, characterized in that: include: The camera monitors the operating environment of the logistics forklift in the factory where it is deployed to obtain an environmental monitoring video stream; The data acquisition module samples the environment monitoring video stream to obtain video sampling data; The abnormal event recognition model extracts inspection key features from the video sampling data, and identifies abnormal logistics events within the field based on the extracted inspection key features; The scheduling strategy inference model performs the following steps on the abnormal events in the field identified by the abnormal event recognition model on all the in-plant logistics forklifts to generate a scheduling strategy for all the in-plant logistics forklifts: Perform cluster analysis on all abnormal logistics events within the site to generate an abnormal classification tree; Obtaining a description of in-plant logistics resources and matching it with the abnormal classification tree to generate an impact assessment value and an impact range assessment value of all in-plant logistics abnormal events on the overall operation of in-plant logistics; Reasoning the forklift task adjustment target and the forklift path planning target according to the impact degree assessment value and the impact range assessment value; Based on the forklift task adjustment target and forklift path planning target obtained by reasoning, a scheduling strategy for all in-plant logistics forklifts is generated.

Images