Storage anomaly detection and intelligent early warning method and system

Abstract

The invention relates to the technical field of storage anomaly judgment, in particular to a storage anomaly detection and intelligent early warning method and system, and provides a two-channel anomaly detection model based on combination of a graph convolutional network and a two-way long-short-term memory network for storage anomaly fault recognition and intelligent early warning. Through a two-channel cooperation mechanism fusing spatial topological feature extraction and time sequence evolution analysis and in combination with an adaptive weight distribution strategy of a multi-modal decision-making layer, accurate capture and dynamic risk assessment of multi-dimensional abnormal features in a storage environment are realized. Therefore, the anomaly detection real-time performance and the early warning accuracy under the complex scene are obviously improved.

Description

Technical Field

[0001] The present invention relates to the technical field of warehouse anomaly judgment, and in particular to a warehouse anomaly detection and intelligent early warning method and system. Background Art

[0002] With the rapid development of logistics automation and intelligent technologies, modern warehousing systems are facing increasingly severe challenges in anomaly detection. In highly dynamic warehousing environments, inventory turnover rates continue to climb, the operational complexity of automated equipment increases significantly, and the range of environmental parameter fluctuations continues to expand. These factors collectively lead to an exponential increase in the diversity and correlation of anomaly events.

[0003] Current warehouse anomaly detection systems primarily utilize traditional detection methods based on fixed thresholds. These systems deploy various sensor networks on shelves, equipment, and the environment to collect real-time data on inventory status, equipment operating parameters, and environmental monitoring. These systems then use pre-set static threshold rules or basic statistical analysis algorithms to identify anomalies and issue early warnings. A typical detection process includes data collection, threshold comparison, anomaly flagging, and alarm triggering. Ultimately, these systems notify management personnel via audible and visual alarms or mobile devices for action.

[0004] However, traditional warehouse anomaly detection systems have obvious deficiencies in their ability to analyze spatiotemporal data correlations, making it difficult to meet the stringent requirements of intelligent warehousing for timely response and accurate early warning. Summary of the Invention

[0005] In view of this, the present application provides a warehouse anomaly detection and intelligent early warning method and system, which solves the technical problems that the existing warehouse anomaly detection system ignores the spatiotemporal correlation characteristics of data, making it difficult to effectively capture abnormal patterns in complex warehouse environments, resulting in low fault identification efficiency of the early warning system and lack of intelligent decision-making capabilities.

[0006] The warehouse anomaly detection and intelligent early warning method provided in this application adopts the following technical solutions: When receiving the historical spatiotemporal data of the warehouse, the historical spatiotemporal data of the warehouse is preprocessed to generate warehouse training data; the warehouse training data is input into the preset initial anomaly detection model for training to generate a dual-path anomaly detection model, wherein the dual-path anomaly detection model includes a spatiotemporal feature extraction layer and a multimodal joint decision layer; Respond to warehouse anomaly detection requests, obtain real-time spatiotemporal data of warehouses, perform preprocessing, and generate target spatiotemporal data; Based on the target spatiotemporal data, feature extraction is performed through the spatiotemporal feature extraction layer to generate a spatiotemporal feature vector; Based on the spatiotemporal feature vectors, a hierarchical alarm signal of storage anomalies is generated through the multimodal joint decision-making layer, and response actions are triggered.

[0007] By adopting the above technical solution, the dual-channel anomaly detection model captures the spatiotemporal correlation of warehouse data through the spatiotemporal feature extraction layer, and then comprehensively evaluates the anomaly probability through the multimodal joint decision-making layer. This enables the system to not only identify abnormal fluctuations in a single data source, but also discover complex abnormal patterns across dimensions and time periods, effectively improving the fault identification efficiency and intelligent decision-making capabilities of warehouse management.

[0008] Preferably, when receiving the storage historical spatiotemporal data, the storage historical spatiotemporal data is preprocessed to generate storage training data, including: Eliminate outliers in the historical spatiotemporal data of the warehouse to obtain first processed data; Using the isolation forest algorithm to remove noise data from the first processed data to obtain second processed data; The second processed data is standardized to generate warehouse training data.

[0009] By adopting the above technical solutions, the outlier removal operation can identify and remove data points that are beyond a reasonable range. The isolation forest algorithm effectively separates local noise in the data distribution through tree structure segmentation depth detection. The normalization process retains the temporal characteristics of the data while eliminating dimensional differences, so that the final generated warehouse training data has better temporal consistency and feature discrimination, providing a high-quality input source for subsequent model training, and improving the robustness and generalization ability of the warehouse anomaly detection model.

[0010] Preferably, outliers in the historical storage spatiotemporal data are removed to obtain first processed data, including: Based on multi-dimensional detection rules, a global scan of the warehouse's historical spatiotemporal data is performed to remove abnormal data and generate preliminary processed data. Non-core fields of the preliminary processed data are removed, and the core fields are interpolated using the historical mean of the same period to obtain intermediate processed data. The intermediate processed data is integrity checked, and the intermediate processed data with integrity check results is used as the first training data.

[0011] By adopting the above technical solution, the multi-dimensional detection rule ensures the comprehensive elimination of abnormal data, and then combines the historical mean interpolation method of the same period to fill in the missing values. Finally, after strict integrity verification, high-quality first-processed data is obtained, which can effectively improve the quality and reliability of historical spatiotemporal data in storage, and provide a more accurate foundation for subsequent data analysis and modeling.

[0012] Preferably, the target spatial data includes target spatial data and target temporal data, and the spatiotemporal feature extraction layer includes a spatial graph feature extraction layer and a temporal feature extraction layer; Based on the target spatiotemporal data, the spatiotemporal feature extraction layer performs feature extraction to generate a spatiotemporal feature vector. Based on the spatiotemporal feature vector, the multimodal joint decision layer generates a graded alarm signal of storage anomaly and triggers a response action, including: Based on the target spatial data, feature extraction is performed through the spatial graph feature extraction layer to generate a global feature vector; Based on the target time data, feature extraction is performed through the temporal feature extraction layer to generate a temporal semantic feature vector; Based on the global feature vector and the temporal semantic feature vector, a hierarchical alarm signal of storage anomaly is generated through the multimodal joint decision layer, and a response action is triggered.

[0013] By adopting the above technical solution, we can fully explore the spatial correlation and temporal evolution laws in the target spatiotemporal data, effectively fuse the output features of the spatial graph feature extraction layer and the temporal feature extraction layer, and form a spatiotemporal feature vector with stronger representation capabilities. It not only retains the topological structure information of the spatial data, but also captures the dynamic change characteristics of the time series, providing a more comprehensive and refined feature representation for post-warehousing anomaly detection and analysis tasks.

[0014] Preferably, the spatial graph feature extraction layer includes a primary residual module, a graph convolution layer, an activation layer, a multi-head attention module, a secondary residual module, a layer normalization layer, and a global average pooling layer connected in sequence. Based on the target spatial data, the spatial graph feature extraction layer performs feature extraction to generate a global feature vector, including: Graph construction is performed on the target spatial data to generate first feature data; Inputting the first feature data into a residual module to perform feature enhancement to obtain feature enhanced data; The feature-enhanced data is input into the graph convolution layer for spatial feature extraction to obtain local feature data; Perform nonlinear transformation on local feature data to generate local linear feature data; Input the local linear feature data into the multi-head self-attention module for weighted aggregation to generate the second feature data; Inputting the first feature data and the second feature data into a quadratic residual module for connection to generate global feature data; The global feature data is input into the global average pooling layer for feature extraction to generate a global feature vector.

[0015] By adopting the above technical solutions, the multi-level structural features of the target spatial data can be effectively captured, the stability of feature transmission can be maintained through residual connections, graph convolution is used to extract spatial local correlations, and the weights of important features are enhanced by combining the multi-head self-attention mechanism. Finally, a global feature vector with strong representation ability is obtained through global average pooling, which realizes the gradual refinement and fusion of spatial features, retaining the spatial topological information of the original data and improving the discriminability and generalization ability of features through deep feature transformation.

[0016] Preferably, the temporal feature extraction layer includes a multi-level temporal feature extraction module, a global average pooling layer, and a bidirectional long short-term memory network layer connected in sequence. Based on the target time data, the temporal feature extraction layer performs feature extraction to generate a temporal semantic feature vector, including: The multi-level temporal feature extraction module performs a convolution operation on the input target time data to generate a multi-scale temporal feature tensor; the global average pooling layer performs a pooling operation on the multi-scale temporal feature tensor to generate a temporal global feature vector; The time series global feature vector is extracted through the bidirectional long short-term memory network layer to generate a time series semantic feature vector.

[0017] By adopting the above technical solutions, the multi-level temporal feature extraction module captures short-term and long-term temporal patterns, the global average pooling layer integrates multi-scale feature information, and the bidirectional long short-term memory network layer further explores the front-end and back-end dependencies of temporal data. The final generated temporal semantic feature vector contains both local temporal details and integrates global contextual information, providing a feature representation with high temporal semantic representation capabilities for spatiotemporal analysis.

[0018] Preferably, the multi-level time series feature extraction module includes multiple time series feature extraction units connected in sequence, and performs convolution operation on the input target time data through the multi-level time series feature extraction module to generate a multi-scale time series feature tensor, including: inputting the target time data into multiple time series feature extraction units connected in sequence for feature extraction, and generating a multi-scale time series feature tensor containing multi-level time series patterns.

[0019] By adopting the above technical solution, the cascade structure of multiple time series feature extraction units gradually extracts time series patterns from local to global, realizing hierarchical feature learning of target time data.

[0020] Preferably, the temporal feature extraction unit includes a convolutional layer, a maximum pooling layer, and a Dropout module connected in sequence, inputs the target time data into the plurality of temporal feature extraction units connected in sequence for feature extraction, and generates a multi-scale temporal feature tensor containing a multi-level temporal pattern, including: Perform primary temporal feature extraction on the target time data through the convolution layer to generate the first feature tensor; The first feature tensor is downsampled in the time dimension through the maximum pooling layer to output the second feature tensor; The Dropout module is used to perform enhanced regularization on the second feature tensor with a preset probability to generate a multi-scale time series feature tensor containing multi-level time series patterns.

[0021] By adopting the above technical solutions, the convolutional layer captures local temporal patterns, the maximum pooling layer enhances the temporal robustness of features and reduces computational complexity, and the Dropout module prevents overfitting through a random inactivation mechanism, giving the model better generalization capabilities. Through the cascaded temporal feature extraction structure, the multi-scale temporal feature tensor ultimately generated not only retains fine-grained temporal characteristics but also integrates high-level temporal semantic information, achieving multi-level and progressive feature abstraction of target temporal data.

[0022] Preferably, the multimodal joint decision layer includes an attention layer and a dual-branch decision layer connected in sequence. Based on the spatiotemporal feature vector, the multimodal joint decision layer generates a hierarchical alarm signal of storage anomaly and triggers a response action, including: fusing the global feature vector and the temporal semantic feature vector through the attention layer to generate a joint feature vector; The dual-branch decision layer is used to perform fault identification and severity prediction on the joint feature vector, generate graded alarm signals for storage anomalies, and trigger emergency response actions.

[0023] By adopting the above technical solution, the attention mechanism is used to adaptively balance the contribution weights of spatial and temporal features, and then the dual-branch structure is used to output accurate anomaly classification results and quantitative severity levels respectively. Ultimately, a graded alarm signal with multi-level semantics is generated, and the matching processing flow is intelligently triggered accordingly, thus building a complete closed loop from feature fusion to decision output.

[0024] The warehouse anomaly detection and intelligent early warning system adopts the following technical solutions: The first data processing module is used to pre-process the historical spatiotemporal data of the warehouse when receiving it, and generate warehouse training data; A model training module is used to input the warehouse training data into a preset initial anomaly detection model for training to generate a dual-path anomaly detection model, wherein the dual-path anomaly detection model includes a spatiotemporal feature extraction layer and a multimodal joint decision layer; The second data processing module responds to the warehouse anomaly detection request, obtains the warehouse real-time spatiotemporal data and performs preprocessing to generate target spatiotemporal data; The extraction module is used to extract features based on the target spatial data through the spatiotemporal feature extraction layer to generate spatiotemporal feature vectors; the decision module is used to generate graded alarm signals of storage anomalies through the multimodal joint decision layer based on the spatiotemporal feature vectors and trigger response actions.

[0025] By adopting the above technical solutions, modular design realizes the whole process from historical data training to real-time detection. The dual-channel anomaly detection model combines the deep extraction of spatiotemporal features with multimodal joint decision-making, which not only ensures the accuracy of anomaly detection, but also realizes the dynamic evaluation of anomaly levels. Ultimately, it forms a complete technical system including data preprocessing, feature learning, intelligent decision-making and emergency linkage, which significantly improves the intelligence level of warehouse management and risk response efficiency.

[0026] In summary, this application includes at least one of the following beneficial technical effects: When receiving the historical spatiotemporal data of the warehouse, the present application pre-processes the historical spatiotemporal data of the warehouse to generate warehouse training data, inputs the warehouse training data into the preset initial anomaly detection model for training, and generates a dual-path anomaly detection model, wherein the dual-path anomaly detection model includes a spatiotemporal feature extraction layer and a multimodal joint decision-making layer, responds to the warehouse anomaly detection request, obtains the real-time spatiotemporal data of the warehouse and pre-processes it, generates target spatiotemporal data, and based on the target spatiotemporal data, performs feature extraction through the spatiotemporal feature extraction layer to generate a spatiotemporal feature vector, and based on the spatiotemporal feature vector, generates a graded alarm signal of the warehouse anomaly through the multimodal joint decision-making layer, and triggers the corresponding response action. By building a dual-path anomaly detection model, we have achieved in-depth mining of the spatiotemporal correlation characteristics of the warehousing environment and fusion of multimodal decisions, greatly improving the accuracy of abnormal pattern recognition and real-time response capabilities in complex scenarios, and providing a closed-loop solution for dynamic perception-intelligent analysis-precise disposal for warehousing management. This solves the technical problems of existing warehousing anomaly detection systems that ignore the spatiotemporal correlation characteristics of data, making it difficult to effectively capture abnormal patterns in complex warehousing environments, resulting in low fault recognition efficiency of the early warning system and a lack of intelligent decision-making capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a flowchart of the steps of the warehouse anomaly detection and intelligent early warning method provided in Example 1 of the present application; Figure 2 This is a flowchart of the steps of the warehouse anomaly detection and intelligent early warning method provided in Example 2 of this application; Figure 3 A flowchart of the method implementation provided in the embodiment of the present application; Figure 4 This is a structural block diagram of the warehouse anomaly detection and intelligent early warning system provided in Example 3 of this application. DETAILED DESCRIPTION

[0028] Please refer to Figure 1 , Figure 1 This is a flowchart of the steps of the warehouse anomaly detection and intelligent early warning method provided in Example 1 of the present invention.

[0029] The warehouse anomaly detection and intelligent early warning method provided in the first embodiment includes: Step 101: When receiving the storage historical spatiotemporal data, pre-process the storage historical spatiotemporal data to generate storage training data.

[0030] It should be noted that historical warehousing spatiotemporal data refers to a multi-source heterogeneous warehousing operation data set recorded within a specific period in the past, including cargo coordinate trajectories, inventory status time series changes, equipment operating parameters, environmental monitoring indicators, and personnel operation logs. These data are associated and aligned through timestamps and spatial location tags.

[0031] Warehouse training data refers to the feature set used to input the initial anomaly detection model for training.

[0032] Step 102: Input the warehouse training data into a preset initial anomaly detection model for training to generate a dual-pathway anomaly detection model, wherein the dual-pathway anomaly detection model includes a spatiotemporal feature extraction layer and a multimodal joint decision layer.

[0033] In this first embodiment, warehouse training data is fed into a preset initial anomaly detection model for fault identification and severity prediction training, generating a dual-path anomaly detection model. The preprocessed target spatiotemporal data is then fed into the trained dual-path anomaly detection model to implement multimodal fault detection for warehouse anomalies. The dual-path anomaly detection model comprises a spatiotemporal feature extraction layer and a multimodal joint decision-making layer.

[0034] Step 103: respond to the warehouse anomaly detection request, obtain the warehouse real-time spatiotemporal data and perform preprocessing to generate target spatiotemporal data.

[0035] It should be noted that real-time spatiotemporal data in warehouses refers to multi-dimensional dynamic data sets collected in real time through devices such as IoT sensors, RFID readers, and video surveillance systems, which can reflect the current inventory distribution, stacking structure, equipment operating status, environmental parameters, and personnel behavior.

[0036] Step 104: Based on the target spatiotemporal data, feature extraction is performed through the spatiotemporal feature extraction layer to generate a spatiotemporal feature vector.

[0037] In the first embodiment of this invention, based on the target spatiotemporal data, the graph convolutional neural network in the spatiotemporal feature extraction layer is combined with the attention mechanism and the bidirectional long short-term memory network (LiLSTM) for joint modeling to capture the global information of the spatial dimension and the sequential dependency of the temporal dimension respectively, generate spatiotemporal feature vectors, and provide the multimodal fusion decision layer with high-quality feature representations with spatial topological structure and temporal evolution laws, thereby significantly improving the accuracy of warehouse anomaly detection.

[0038] Step 105: Based on the spatiotemporal feature vector, a hierarchical alarm signal of storage anomaly is generated through the multimodal joint decision layer, and a response action is triggered.

[0039] In the first embodiment, based on the spatiotemporal feature vector, the fault category and warning level of the warehouse anomaly detection are output through the multimodal joint decision layer, and the corresponding response action is triggered.

[0040] In the first embodiment of the present invention, the fault categories specifically include abnormal quantity of goods, abnormal stacking of goods, abnormal equipment and machinery, abnormal environmental safety and illegal operation of personnel, and the warning levels include low risk, medium risk and high risk. Among them, the alarm signal includes the fault category and the warning level. The warning level is directly related to the response action, which is used to automatically dispatch warehouse emergency resources. It should be noted that triggering the corresponding response action specifically includes: low-risk warning triggering system log recording and manual review process, medium-risk warning triggering equipment speed reduction and local area blockade, and high-risk warning immediately starting the whole warehouse shutdown plan and simultaneously notifying the security management terminal to intervene.

[0041] In the present invention, when the historical spatiotemporal data of the warehouse is received, the historical spatiotemporal data of the warehouse is preprocessed to generate warehouse training data, and the warehouse training data is input into a preset initial anomaly detection model for training to generate a dual-path anomaly detection model, wherein the dual-path anomaly detection model includes a spatiotemporal feature extraction layer and a multimodal joint decision layer, responds to the warehouse anomaly detection request, obtains the real-time spatiotemporal data of the warehouse and preprocesses it to generate target spatiotemporal data, and based on the target spatiotemporal data, performs feature extraction through the spatiotemporal feature extraction layer to generate a spatiotemporal feature vector, and based on the spatiotemporal feature vector, generates a graded alarm signal of the warehouse anomaly through the multimodal joint decision layer, and triggers the corresponding response action, which solves the technical problem that the existing warehouse anomaly detection system ignores the spatiotemporal correlation characteristics of the data, is difficult to effectively capture the abnormal pattern in the complex warehouse environment, resulting in low fault recognition efficiency and lack of intelligent decision-making ability of the early warning system. This paper proposes a dual-pathway anomaly detection model based on the combination of graph convolutional networks and bidirectional long short-term memory networks for warehouse anomaly detection fault identification and intelligent early warning. Through the dual-pathway collaborative mechanism that integrates spatial topological feature extraction and time series evolution analysis, combined with the adaptive weight allocation strategy of the multimodal decision layer, it achieves accurate capture of multi-dimensional abnormal features in the warehouse environment and dynamic risk assessment, thereby significantly improving the real-time performance of anomaly detection and early warning accuracy in complex scenarios, and ultimately forming an explainable and scalable intelligent warehouse safety control system.

[0042] Please refer to Figure 2 , Figure 2 This is a flowchart of the steps of the warehouse anomaly detection and intelligent early warning method provided in Example 2 of the present invention.

[0043] The warehouse anomaly detection and intelligent early warning method provided in the second embodiment includes: Step 201: When receiving the storage historical spatiotemporal data, pre-process the storage historical spatiotemporal data to generate storage training data.

[0044] Specifically, step 201 includes the following sub-steps: S11. When receiving the historical spatiotemporal data of the warehouse, perform a global scan on the historical spatiotemporal data of the warehouse based on the multi-dimensional detection rules, eliminate abnormal data, and generate preliminary processed data.

[0045] It should be noted that multi-dimensional detection rules refer to a data verification system based on multiple independent evaluation criteria, which is used to verify the rationality of warehouse spatiotemporal data from different perspectives to avoid misjudgments or missed detections caused by a single detection method. The multi-dimensional detection rules are detailed in the following table: Table 1. Comparison table of multi-dimensional detection rules Abnormal data refers to invalid or erroneous recorded data that exceeds the fluctuation range of preset detection indicators or clearly violates business logic.

[0046] S12. Eliminate non-core fields of the preliminary processed data, and interpolate the core fields using the historical mean of the same period to obtain intermediate processed data.

[0047] It should be noted that non-core fields refer to auxiliary data fields that are not directly related to the core business logic of the warehouse. In this embodiment, non-core fields specifically include secondary information such as equipment log number and operator employee number.

[0048] Core fields refer to key business data indicators that directly affect warehouse spatiotemporal data analysis and decision-making. In this embodiment, core fields specifically include basic data dimensions such as cargo coordinate location, timestamp record, inventory status identification, cargo weight information, etc.

[0049] S13: Perform integrity check on the intermediate processed data, and use the intermediate processed data with integrity check results as the first training data.

[0050] Specifically, if the verification result of the intermediate processed data is incomplete, the fallback mechanism is triggered, and steps S11 and S12 are repeated for a second repair until the preset integrity threshold requirement is met.

[0051] S14. Using the isolation forest algorithm to remove noise data from the first processed data to obtain second processed data.

[0052] It should be noted that Isolation Forest is an ensemble learning algorithm based on random binary trees, which is used to efficiently detect outliers in data.

[0053] The working principle of the isolation forest algorithm is to randomly divide the feature space so that outliers are isolated more quickly due to their small number and large feature differences, thereby calculating the path length as the anomaly score.

[0054] S15. Standardize the second processed data to generate warehouse training data.

[0055] In the second embodiment, when the historical spatiotemporal data of the warehouse is received, the historical spatiotemporal data of the warehouse is globally scanned based on the multi-dimensional detection rules, abnormal data is eliminated, and preliminary processing data is generated. The non-core fields of the preliminary processing data are eliminated, and the core fields are interpolated using the historical mean of the same period to obtain intermediate processing data. The intermediate processing data is checked for integrity, and the intermediate processing data with a complete check result is used as the first training data. The isolation forest algorithm is used to remove noise data from the first processing data to obtain second processing data. The second processing data is standardized to generate warehouse training data, thereby achieving efficient cleaning and feature enhancement of the historical spatiotemporal data of the warehouse, effectively solving the problems of noise interference, field redundancy and missing values ​​in the original data, and providing a high-quality data foundation for subsequent anomaly detection model training.

[0056] Step 202: Input the warehouse training data into a preset initial anomaly detection model for training to generate a dual-path anomaly detection model, wherein the dual-path anomaly detection model includes a spatiotemporal feature extraction layer and a multimodal joint decision layer.

[0057] Furthermore, step 202 may include the following sub-steps: S21. Input the warehouse training data into the initial anomaly detection model to generate the corresponding training refined feature vector.

[0058] The training refined feature vector refers to the feature vector generated by inputting warehouse training data into the initial anomaly detection model, which is used to determine whether the initial anomaly detection model has been trained.

[0059] In the second embodiment, the generated warehouse training data is input into the initial anomaly detection model to generate corresponding training refined feature vectors.

[0060] S22. Calculate the training loss value between the training refined feature vector and the associated standard refined feature vector.

[0061] The standard refined feature vector refers to the standard refined feature vector corresponding to the warehouse training data.

[0062] The training loss value refers to the loss value between the training refined feature vector and the standard refined feature vector.

[0063] In the second embodiment, a training loss value between a training refined feature vector and an associated standard refined feature vector is calculated.

[0064] S23. Compare the training loss value with a preset training standard threshold.

[0065] The training standard threshold refers to the threshold used to determine whether the initial anomaly detection model is trained.

[0066] In the second embodiment, the training loss value is compared with a preset training standard threshold.

[0067] S24. If the training loss value is less than or equal to the training standard threshold, the training is stopped and a dual-path anomaly detection model is generated.

[0068] In the second embodiment, if the training loss value is less than or equal to the training standard threshold, the training is stopped and a dual-pathway anomaly detection model is generated.

[0069] Furthermore, step 202 may also include the following sub-steps: S25. If the training loss value is greater than the training standard threshold, the parameters of the initial anomaly detection model are adjusted according to the preset gradient, and the process jumps to the step of inputting the warehouse training data into the initial anomaly detection model to generate the corresponding training refined feature vector.

[0070] In addition, in the second embodiment, the spatiotemporal feature extraction layer includes a spatial graph feature extraction layer and a temporal feature extraction layer.

[0071] Step 203: respond to the warehouse anomaly detection request, obtain the warehouse real-time spatiotemporal data and perform preprocessing to generate target spatiotemporal data.

[0072] Among them, the step of preprocessing the warehouse real-time spatiotemporal data to obtain the target spatiotemporal data is similar to sub-steps S11-S15 and will not be repeated here.

[0073] In addition, in the second embodiment, the target space data includes target space data and target time data.

[0074] Step 204: Based on the target spatial data, feature extraction is performed through the spatial graph feature extraction layer to generate a global feature vector.

[0075] Further, please refer to Figure 3 The spatial graph feature extraction layer includes a first residual module, a graph convolution layer, an activation layer, a multi-head attention module, a second residual module, a layer normalization layer, and a global average pooling layer connected in sequence. Step 204 may include the following sub-steps: S41. Graph construction is performed on the target spatial data to generate first feature data.

[0076] In the second embodiment, the cargo storage units in the target spatial data are used as vertices, and the spatial distance, stacking relationship and transportation path between the cargoes are used as edges. The k-nearest neighbor algorithm is used to determine the spatial adjacency relationship, and a dynamic weighted graph structure is constructed. The edge weights are calculated based on the spatial distance between the cargoes and the correlation between the transportation paths. Three-dimensional spatial coordinates, RFID tags, weight distribution and environmental parameters are introduced as node attributes, and finally a multi-level topological graph reflecting the real-time spatial relationship of the warehouse and the attributes of the cargo is generated.

[0077] S42: Input the first feature data into a residual module to perform feature enhancement to obtain feature enhanced data.

[0078] In this second example, a 1×1 graph convolution is used to linearly project the input features, expanding the feature dimensions while maintaining the topological structure. The original features are then added to the projected features via cross-layer identity mapping to form a residual connection. Finally, batch normalization and LeakyReLU activation are performed. This enhances feature expression while preserving the original spatial relationships. This allows the output feature-enhanced data to maintain the salience of key anomaly features while avoiding the oversmoothing caused by multiple neighborhood aggregations in graph convolutional neural networks. In specific experiments, the improvement in node feature similarity was kept below 15%.

[0079] S43. Input the feature enhancement data into the graph convolution layer to extract spatial features and obtain local feature data.

[0080] It should be noted that this embodiment uses a two-layer Chebyshev graph convolution kernel to extract spatial features in the frequency domain. By limiting the receptive field range (K = 2nd-order neighborhood), it balances the capture of local details with the risk of oversmoothing. At the same time, it introduces an attention weight adjustment mechanism based on the cargo movement path. This allows the output local feature data to effectively characterize abnormal deformation of cargo stacking while distinguishing reasonable position changes caused by normal operations. In this second embodiment, the feature response value for tilt angles > 15° increased by nearly 40%.

[0081] S44. Perform nonlinear transformation on the local feature data to generate local linear feature data.

[0082] In the second embodiment, the local feature data is input into the ReLU activation layer for nonlinear transformation. The negative noise is filtered out through thresholding, and the positive activation features are retained to enhance the model's sensitivity to abnormal events. At the same time, the adaptive gradient clipping technology is combined to control the risk of gradient explosion, and finally local linear feature data is generated, so that the response value of key abnormal features (such as cargo displacement or stacking tilt) is significantly improved, while the fluctuation range of normal features remains in a reasonable range.

[0083] S45. Input the local linear feature data into the multi-head self-attention module for weighted aggregation to generate second feature data.

[0084] In the second embodiment of the present invention, a multi-head attention mechanism is used to obtain the long-range dependency relationship between cargo nodes. The attention weights of spatially adjacent nodes, transportation path-related nodes, and nodes with similar environmental parameters are calculated separately by eight parallel attention heads. The second feature data finally generated by aggregation contains both the fine-grained features of the local stacking structure and the global collaborative relationship of cross-regional cargo.

[0085] It should be noted that the use of 8 attention heads in this embodiment 2 is the optimal configuration verified by experiments, specifically based on two considerations: (1) ablation experiments show that when the number of attention heads increases from 4 to 8, the feature capture capability of irregular cargo stacking (such as special-shaped stacking with a height difference of >1.2m) is improved by 31%, while further increasing it to 12 heads only brings a marginal gain of 3%; (2) to meet the requirements of warehousing scene characteristics, 8 independent attention heads can focus on association patterns of different dimensions such as spatial neighborhood (2 heads), transportation path (3 heads), environmental parameters (2 heads) and dynamic behavior (1 head), which is in line with the multimodal characteristics of warehousing data.

[0086] S46: Input the first feature data and the second feature data into a quadratic residual module for connection to generate global feature data.

[0087] In the second embodiment, the quadratic residual module is an improved design for the "information bottleneck" problem existing in the traditional residual network during deep feature fusion. Its core idea is to deeply couple the first feature data (low-level local features) and the second feature data (high-level semantic features) through a hierarchical cascade residual structure and a cross-scale feature interaction mechanism to generate global feature data with both position details and semantic information. This not only solves the problem of insufficient feature expression of a single residual connection in complex scenarios, but also avoids the destruction of existing features by the introduced attention mechanism.

[0088] S47. Input the global feature data into the global average pooling layer for feature extraction to generate a global feature vector.

[0089] In the second embodiment, the global average pooling layer compresses the N×d-dimensional graph feature matrix into a 1×d-dimensional global feature vector by calculating the mean of all feature channels along the node dimension. This eliminates node position sensitivity while retaining key information of the entire graph, thereby improving the model's robustness to topological changes in the graph structure.

[0090] The global mean V of each feature channel c , specifically: in, represents the eigenvalue of the i-th node in the c-th channel, and N is the total number of nodes.

[0091] Step 205: Based on the target time data, feature extraction is performed through the temporal feature extraction layer to generate a temporal semantic feature vector.

[0092] Further, please refer to Figure 3 The temporal feature extraction layer includes a multi-level temporal feature extraction module, a global average pooling layer, and a bidirectional long short-term memory network layer connected in sequence. Step 205 may include the following sub-steps: S51. Perform a convolution operation on the input target time data through a multi-level time series feature extraction module to generate a multi-scale time series feature tensor.

[0093] In the second embodiment, the multi-level temporal feature extraction module includes two sequentially connected temporal feature extraction units, and the two temporal feature extraction units are used to perform convolution operations on the target time data in turn to generate a multi-scale temporal feature tensor.

[0094] The temporal feature extraction unit includes a convolutional layer, a maximum pooling layer, and a Dropout module connected in sequence. Step S51 may include the following sub-steps: S101, performing primary time series feature extraction on the target time data through a convolutional layer to generate a first feature tensor; S102, downsampling the first feature tensor in the time dimension through a maximum pooling layer, and outputting a second feature tensor; S103 , performing enhanced regularization processing on the second feature tensor with a preset probability through a Dropout module to generate a multi-scale time series feature tensor containing a multi-level time series pattern.

[0095] It's worth noting that the structure of each time series feature extraction unit is consistent; it can be understood as follows: the target time data is sequentially input into two time series feature extraction units, the first time series feature extraction unit generates a first multi-scale time series feature tensor, and then the first multi-scale time series feature tensor is input into the second time series feature extraction unit, which outputs the final multi-scale time series feature tensor. It should be noted that the multi-level time series feature extraction module can include three, but no more than five, sequentially connected time series feature extraction units.

[0096] S52. Perform a pooling operation on the multi-scale time series feature tensor through a global average pooling layer to generate a time series global feature vector.

[0097] S53. Extract features from the temporal global feature vector through a bidirectional long short-term memory network layer (LiLSTM) to generate the temporal semantic feature vector.

[0098] In the second embodiment, the bidirectional long short-term memory network layer adopts a parallel structure of forward and backward long short-term memory networks, and performs contextual semantic modeling on the temporal global feature vector from the forward and reverse time dimensions through a gating mechanism, and finally concatenates the bidirectional hidden state vector to form a temporal semantic feature vector with a dimension of 2H (H is the number of hidden units in a single-layer long short-term memory network).

[0099] The gating mechanism is the core design of the Long Short-Term Memory (LSTM) network. It controls the flow of information and addresses the long-term dependency and vanishing / exploding gradient problems in traditional recurrent neural networks. Essentially, it dynamically regulates the retention and forgetting of information through a learnable "gate" structure, thereby enabling the modeling of long-range dependencies in time series data.

[0100] Among them, in the bidirectional long short-term memory network, the forward and backward long short-term memory networks both use a gating mechanism independently: the gating mechanism of the bidirectional long short-term memory network (BiLSTM) realizes dynamic control of information through a carefully designed gate structure. Its core lies in that the forward and backward long short-term memory networks can independently decide whether to retain or forget information when they operate independently. In the forward long short-term memory network, the network processes the input sequence step by step along the time order, and selectively updates the internal memory state through the collaborative work of the forget gate, input gate and output gate. The forget gate is responsible for evaluating the retention value of the memory content of the previous moment, the input gate filters the new information worth remembering in the current input, and the output gate controls the contribution of the current memory state to the final hidden state. This embodiment enables the network to capture long-distance temporal dependencies through the gating system, effectively avoiding the gradient disappearance problem commonly seen in long sequence processing.

[0101] Meanwhile, the reverse LSTM network operates in a completely symmetrical manner, except that its processing order is opposite to the flow of time. This reverse processing gives the network the unique ability to infer the possible meaning of the current state from future information. The gating mechanism in the reverse network also includes a forget gate, an input gate, and an output gate. These gates are independent of the corresponding structures in the forward network and specifically learn how to understand and process temporal information from a backward perspective. The processing in these two directions is independent but complementary. At each time point in the sequence, the forward LSTM network captures the accumulated information from the beginning of the sequence to the current moment, while the reverse LSTM network integrates the context from the end of the sequence back to the current moment.

[0102] Ultimately, the bidirectional long short-term memory network concatenates and merges the hidden states in both directions, integrating the bidirectional contextual representation of the time series. The forward state contains the trajectory of historical development and causal logic, while the backward state carries the prediction of future trends and the understanding of results. It can simultaneously reflect the historical causes and future impacts of time series data.

[0103] Step 206: Based on the global feature vector and the temporal semantic feature vector, a hierarchical alarm signal of storage anomaly is generated through the multimodal joint decision layer, and a response action is triggered.

[0104] Further, please refer to Figure 3 The multimodal joint decision layer includes an attention layer and a dual-branch decision layer connected in sequence. Step 206 may include the following sub-steps: S61. The global feature vector and the temporal semantic feature vector are fused through the attention layer to generate a joint feature vector.

[0105] In the second embodiment, the attention layer realizes the deep interaction between the global feature vector and the temporal semantic feature vector through a dynamic weight allocation mechanism, that is, firstly, the similarity matrix of the two features in the latent space is calculated, and then the attention weight is generated through softmax normalization, and finally a joint feature vector with cross-modal correlation characteristics is generated in a weighted summation manner, so that the system can automatically focus on the feature combination that is most relevant to the current warehouse status. For example, when an abnormal cargo stacking rate is detected, the attention to the temporal features of the relevant time period will be enhanced.

[0106] S62. Perform fault identification and severity prediction on the joint feature vector through the dual-branch decision layer, generate a graded alarm signal of storage anomaly, and trigger a response action.

[0107] Furthermore, the dual-branch decision layer includes a first decision layer and a second decision layer.

[0108] The first decision layer includes a fully connected network and a softmax classifier. The fully connected network and the softmax classifier perform feature transformation and normalization on the joint feature vector, and output the probability distribution of the abnormal category, thereby accurately identifying specific failure modes in the warehousing system.

[0109] The second decision layer includes a fully connected network and a sigmoid function. It performs nonlinear mapping on the same joint feature vector through the fully connected network and the sigmoid function, outputs a continuous severity score between 0 and 1, and quantitatively evaluates the potential harm of the fault.

[0110] In this second embodiment, the probability distribution of abnormal categories output by the first decision layer is first analyzed. When the probability value of a category exceeds a preset judgment threshold, a fault of that category is determined to have occurred. Within the probability distribution of abnormal categories output by the first decision layer, the judgment thresholds for each fault type are set as follows: 0.85-0.90 for abnormal cargo quantity, 0.80-0.85 for abnormal cargo stacking, 0.70-0.75 for abnormal equipment and machinery, 0.90-0.95 for abnormal environmental safety, and 0.65-0.70 for illegal personnel operation.

[0111] Simultaneously, the severity score output by the second decision layer is compared with the pre-set risk threshold, and the risk level is determined based on the score range. The coordinated output of the two-branch decision layer comprehensively determines whether an anomaly exists and determines the alert level. The severity score output by the second decision layer is compared with the grading threshold: a score below 0.3 is considered low risk, triggering only system logging and manual review; a score between 0.3 and 0.6 is considered medium risk, triggering equipment downtime and localized lockdown; a score exceeding 0.6 is considered high risk, immediately initiating a shutdown plan and simultaneously notifying the security management terminal for intervention. Specifically, for two high-risk fault categories, mechanical anomalies and environmental safety anomalies, the system employs stricter linkage criteria: a probability exceeding 0.7 and a score greater than 0.5 upgrades to a medium-risk alert, while a score of 0.4 may trigger a high-risk response. It should be noted that this dual-threshold assessment mechanism is trained based on historical operational data and optimizes the balance between precision and recall through receiver operating characteristic (ROC) curve analysis.

[0112] In the present invention, when the historical spatiotemporal data of the warehouse is received, the historical spatiotemporal data of the warehouse is preprocessed to generate warehouse training data, and the warehouse training data is input into a preset initial anomaly detection model for training to generate a dual-path anomaly detection model, wherein the dual-path anomaly detection model includes a spatiotemporal feature extraction layer and a multimodal joint decision layer, responds to the warehouse anomaly detection request, obtains the real-time spatiotemporal data of the warehouse and preprocesses it to generate target spatiotemporal data, and based on the target spatiotemporal data, performs feature extraction through the spatiotemporal feature extraction layer to generate a spatiotemporal feature vector, and based on the spatiotemporal feature vector, generates a graded alarm signal of the warehouse anomaly through the multimodal joint decision layer, and triggers the corresponding response action, which solves the technical problem that the existing warehouse anomaly detection system ignores the spatiotemporal correlation characteristics of multi-source data, is difficult to effectively capture the abnormal pattern in the complex warehouse environment, resulting in low fault recognition efficiency and lack of intelligent decision-making ability of the early warning system. This paper proposes a dual-pathway anomaly detection model based on the combination of graph convolutional networks and bidirectional long short-term memory networks for warehouse anomaly detection fault identification and intelligent early warning. Through the dual-pathway collaborative mechanism that integrates spatial topological feature extraction and time series evolution analysis, combined with the adaptive weight allocation strategy of the multimodal decision layer, it achieves accurate capture of multi-dimensional abnormal features in the warehouse environment and dynamic risk assessment, thereby significantly improving the real-time performance of anomaly detection and early warning accuracy in complex scenarios, and ultimately forming an explainable and scalable intelligent warehouse safety control system.

[0113] Please refer to Figure 4 , Figure 4 This is a flowchart of the steps of the warehouse anomaly detection and intelligent early warning system provided in Example 3 of the present invention.

[0114] The first data processing module 301 is used to pre-process the historical spatiotemporal data of the warehouse upon receiving it and generate warehouse training data; The model training module 302 is used to input warehouse training data into a preset initial anomaly detection model for training, generating a dual-path anomaly detection model, wherein the dual-path anomaly detection model includes a spatiotemporal feature extraction layer and a multimodal joint decision layer; the second data processing module 303 responds to warehouse anomaly detection requests, obtains real-time spatiotemporal data from the warehouse, performs preprocessing, and generates target spatiotemporal data; An extraction module 304 is configured to extract features based on the target spatial data through a spatiotemporal feature extraction layer to generate a spatiotemporal feature vector; The decision module 305 is used to generate a graded alarm signal of storage anomaly based on the spatiotemporal feature vector through a multimodal joint decision layer and trigger corresponding emergency response actions.

[0115] Furthermore, warehouse historical spatiotemporal data refers to a multi-source heterogeneous warehouse operation data set recorded within a specific period in the past, including cargo coordinate trajectories, inventory status time series changes, equipment operating parameters, environmental monitoring indicators, and personnel operation logs. These data are associated and aligned through timestamps and spatial location tags.

[0116] Warehouse training data refers to the feature set used to input the initial anomaly detection model for training.

[0117] Furthermore, the spatial graph feature extraction layer includes a first residual module, a graph convolution layer, an activation layer, a multi-head attention module, a second residual module, a layer normalization layer and a global average pooling layer connected in sequence. Furthermore, the extraction module includes: A first feature data submodule is used to construct a graph of the target spatial data and generate first feature data; A feature enhancement data submodule, configured to input the first feature data into a primary residual module for feature enhancement to obtain feature enhancement data; The local feature data submodule is used to input the feature enhancement data into the graph convolution layer for spatial feature extraction to obtain local feature data; The local linear feature data submodule is used to perform nonlinear transformation on the local feature data to generate local linear feature data; the second feature data submodule is used to input the local linear feature data into the multi-head self-attention module for weighted aggregation to generate second feature data; A global feature data submodule, configured to input the first feature data and the second feature data into a quadratic residual module for connection to generate global feature data; The global feature vector submodule is used to input global feature data into the global average pooling layer for feature extraction and generate a global feature vector.

[0118] Furthermore, the temporal feature extraction layer includes a multi-level temporal feature extraction module, a global average pooling layer, and a bidirectional long short-term memory network layer connected in sequence. Furthermore, the extraction module also includes: The multi-scale time series feature tensor submodule is used to perform convolution operations on the input target time data through the multi-level time series feature extraction module to generate a multi-scale time series feature tensor; The time series global feature vector submodule is used to perform pooling operations on multi-scale time series feature tensors through a global average pooling layer to generate a time series global feature vector; The temporal semantic feature vector sub-module is used to extract features from the temporal global feature vector through a bidirectional long short-term memory network layer to generate a temporal semantic feature vector.

[0119] Furthermore, the time series feature extraction unit includes a convolutional layer, a maximum pooling layer, and a Dropout module connected in sequence. Furthermore, the multi-scale time series feature tensor submodule includes: A first feature tensor unit is used to extract primary time series features from the target time data through a convolutional layer to generate a first feature tensor; A second feature tensor unit is used to downsample the first feature tensor in the time dimension through a maximum pooling layer and output a second feature tensor; The multi-scale time series feature tensor unit is used to perform enhanced regularization processing on the second feature tensor with a preset probability through the Dropout module to generate a multi-scale time series feature tensor containing multi-level time series patterns.

[0120] Furthermore, the multimodal joint decision layer includes an attention layer and a dual-branch decision layer connected in sequence. Furthermore, the decision module includes: The joint feature vector submodule is used to fuse the global feature vector and the temporal semantic feature vector through the attention layer to generate a joint feature vector; The alarm submodule is used to perform fault identification and severity prediction on the joint feature vector through a dual-branch decision layer, generate graded alarm signals for storage anomalies, and trigger response actions.

[0121] In the present invention, when the historical spatiotemporal data of the warehouse is received, the historical spatiotemporal data of the warehouse is preprocessed to generate warehouse training data, and the warehouse training data is input into a preset initial anomaly detection model for training to generate a dual-path anomaly detection model, wherein the dual-path anomaly detection model includes a spatiotemporal feature extraction layer and a multimodal joint decision layer, responds to the warehouse anomaly detection request, obtains the real-time spatiotemporal data of the warehouse and preprocesses it to generate target spatiotemporal data, and based on the target spatiotemporal data, performs feature extraction through the spatiotemporal feature extraction layer to generate a spatiotemporal feature vector, and based on the spatiotemporal feature vector, generates a graded alarm signal of the warehouse anomaly through the multimodal joint decision layer, and triggers the corresponding response action, which solves the technical problem that the existing warehouse anomaly detection system ignores the spatiotemporal correlation characteristics of multi-source data, is difficult to effectively capture the abnormal pattern in the complex warehouse environment, resulting in low fault recognition efficiency and lack of intelligent decision-making ability of the early warning system. This paper proposes a dual-pathway anomaly detection model based on the combination of graph convolutional networks and bidirectional long short-term memory networks for warehouse anomaly detection fault identification and intelligent early warning. Through the dual-pathway collaborative mechanism that integrates spatial topological feature extraction and time series evolution analysis, combined with the adaptive weight allocation strategy of the multimodal decision layer, it achieves accurate capture of multi-dimensional abnormal features in the warehouse environment and dynamic risk assessment, thereby significantly improving the real-time performance of anomaly detection and early warning accuracy in complex scenarios, and ultimately forming an explainable and scalable intelligent warehouse safety control system.

[0122] The above are all preferred embodiments of the present application, and are not intended to limit the scope of protection of the present application. Therefore, any equivalent changes made based on the structure, shape, and principle of the present application should be included in the scope of protection of the present application.

Claims

1. Warehouse anomaly detection and intelligent early warning method, characterized by: include: When receiving the acquired warehouse historical spatiotemporal data, preprocessing the warehouse historical spatiotemporal data to generate warehouse training data; Inputting the warehouse training data into a preset initial anomaly detection model for training to generate a dual-path anomaly detection model, wherein the dual-path anomaly detection model includes a spatiotemporal feature extraction layer and a multimodal joint decision layer; Respond to warehouse anomaly detection requests, obtain real-time spatiotemporal data of warehouses, perform preprocessing, and generate target spatiotemporal data; Based on the target spatiotemporal data, feature extraction is performed through the spatiotemporal feature extraction layer to generate a spatiotemporal feature vector; Based on the spatiotemporal feature vector, a hierarchical alarm signal of storage anomaly is generated through the multimodal joint decision layer, and a response action is triggered.

2. The warehouse anomaly detection and intelligent early warning method according to claim 1 is characterized in that: When receiving the storage historical spatiotemporal data, preprocessing the storage historical spatiotemporal data to generate storage training data includes: Eliminating abnormal values ​​in the warehouse historical spatiotemporal data to obtain first processed data; Using an isolation forest algorithm to remove noise data from the first processed data to obtain second processed data; The second processed data is standardized to generate the warehouse training data.

3. The warehouse anomaly detection and intelligent early warning method according to claim 2 is characterized in that: The step of removing abnormal values ​​from the storage historical spatiotemporal data to obtain first processed data includes: Perform a global scan of the warehouse historical spatiotemporal data based on multi-dimensional detection rules, eliminate abnormal data, and generate preliminary processed data; Eliminating non-core fields of the preliminary processed data, and interpolating the core fields using historical averages of the same period to obtain intermediate processed data; An integrity check is performed on the intermediate processed data, and the intermediate processed data with a complete check result is used as the first training data.

4. The warehouse anomaly detection and intelligent early warning method according to claim 1 is characterized in that: The target spatial data includes target spatial data and target time data, and the spatiotemporal feature extraction layer includes a spatial graph feature extraction layer and a temporal feature extraction layer; The step of extracting features based on the target spatiotemporal data through the spatiotemporal feature extraction layer to generate a spatiotemporal feature vector, and generating a graded alarm signal of storage anomaly based on the spatiotemporal feature vector through the multimodal joint decision layer to trigger a response action includes: Based on the target spatial data, performing feature extraction through the spatial graph feature extraction layer to generate a global feature vector; Based on the target time data, feature extraction is performed through the temporal feature extraction layer to generate a temporal semantic feature vector; Based on the global feature vector and the temporal semantic feature vector, a hierarchical alarm signal of storage anomaly is generated through the multimodal joint decision layer, and a response action is triggered.

5. The warehouse anomaly detection and intelligent early warning method according to claim 4 is characterized in that: The spatial graph feature extraction layer includes a primary residual module, a graph convolution layer, an activation layer, a multi-head attention module, a secondary residual module, a layer normalization layer, and a global average pooling layer connected in sequence. The feature extraction based on the target spatial data is performed through the spatial graph feature extraction layer to generate a global feature vector, including: Performing graph construction on the target space data to generate first feature data; Inputting the first feature data into the primary residual module for feature enhancement to obtain feature enhanced data; Inputting the feature enhancement data into the graph convolution layer to extract spatial features to obtain local feature data; Performing nonlinear transformation on the local feature data to generate local linear feature data; Inputting the local linear feature data into the multi-head self-attention module for weighted aggregation to generate second feature data; Inputting the first feature data and the second feature data into the quadratic residual module for connection to generate global feature data; The global feature data is input into the global average pooling layer for feature extraction to generate a global feature vector.

6. The warehouse anomaly detection and intelligent early warning method according to claim 4 is characterized in that: The temporal feature extraction layer includes a multi-level temporal feature extraction module, a global average pooling layer, and a bidirectional long short-term memory network layer connected in sequence. The temporal feature extraction layer performs feature extraction based on the target time data to generate a temporal semantic feature vector, including: Performing a convolution operation on the input target time data through the multi-level time series feature extraction module to generate a multi-scale time series feature tensor; Performing a pooling operation on the multi-scale time series feature tensor through the global average pooling layer to generate a time series global feature vector; The temporal global feature vector is subjected to feature extraction through the bidirectional long short-term memory network layer to generate the temporal semantic feature vector.

7. The warehouse anomaly detection and intelligent early warning method according to claim 6 is characterized in that: The multi-level temporal feature extraction module includes a plurality of sequentially connected temporal feature extraction units, and the multi-level temporal feature extraction module performs a convolution operation on the input target time data to generate a multi-scale temporal feature tensor, including: The target time data is input into a plurality of sequentially connected time series feature extraction units for feature extraction, so as to generate a multi-scale time series feature tensor containing multi-level time series patterns.

8. The warehouse anomaly detection and intelligent early warning method according to claim 7 is characterized in that: The temporal feature extraction unit includes a convolutional layer, a maximum pooling layer, and a Dropout module connected in sequence. The target time data is input into the plurality of temporal feature extraction units connected in sequence for feature extraction to generate a multi-scale temporal feature tensor containing a multi-level temporal pattern, including: Performing primary time series feature extraction on the target time data through the first convolutional layer to generate a first feature tensor; Downsampling the first feature tensor in the time dimension through the maximum pooling layer to output a second feature tensor; The Dropout module is used to perform enhanced regularization processing on the second feature tensor with a preset probability to generate a multi-scale time series feature tensor containing multi-level time series patterns.

9. The warehouse anomaly detection and intelligent early warning method according to claim 4 is characterized in that: The multimodal joint decision layer includes an attention layer and a dual-branch decision layer connected in sequence. Based on the global feature vector and the temporal semantic feature vector, the multimodal joint decision layer generates a hierarchical alarm signal of storage anomaly and triggers a response action, including: fusing the global feature vector and the temporal semantic feature vector through the attention layer to generate a joint feature vector; The dual-branch decision layer performs fault identification and severity prediction on the joint feature vector, generates a graded alarm signal of storage anomaly, and triggers a response action.

10. A warehouse anomaly detection and intelligent early warning system, characterized in that: include: A first data processing module is configured to pre-process the received storage historical spatiotemporal data to generate storage training data; A model training module is used to input the warehouse training data into a preset initial anomaly detection model for training to generate a dual-path anomaly detection model, wherein the dual-path anomaly detection model includes a spatiotemporal feature extraction layer and a multimodal joint decision layer; The second data processing module responds to the warehouse anomaly detection request, obtains the warehouse real-time spatiotemporal data and performs preprocessing to generate target spatiotemporal data; An extraction module, configured to perform feature extraction based on the target spatial data through the spatiotemporal feature extraction layer to generate a spatiotemporal feature vector; A decision module is used to generate a graded alarm signal of storage anomaly based on the spatiotemporal feature vector through the multimodal joint decision layer and trigger a response action.

Images