Temperature prediction method and system for multi-mode AIGC cold-chain logistics monitoring platform

Abstract

The invention relates to the technical field of cold-chain logistics monitoring, and discloses a temperature prediction method and system for a multi-modal AIGC cold-chain logistics monitoring platform, and the method comprises the steps: carrying out the multi-modal data collection and data fusion of a cold-chain logistics transportation process based on a cold-chain logistics monitoring platform, and obtaining a multi-modal fusion data matrix; inputting an AIGC model to carry out feature extraction and inter-modal feature weight calculation to obtain a multi-modal feature importance vector; performing cross-modal weighted fusion calculation and associated feature map construction to obtain a temperature state feature sequence; inputting a multi-modal deep belief network to carry out probability distribution modeling to obtain a temperature prediction probability distribution sequence; performing time sequence characteristic analysis to obtain a temperature prediction interval sequence; and performing dynamic threshold judgment to obtain a multi-stage temperature early warning trigger signal. According to the invention, the temperature prediction accuracy of the cold-chain logistics monitoring platform is improved.

Description

Technical Field

[0001] The present invention relates to the field of cold chain logistics monitoring technology, and in particular to a temperature prediction method and system for a multi-modal AIGC cold chain logistics monitoring platform. Background Art

[0002] With the rapid development of the cold chain logistics industry, temperature control has become a key factor in ensuring product quality. Traditional cold chain logistics monitoring systems mainly rely on single temperature sensor data and lack the ability to comprehensively analyze multi-source data, resulting in low temperature prediction accuracy and failure to promptly detect potential temperature anomaly risks. In addition, existing monitoring systems generally have problems such as insufficient data fusion and insufficient correlation analysis between modalities, making it difficult to effectively use auxiliary information such as images and door status to accurately predict temperature change trends.

[0003] At present, the temperature prediction methods for cold chain logistics mainly use statistical models or simple machine learning algorithms. These methods have limited feature extraction capabilities and strong certainty of prediction results when processing multimodal data. Especially in complex transportation environments, a single model is difficult to accurately capture the dynamic characteristics of temperature changes, and lacks uncertainty assessment of prediction results, which can easily lead to problems such as untimely warnings or high false alarm rates. Summary of the invention

[0004] The present invention provides a temperature prediction method and system for a multi-modal AIGC cold chain logistics monitoring platform, and the present invention improves the temperature prediction accuracy of the cold chain logistics monitoring platform.

[0005] In a first aspect, the present invention provides a temperature prediction method for a multimodal AIGC cold chain logistics monitoring platform, the temperature prediction method for the multimodal AIGC cold chain logistics monitoring platform comprising: Based on the cold chain logistics monitoring platform, multimodal data collection and data fusion are carried out on the cold chain logistics transportation process to obtain a multimodal fusion data matrix; Inputting the multimodal fusion data matrix into the AIGC model to perform feature extraction and inter-modal feature weight calculation to obtain a multimodal feature importance vector; Performing cross-modal weighted fusion calculation and associated feature graph construction on the multimodal feature importance vector and the multimodal fusion data matrix to obtain a temperature state feature sequence; Inputting the temperature state feature sequence into a multimodal deep belief network for probability distribution modeling to obtain a temperature prediction probability distribution sequence; Performing time series feature analysis on the temperature prediction probability distribution sequence to obtain a temperature prediction interval sequence; Dynamic threshold determination is performed on the temperature prediction interval sequence to obtain a multi-level temperature warning trigger signal.

[0006] In a second aspect, the present invention provides a temperature prediction system for a multimodal AIGC cold chain logistics monitoring platform, the temperature prediction system for the multimodal AIGC cold chain logistics monitoring platform comprising: The acquisition module is used to collect and fuse multimodal data of the cold chain logistics transportation process based on the cold chain logistics monitoring platform to obtain a multimodal fusion data matrix; A calculation module, used for inputting the multimodal fusion data matrix into the AIGC model to perform feature extraction and inter-modal feature weight calculation to obtain a multimodal feature importance vector; A construction module is used to perform cross-modal weighted fusion calculation and associated feature graph construction on the multimodal feature importance vector and the multimodal fusion data matrix to obtain a temperature state feature sequence; A modeling module, used for inputting the temperature state feature sequence into a multimodal deep belief network for probability distribution modeling to obtain a temperature prediction probability distribution sequence; An analysis module, used for performing time series feature analysis on the temperature prediction probability distribution sequence to obtain a temperature prediction interval sequence; The determination module is used to perform dynamic threshold determination on the temperature prediction interval sequence to obtain a multi-level temperature warning trigger signal.

[0007] The third aspect of the present invention provides a computer device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the computer device executes the temperature prediction method of the above-mentioned multimodal AIGC cold chain logistics monitoring platform.

[0008] The fourth aspect of the present invention provides a computer-readable storage medium, which stores instructions. When the computer-readable storage medium is run on a computer, it enables the computer to execute the temperature prediction method of the above-mentioned multimodal AIGC cold chain logistics monitoring platform.

[0009] In the technical solution provided by the present invention, the accuracy of temperature prediction is enhanced through multimodal data fusion and deep feature extraction of the AIGC model, and the cross-modal attention mechanism and particle swarm optimization algorithm are used to realize the adaptive weight calculation of inter-modal features, which effectively solves the heterogeneity problem of multi-source data; based on the adaptive density peak clustering algorithm, the inter-modal correlation feature map is constructed to enhance the model's recognition ability of temperature change patterns; the probability distribution modeling of the multimodal deep belief network is used to realize the quantification of the uncertainty of the temperature prediction results, and the combination of quantile regression and long short-term memory units improves the model's ability to capture temperature time series characteristics; the dynamic threshold judgment and multi-level early warning mechanism are used to reduce the false alarm rate of the system. The present invention has strong adaptability and scalability, and can support the temperature monitoring needs of different types of cold chain logistics scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0011] Figure 1 Schematic diagram of the steps of the temperature prediction method of the multi-modal AIGC cold chain logistics monitoring platform in an embodiment of the present invention; Figure 2 It is a schematic diagram of the structure of the temperature prediction system of the multi-modal AIGC cold chain logistics monitoring platform in an embodiment of the present invention; Figure 3 It is a schematic block diagram of the structure of a computer device in an embodiment of the present invention. DETAILED DESCRIPTION

[0012] An embodiment of the present invention provides a temperature prediction method and system for a multimodal AIGC cold chain logistics monitoring platform. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0013] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 , an embodiment of the temperature prediction method of the multi-modal AIGC cold chain logistics monitoring platform in the embodiment of the present invention includes: Step S1: Based on the cold chain logistics monitoring platform, multimodal data collection and data fusion are performed on the cold chain logistics transportation process to obtain a multimodal fusion data matrix; It is understandable that the execution subject of the present invention may be a temperature prediction system of the multimodal AIGC cold chain logistics monitoring platform, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking the server as the execution subject as an example.

[0014] Specifically, multimodal data collection is performed on the cold chain logistics monitoring platform, including temperature data, humidity data, door status data and image data. In the data collection stage, sensors monitor the changes of various data in the transportation process in real time. Temperature and humidity sensors collect environmental data, while door status sensors record the timing of each door opening and closing. The monitoring camera captures the image of the goods to form visual data that can be used for subsequent analysis. The temperature data and humidity data are segmented and standardized according to the cross-modal feature distribution in the time dimension. Using techniques such as Z-score standardization or Min-Max normalization, data of different dimensions are converted into comparable forms to obtain temperature and humidity feature data. The temperature and humidity feature data are cross-modally normalized to generate a normalized temperature and humidity matrix. The door status data is time-series encoded, and the opening and closing status of each door is mapped to the time series. Feature mapping is performed in combination with the timestamp of the normalized temperature and humidity matrix to obtain the time series data of the door status. At the same time, the image data is segmented in time series, and the time index is established according to the sampling interval of the normalized temperature and humidity matrix to form the image feature time series data, ensuring the temporal consistency of the image data and other modal data. The door state time series data is aligned with the normalized temperature and humidity matrix in the time dimension. At the same time, the image feature time series data is resampled to make it consistent with the time reference of the normalized temperature and humidity matrix to obtain aligned image feature data. A time reference sequence is established based on the normalized temperature and humidity matrix so that the aligned door state data and image feature data can be synchronized in time series. Through the time series synchronization process, multimodal time series features are obtained, which comprehensively reflect the temperature, humidity, door state and visual information during the transportation process. Through cross-modal feature fusion, the correlation analysis between modalities is used to integrate the data and obtain a multimodal fusion data matrix.

[0015] Step S2: input the multimodal fusion data matrix into the AIGC model to perform feature extraction and inter-modal feature weight calculation to obtain a multimodal feature importance vector; Specifically, the multimodal fusion data matrix is ​​input into the encoding layer of the AIGC model. At this stage, the encoding layer consists of a three-layer Transformer structure, with 8 attention heads set in each layer and a hidden layer dimension of 512. Through this structure, the model can capture the potential complex feature relationships in the data and obtain the initial feature representation. The initial feature representation is processed through the cross-modal attention layer. The design of the cross-modal attention layer includes a query matrix, a key matrix, and a value matrix, all of which have dimensions of 512×512. Through this structure, the model effectively interacts and fuses information between different modalities to generate cross-modal attention features that reflect the importance and relevance of each modal data in a specific context. The cross-modal attention features are input into the decoding layer of the AIGC model. The decoding layer adopts a bidirectional attention network structure, which includes 4 decoding blocks, and the output dimension of each decoding block is set to 256. The bidirectional structure can infer the current features from the past state and consider future context information to improve the prediction ability of the model and obtain a decoded feature sequence. A feature weight calculation unit is constructed based on the decoded feature sequence, and the feature dimension information is mapped to the initial position of the particle swarm. As an intelligent optimization algorithm, the particle swarm optimization algorithm can find the optimal solution by simulating the behavior of particle groups in nature. The initial weight distribution is continuously optimized through iterative calculation to more accurately reflect the importance of each feature and obtain the optimized weight vector. The optimized weight vector is input into the feature fusion layer of the AIGC model. The design of the feature fusion layer includes two fully connected layers, the output dimension of the first layer is 128, and the output dimension of the second layer is 64. Through the fusion process, the model can integrate the optimized feature information to generate a fused feature vector. In order to extract key information, the fused feature vector is processed by the attention pooling layer of the AIGC model. The pooling layer combines the self-attention mechanism and the cross-attention mechanism to effectively extract and integrate information between different modalities. After pooling, the feature dimension is compressed to 32 to obtain a multimodal feature vector. The multimodal feature vector is input into the importance evaluation layer of the AIGC model and normalized by the sigmoid activation function to obtain a multimodal feature importance vector. Through this process, the model can quantitatively evaluate the importance of features, ensuring that reasonable judgments and decisions can be made based on the weights of each feature in temperature prediction, thereby improving the accuracy and reliability of temperature prediction.

[0016] Step S3, performing cross-modal weighted fusion calculation on the multimodal feature importance vector and the multimodal fusion data matrix and constructing a related feature graph to obtain a temperature state feature sequence; Specifically, the multimodal feature importance vector and the multimodal fusion data matrix are cross-modal feature decomposition, and the shared features between modalities and modality-specific features are extracted from the fusion data. Through decomposition, the common features and unique features of each modal data are identified. A cross-modal weighted matrix is ​​constructed based on the shared features between modalities and modality-specific features, which reflects the weight coefficients between different modalities. In the weighted operation process, combined with the additive attention mechanism, the modal interaction features are effectively calculated by applying the weight coefficients between modalities to obtain the weighted fusion features. Based on the weighted fusion features, the local density of data points is calculated to obtain the local density distribution, which reveals the concentration of data points in a specific area. According to the local density distribution, the minimum distance of high-density points is calculated, and the shortest path between different density points is calculated through the distance matrix to obtain the density distance matrix. Through the relationship between density and distance, the mutual connection and distribution characteristics between data points are analyzed. Peak detection is performed on the density distance matrix, and the number of cluster centers is determined by the product of local density and minimum distance. In this way, the density peak points in the data are accurately identified to form a density peak point set. Based on the peak points, a feature association graph is constructed, in which the attribution relationship of data points is calculated by relative density ratio and Euclidean distance. By quantifying the relationship between data points, the model can understand the interaction between features. The temperature state information in the association feature graph is arranged in time series. The temperature state changes dynamically, and the data is sorted in time series to capture the temperature change trend. The temperature state information is effectively organized into a temperature state feature sequence by serialization reconstruction through the preset time window.

[0017] Step S4, inputting the temperature state feature sequence into the multimodal deep belief network for probability distribution modeling to obtain a temperature prediction probability distribution sequence; Specifically, the temperature state feature sequence is input into the multimodal deep belief network. The structure of the network consists of three parts: feature encoder, cross-modal fusion and probability generator. The feature encoder adopts a stacked restricted Boltzmann machine structure, which can effectively extract the deep features of the data through multi-layer nonlinear transformation. The cross-modal fusion is based on a contrastive learning structure. By constructing positive and negative sample pairs, the model's understanding of the relationship between features is improved. The probability generator uses a generative adversarial network structure and uses adversarial training to enhance the accuracy and robustness of temperature prediction. The temperature state feature sequence is input into the feature encoder. The feature encoder consists of three layers of restricted Boltzmann machines, in which the number of visible units in the first layer is 256, the number of hidden units is 128, the number of hidden units in the second layer is 64, and the number of hidden units in the third layer is 32. Each layer uses a sigmoid activation function to perform nonlinear transformation. Through this process, the model can compress the input temperature state feature sequence into a deep feature representation, which represents more abstract concepts and patterns in the data. The deep feature representation is contrastively learned through a cross-modal fusion. The cross-modal fusion unit contains feature extraction units for positive and negative sample pairs. Each unit consists of two fully connected layers. The output dimension of the first layer is 64, and the output dimension of the second layer is 32. The ReLU activation function is used. Through contrastive learning, the model's ability to distinguish features between different modalities is improved, and a contrast feature vector is obtained. The contrast feature vector is input into the generative network of the probability generator. The generative network contains four fully connected layers, and the output dimensions of each layer are 128, 256, 512, and 256, respectively. Between each layer, the LeakyReLU activation function and batch normalization layer are used to improve the stability and efficiency of training. Through this process, the generative network can transform the contrast feature vector into a sequence of generated samples, which simulates the potential distribution of temperature states. The sample sequence is input into the discriminative network of the probability generator. The discriminative network consists of three convolutional layers with a convolution kernel size of 3×3 and the number of channels of 64, 128, and 256, respectively. Each layer is followed by a maximum pooling layer to reduce the feature dimension and improve the abstractness of the feature. The discrimination result is output through the fully connected layer to obtain the adversarial training loss. This loss quantifies the gap between the generated samples and the real samples. Based on the adversarial training loss, the generative network and the discriminative network will be optimized alternately. Through adversarial training, the generative network continuously adjusts the quality of the generated samples to deceive the discriminative network, while the discriminative network continuously improves its own discrimination ability and promotes the learning process of the entire model. After several rounds of alternating optimization, an optimized probabilistic generative model is obtained. The output of the optimized probabilistic generative model is converted to a probability distribution through a kernel density function. This process is achieved through a kernel density function. The kernel density estimation effectively converts the discrete values ​​of the generated samples into a continuous probability distribution to obtain a temperature probability distribution. The temperature probability distribution is arranged in chronological order to form a temperature prediction probability distribution sequence.This sequence can provide real-time temperature prediction for subsequent decision support systems, ensuring that the temperature remains within a safe range during cold chain logistics transportation and minimizing losses caused by temperature fluctuations.

[0018] Step S5, performing time series feature analysis on the temperature prediction probability distribution sequence to obtain a temperature prediction interval sequence; Specifically, the time series decomposition of the temperature prediction probability distribution sequence is performed through the quantile calculation unit to extract the time series quantile data in the sequence. The time series quantile data can provide detailed information about the temperature change and reveal the relative position of the temperature at different time points. The time series quantile data is input into the long short-term memory unit (LSTM). The dimensions of the input gate, forget gate and output gate of the LSTM unit are all 128, so that the model can effectively learn and memorize the long-term dependencies in the time series. Through this step, the model can output a time series state sequence that reflects the trend of temperature changes in the time dimension. The cross-modal attention unit is used to perform feature fusion on the time series state sequence. The design of the cross-modal attention unit includes 4 attention heads, and the output dimension of each head is 32. The multi-head attention mechanism can process different feature subsets in parallel and enhance the model's ability to capture information between modalities. In this way, the generated attention feature sequence reflects the importance of the temperature state at different time points. Feature mapping dimensionality reduction is performed on the attention feature sequence. High-dimensional data is compressed to a smaller dimension so that subsequent calculations are more efficient and easy to process. Through this process, a mapping feature vector is obtained. Perform numerical transformation on the mapping feature vector, calculate the temperature values ​​corresponding to different quantiles, and obtain multi-quantile temperature data. Perform temperature interval boundary prediction on multi-quantile temperature data. Generate interval prediction results based on the calculated temperature quantile information. Test the interval prediction results to ensure that the obtained interval prediction results are valid in practical applications. During the test process, the validity of the prediction results is determined according to certain standards or thresholds, and unreasonable predictions are excluded. Reconstruct the valid temperature intervals in chronological order. To ensure the smoothness and continuity of the data, connect the discrete prediction points through linear interpolation, fill the gaps between the discrete points, and generate a more continuous temperature prediction interval sequence.

[0019] Step S6: Perform dynamic threshold determination on the temperature prediction interval sequence to obtain a multi-level temperature warning trigger signal.

[0020] Specifically, the interval width of the temperature prediction interval sequence is calculated. The relative interval width value is obtained by dividing the difference between the upper and lower limits by the median, and the amplitude of the temperature fluctuation is quantified to form an interval width sequence. Based on the interval width sequence, a dynamic threshold calculation matrix is ​​constructed to smooth the interval sequence, thereby eliminating the noise caused by data fluctuations and improving the accuracy of subsequent analysis. Outlier detection is performed on the smoothed interval sequence. By calculating the mean and standard deviation of the sequence, a dynamic fluctuation range is established to obtain the fluctuation threshold boundary. This fluctuation threshold boundary is the key basis for judging whether the temperature prediction interval is abnormal. On this basis, the fluctuation threshold boundary is numerically mapped with the temperature control standard of cold chain logistics to obtain the temperature control boundary. The temperature control boundary is graded, and the model needs to set multi-level warning thresholds, which help to grade the risk of temperature changes. Based on the multi-level warning threshold, the temperature prediction interval is detected for abnormality. By calculating the overlap between the prediction interval and the warning threshold, the abnormality degree value is obtained. The relationship between the temperature change and the warning threshold is quantified, which can intuitively reflect the abnormality of the temperature state. The abnormality degree value is input into the warning level classifier. Through the weighted voting method, the classifier can calculate the probability of different warning levels, ensure that the influence of each warning level is reasonably evaluated, and obtain the warning level sequence. Prioritize the warning level sequence. The sorting process combines the abnormality degree value and duration for comprehensive scoring to ensure that high-priority warning signals can be responded to in a timely manner. Through comprehensive analysis, a multi-level temperature warning trigger signal can be effectively formed, thereby establishing a dynamic temperature monitoring system in the cold chain logistics process. This system can not only reflect temperature changes in real time, but also guide relevant personnel to take appropriate actions according to different warning levels to ensure that the goods are kept within a safe temperature range during transportation, minimizing the risks and losses caused by temperature fluctuations.

[0021] In the embodiments of the present invention, the accuracy of temperature prediction is enhanced through multimodal data fusion and deep feature extraction of the AIGC model. The cross-modal attention mechanism and the particle swarm optimization algorithm are used to realize the adaptive weight calculation of inter-modal features, which effectively solves the heterogeneity problem of multi-source data. The inter-modal correlation feature map is constructed based on the adaptive density peak clustering algorithm, which enhances the model's recognition ability of temperature change patterns. The probability distribution modeling of the multimodal deep belief network is used to realize the quantification of the uncertainty of the temperature prediction results. The combination of quantile regression and long short-term memory units improves the model's ability to capture temperature time series characteristics. The dynamic threshold judgment and multi-level early warning mechanism are used to reduce the false alarm rate of the system. The present invention has strong adaptability and scalability, and can support the temperature monitoring needs of different types of cold chain logistics scenarios.

[0022] In a specific embodiment, the process of executing step S1 may specifically include the following steps: Based on the cold chain logistics monitoring platform, the temperature data, humidity data, door status data and image data of the cold chain logistics transportation process are collected; The temperature data and humidity data are segmented, and data normalization is performed according to the cross-modal feature distribution in the time dimension to obtain temperature and humidity feature data. The temperature and humidity feature data are cross-modally normalized to obtain a normalized temperature and humidity matrix. The door state data is time-series encoded, and feature mapping is performed in combination with the timestamp of the normalized temperature and humidity matrix to obtain the door state time series data. The image data is time-series segmented, and a time index is established according to the sampling interval of the normalized temperature and humidity matrix to obtain the image feature time series data. Align the door state time series data with the normalized temperature and humidity matrix in the time dimension to obtain aligned door state data, and resample the image feature time series data to obtain aligned image feature data; A time reference sequence is established based on the normalized temperature and humidity matrix. The alignment gate state data and the alignment image feature data are synchronized to obtain multimodal time series features. Cross-modal feature fusion is performed on the multimodal time series features. Data integration is performed through inter-modal correlation analysis to obtain a multimodal fusion data matrix.

[0023] Specifically, based on the cold chain logistics monitoring platform, temperature data, humidity data, door status data and image data are collected during the cold chain logistics transportation process. Temperature data is usually stored in the form of a time series, indicating the temperature readings at different time points during the transportation process. The same is true for humidity data, reflecting changes in ambient humidity. Door status data indicates the time information of the storage and access status of goods during cold chain transportation, and image data provides visual environmental monitoring information, such as the internal conditions of the carriage or the opening of the door. The temperature data and humidity data are segmented and divided according to the time dimension. By analyzing the cross-modal feature distribution of the time dimension, the change patterns in different time periods can be effectively identified. To ensure the accuracy of subsequent analysis, data standardization is performed using the formula: ; in, is the original data, is the mean of the data, is the standard deviation of the data, and the result is is the normalized data. The temperature and humidity data are processed into relative standard values ​​so that the subsequent analysis is not affected by different dimensions, and the temperature and humidity feature data are obtained. These feature data are cross-modal normalized and calculated to generate a normalized temperature and humidity matrix. The door state data is time-series encoded to convert the original door state information into a form that can be used for time series analysis. Combined with the timestamp of the normalized temperature and humidity matrix, the door state time series data is obtained through feature mapping technology. For example, assuming that the door state data is "open" and "closed", it is converted into a binary sequence through the timestamp, for example, 1 represents "open" and 0 represents "closed". Similarly, the image data is time-series segmented, and a time index is established according to the sampling interval of the normalized temperature and humidity matrix to obtain the image feature time series data. This index ensures that each frame of the image corresponds to the temperature and humidity data, maintaining the time consistency of the data. The door state time series data is aligned with the normalized temperature and humidity matrix in the time dimension to ensure that the time labels of the two are consistent, and the aligned door state data is obtained. At the same time, the image feature time series data is resampled to ensure its temporal consistency with the temperature, humidity and door status data, and to generate aligned image feature data. On this basis, a time reference sequence is established based on the normalized temperature and humidity matrix to synchronize different types of data. Through this reference sequence, the aligned door status data and aligned image feature data are synchronized in time to obtain multimodal time series features. Multimodal time series features include information on temperature, humidity, door status and image data. Cross-modal feature fusion is performed on multimodal time series features, and data integration is performed through correlation analysis between modalities. This process is achieved through the following formula: ; in, is the fused feature, is the weight of the mode, are the feature vectors of different modes. By adjusting the weights between the modes, a certain feature occupies a more important position in the final fusion result, and finally a multimodal fusion data matrix is ​​obtained. For example, when the temperature data of cold chain transportation detects an abnormal increase, the image data shows that the goods are stored incorrectly or the carriage is not closed during transportation, and the door status data reflects the frequent opening and closing status. Through the above process, the data of different modes are integrated together for comprehensive analysis, so as to better monitor and manage the temperature and environmental conditions in the cold chain logistics process.

[0024] In a specific embodiment, the process of executing step S2 may specifically include the following steps: The multimodal fusion data matrix is ​​input into the encoding layer of the AIGC model. The encoding layer contains a three-layer Transformer structure, each layer has 8 attention heads, and the hidden layer dimension is 512 to obtain the initial feature representation; The initial feature representation is processed through a cross-modal attention layer, which contains a query matrix, a key matrix, and a value matrix. The matrix dimensions are all 512×512, and the cross-modal attention features are obtained. The cross-modal attention features are input into the decoding layer of the AIGC model. The decoding layer adopts a bidirectional attention network structure, which contains 4 decoding blocks. The output dimension of each decoding block is 256, and a decoding feature sequence is obtained. Based on the decoded feature sequence, a feature weight calculation unit is constructed to map the feature dimension information to the initial position of the particle swarm to obtain the initial weight distribution, and the initial weight distribution is iteratively calculated through the particle swarm optimization algorithm to obtain the optimized weight vector; The optimized weight vector is input into the feature fusion layer of the AIGC model. The feature fusion layer contains two fully connected layers. The output dimension of the first layer is 128 and the output dimension of the second layer is 64, and the fused feature vector is obtained. The fused feature vector is processed by the attention pooling layer of the AIGC model. The attention pooling layer includes a self-attention mechanism and a cross-attention mechanism. The dimension after pooling is 32, and a multimodal feature vector is obtained. The multimodal feature vector is input into the importance evaluation layer of the AIGC model and normalized by the sigmoid activation function to obtain a multimodal feature importance vector.

[0025] Specifically, the multimodal fusion data matrix is ​​input into the encoding layer of the AIGC model, which contains a three-layer Transformer structure. Each layer of Transformer consists of multiple attention heads. For example, 8 attention heads are set to process multiple information in parallel at the same level, enhancing the model's ability to capture the relationship between different modalities. The hidden layer dimension is set to 512, indicating the feature representation capability of each attention head. Through the setting, the initial feature representation is expressed by the following formula: ; in, is the initial feature representation, is the input multimodal fusion data matrix. By stacking multiple layers of Transformer, the model can capture richer feature information. The initial feature representation is processed by the cross-modal attention layer. In this layer, the cross-modal attention layer consists of a query matrix, a key matrix, and a value matrix, and the matrix dimensions are all 512×512. The query matrix is ​​used to extract the key parts of the features, the key matrix is ​​used to identify the importance of the features, and the value matrix contains the information that needs to be aggregated. This process is expressed by the following formula: ; in, , , denote the query matrix, key matrix and value matrix respectively, is the dimension of the key, is the obtained cross-modal attention feature. By calculating the attention weight, the model can extract the most relevant information based on the weighted features of the relationship between different modalities and generate cross-modal attention features. The cross-modal attention features are input into the decoding layer of the AIGC model. The decoding layer adopts a bidirectional attention network structure, which contains 4 decoding blocks, and the output dimension of each decoding block is 256. This design allows the model to pay attention to the current input information when decoding, and also comprehensively consider the past information, thereby improving the accuracy of decoding. The output of the decoded feature sequence is expressed as: ; in, To decode the feature sequence, after processing by 4 decoding blocks, the model can extract high-dimensional semantic information from the cross-modal features. Based on the decoded feature sequence, a feature weight calculation unit is constructed. The feature dimension information is mapped to the initial position of the particle swarm, and the particle swarm optimization is performed based on this. The particle swarm optimization algorithm is an optimization technology based on swarm intelligence. By iteratively calculating the initial weight distribution, the optimized weight vector is obtained. The specific update process is expressed as: ; ; in, is the particle speed, is the position of the particle, is the inertia weight, and are learning factors, and is a random number, is the historical optimal position of the particle, is the global optimal position, is the number of iterations. The obtained optimized weight vector can more effectively reflect the importance of each feature. The optimized weights are input into the feature fusion layer of the AIGC model. The feature fusion layer contains two fully connected layers. The output dimension of the first layer is set to 128, and the output dimension of the second layer is set to 64. The features are further compressed and extracted through the linear transformation of the fully connected layer. The fused feature vector is expressed as: ; in, is the fusion feature vector, and is the weight matrix of the fully connected layer. The fused feature vector is processed by the attention pooling layer of the AIGC model. The attention pooling layer combines the self-attention mechanism and the cross-attention mechanism so that the relationship between different features can be fully utilized. After pooling, the output dimension is 32, and a multimodal feature vector is obtained. This process is done by the formula: ; in, is the feature vector after pooling. This process effectively integrates information from multiple modalities to generate a compact and informative feature representation. The multimodal feature vector is input into the importance evaluation layer of the AIGC model. In this layer, the sigmoid activation function is used for normalization to obtain the multimodal feature importance vector to ensure that the influence of each feature in the model can be reasonably reflected. The calculation formula is: ; in, is the multimodal feature importance vector, is the weight matrix, is the bias, is the sigmoid function. The multimodal feature importance vector output by the model reflects the relative importance of each modal feature and provides a basis for subsequent decision-making and analysis. For example, in cold chain logistics, temperature, humidity, door status, and image data jointly affect the safety of goods. The above process model can synthesize this information and derive the importance of different features, so as to more accurately identify potential risks and anomalies during the monitoring process.

[0026] In a specific embodiment, the process of executing step S3 may specifically include the following steps: Perform cross-modal feature decomposition on the multimodal feature importance vector and the multimodal fusion data matrix to obtain inter-modal shared features and modality-specific features; Based on the shared features and modality-specific features between modalities, a cross-modal weighted matrix is ​​constructed to obtain the inter-modal weight coefficients. The multimodal fusion data matrix is ​​weighted by the inter-modal weight coefficients, and the modal interaction features are calculated in combination with the additive attention mechanism to obtain the weighted fusion features. The local density of data points is calculated based on the weighted fusion features to obtain the local density distribution, and the minimum distance of high-density points is calculated based on the local density distribution. The shortest path between points of different densities is calculated through the distance matrix to obtain the density distance matrix. Peak detection is performed on the density distance matrix. The number of cluster centers is determined by the product of local density and minimum distance to obtain the density peak point set. Based on the density peak point set, a feature association graph is constructed. The data point attribution relationship is calculated by relative density ratio and Euclidean distance to obtain an association feature graph. The temperature state information in the associated feature graph is arranged in time series and serialized and reconstructed through a preset time window to obtain a temperature state feature sequence.

[0027] Specifically, by linearly transforming the multimodal feature importance vector and the multimodal fusion data matrix, we can obtain the shared feature matrix and the modality-specific feature matrix. The shared feature matrix is ​​set to and the mode-specific feature matrix is , then it is expressed as: ; in, is the multimodal feature importance vector, and are weight matrices for shared features and modality-specific features, respectively. This approach allows the model to extract features that are common to all modalities from the fused data while retaining the unique information of each modality. A cross-modal weighting matrix is ​​constructed based on the shared features and modality-specific features between modalities to obtain the weight coefficients between modalities. The weight matrix is ​​set to , calculated by the following formula: ; By combining shared features and modality-specific features, we ensure that the weights between different modalities can reflect their importance in data fusion. We perform weighted operations on the multimodal fusion data matrix, and calculate the modality interaction features in combination with the additive attention mechanism to obtain the weighted fusion features, which are expressed by the following formula: ; in, is the multimodal fusion data matrix, is a weighted fusion feature. The application of the additive attention mechanism allows each modality to take into account the influence of other modalities when outputting features, better reflecting the relevance of different data sources. Based on the weighted fusion feature, the local density of each data point is calculated to obtain the local density distribution. The local density calculation is implemented by the following formula: ; in, Indicate point The local density of Yes The neighbor set of is a kernel function, such as a Gaussian kernel. In this way, the local density distribution of each data point is obtained, and then the high-density area is identified. According to the local density distribution, the minimum distance of the high-density points is calculated to form a distance matrix. The minimum distance calculation is expressed as: ; This distance represents the distance at a specific point The distance between the nearest point and its neighbors. Through these distance calculations, a density distance matrix is ​​constructed. , which can reveal the relative position relationship between each data point. By finding the local maximum value, the density distance matrix is ​​peak detected, and the number of cluster centers is determined by the product of the local density and the minimum distance, and the density peak point set is obtained. It is expressed as: ; in, is the cluster center set. Based on the density peak point set, a feature association graph is constructed, and the attribution relationship of the data points is calculated by the relative density ratio and Euclidean distance. This process is expressed by the following formula: ; in, Indicate point and Point The similarity between is a parameter that controls the range of similarity. Through the association graph, the relationship between data points is clarified, and analysis is performed based on these relationships. The temperature state information in the association feature graph is arranged in time series, and serialized and reconstructed through a preset time window to obtain a temperature state feature sequence. The association feature graph is indexed in time series, and the data is divided into corresponding sequences according to the length of the time window. It can be expressed as: ; in, is the characteristic sequence of temperature state in the time window, is the length of the window. The construction of the temperature state feature sequence can ensure that it reflects the dynamic changes in time.

[0028] In a specific embodiment, the process of executing step S4 may specifically include the following steps: The temperature state feature sequence is input into a multimodal deep belief network, which includes a feature encoder, a cross-modal fusion device and a probability generator. The feature encoder adopts a stacked restricted Boltzmann machine structure, the cross-modal fusion device adopts a contrastive learning structure, and the probability generator adopts a generative adversarial network structure. The temperature state feature sequence is input into the feature encoder. The feature encoder contains three layers of restricted Boltzmann machines. The number of visible units in the first layer is 256, the number of hidden units is 128, the number of hidden units in the second layer is 64, and the number of hidden units in the third layer is 32. Each layer uses a sigmoid activation function to obtain a deep feature representation. The deep feature representation is contrastively learned through a cross-modal fuser. The cross-modal fuser contains feature extraction units for positive sample pairs and negative sample pairs. Each unit contains two fully connected layers. The output dimension of the first layer is 64, and the output dimension of the second layer is 32. The ReLU activation function is used to obtain the contrast feature vector. The contrast feature vector is input into the generative network of the probability generator. The generative network contains four fully connected layers with output dimensions of 128, 256, 512, and 256 respectively. Each layer uses the LeakyReLU activation function and batch normalization layer to obtain the generated sample sequence. The generated sample sequence is input into the discriminant network of the probability generator. The discriminant network contains three convolutional layers with a convolution kernel size of 3×3 and the number of channels of 64, 128, and 256 respectively. Each layer is followed by a maximum pooling layer, and finally the discrimination result is output through a fully connected layer to obtain the adversarial training loss. Based on the adversarial training loss, the generative network and the discriminative network are alternately optimized to obtain the optimized probability generation model. The output of the optimized probability generation model is converted into a probability distribution through a kernel density function to obtain a temperature probability distribution, and the temperature probability distribution is arranged in chronological order to obtain a temperature prediction probability distribution sequence.

[0029] Specifically, the temperature state feature sequence is input into the multimodal deep belief network, which includes a feature encoder, a cross-modal fusion unit, and a probability generator. The feature encoder adopts a stacked restricted Boltzmann machine structure to extract latent features by converting the input temperature state feature sequence into a deep feature representation. The structure of the feature encoder is expressed as: ; ; ; in, is the input temperature state characteristic sequence, and They are The weights and biases of the layers, For the The hidden feature representation of the layer Represents the sigmoid activation function. The feature encoder consists of three layers of restricted Boltzmann machines, where the number of visible units in the first layer is 256, the number of hidden units is 128, the number of hidden units in the second layer is 64, and the number of hidden units in the third layer is 32. After this process, the input temperature state feature sequence is converted into a deep feature representation. The deep feature representation is input into the cross-modal fusion device for contrastive learning. The cross-modal fusion device is designed to contain feature extraction units for positive sample pairs and negative sample pairs. Each unit consists of two fully connected layers, where the output dimension of the first layer is 64, the output dimension of the second layer is 32, and the ReLU activation function is used. The goal of contrastive learning is to maximize the similarity between positive sample pairs and minimize the similarity between negative sample pairs. The specific calculation formula is as follows: ; in, represents the loss function, is the set of positive sample pairs, is the set of negative sample pairs, is the feature vector obtained by the feature extraction unit, A hyperparameter to balance the impact of positive and negative samples. Through contrastive learning, the features related to the task can be better extracted. The contrast feature vector is input into the generative network of the probability generator. The generative network consists of four fully connected layers with output dimensions of 128, 256, 512 and 256 respectively. Each layer uses the LeakyReLU activation function and batch normalization layer. The calculation of the generative network is expressed as: LeakyReLU (output 128); LeakyReLU (output 256); LeakyReLU (output 512); LeakyReLU (output 256); The generator network outputs the generated sample sequence, which is then input into the discriminant network for adversarial training. The discriminant network consists of three convolutional layers, each with a convolution kernel size of 3×3 and 64, 128, and 256 channels, respectively. Each layer is followed by a maximum pooling layer, and the discriminant result is output through a fully connected layer. The calculation formula is: ; in, represents the output of the discriminant network, and are the weights and biases of the discriminant network respectively. Through adversarial training, the authenticity of the generated samples is evaluated and the adversarial training loss is calculated: ; in, represents the loss function of the discriminant network, is the real data distribution, is the distribution of generated samples. Through back propagation, the weights of the generative network and the discriminative network will be optimized alternately to achieve the optimized probability generation model. The output of the optimized probability generation model is converted to a probability distribution through the kernel density function. The formula for kernel density estimation is: ; in, Indicates at point The probability density estimate at is the kernel function, is the bandwidth parameter. By arranging the temperature probability distribution in chronological order, the obtained temperature prediction probability distribution sequence is expressed as: ; in, For at time point The temperature probability distribution.

[0030] In a specific embodiment, the process of executing step S5 may specifically include the following steps: The temperature prediction probability distribution sequence is decomposed in time series through the quantile calculation unit to obtain the time series quantile data, and the time series quantile data is input into the long short-term memory unit. The input gate dimension of the long short-term memory unit is 128, the forget gate dimension is 128, and the output gate dimension is 128 to obtain the time series state sequence; The feature fusion of the temporal state sequence is performed through the cross-modal attention unit. The cross-modal attention unit contains 4 attention heads, and the output dimension of each head is 32, and the attention feature sequence is obtained; Perform feature mapping dimensionality reduction on the attention feature sequence to obtain a mapping feature vector, perform numerical transformation on the mapping feature vector, calculate the temperature values ​​corresponding to different quantiles, and obtain multi-quantile temperature data; The temperature interval boundaries are predicted for multi-quantile temperature data to obtain interval prediction results, and the interval prediction results are tested to obtain valid temperature intervals. The valid temperature intervals are reconstructed in chronological order, and discrete prediction points are connected by linear interpolation to obtain a temperature prediction interval sequence.

[0031] Specifically, the sequence is decomposed into time series by the quantile calculation unit to extract the time series quantile data. Assume that the temperature prediction probability distribution sequence is , by calculating at each time point The quantile is calculated by the following formula: ; in, Indicates Percentile points, represents the temperature prediction value after sorting, is the total number of predicted values, and is the number of quantiles required. Input the time series quantile data into the long short-term memory (LSTM) unit. The LSTM unit is specially designed to process time series data and can effectively capture long-term dependencies. The calculation process of LSTM includes input gate, forget gate and output gate. Assume that the input data is , the hidden state is , the calculation of the LSTM unit can be expressed as: ; ; ; ; ; ; in, , and are the activation values ​​of the input gate, forget gate, and output gate, respectively. is the cell state of the current unit, is the hidden state. The dimension of the input gate is set to 128, and the dimensions of the forget gate and output gate are also 128, so that LSTM can capture the dynamic changes of time series data and finally generate a time series state sequence . The feature fusion of the temporal state sequence is performed through the cross-modal attention unit. The design of the cross-modal attention mechanism enables the model to focus on different parts of the input more effectively. Assuming that the cross-modal attention unit contains 4 attention heads and the output dimension of each head is 32, the following formula is used to calculate the attention feature sequence: ; in, It is Attention head at time point The output, , and denote query, key, and value matrices respectively, is the dimension of the key. After concatenating the outputs of all heads together, we get the attention feature sequence Perform feature mapping dimensionality reduction on the attention feature sequence to obtain the mapping feature vector , and perform numerical transformation on these eigenvectors to calculate the temperature values ​​corresponding to different quantiles. Assume that the mapping eigenvector is , which is converted into a temperature value using the following formula: ; in, is a mapping function that converts the feature vector into the actual temperature value. In this way, multi-quantile temperature data is obtained . Predict the temperature interval boundaries for multi-quantile temperature data. Use regression models or time series analysis methods. Assuming that the temperature interval boundaries are predicted by a linear regression model, it can be expressed as follows: ; in, Indicated in The predicted temperature at a time point, is a model parameter, thereby generating interval prediction results. The prediction results are tested to determine the valid temperature interval. The determination of the valid temperature interval is based on the set temperature range, for example, the safe temperature interval is set to [2°C, 8°C]. If the predicted temperature interval intersects this range, the interval is considered valid. The valid temperature interval will be reconstructed in time, the intervals will be arranged in chronological order, and the discrete prediction points will be connected using linear interpolation to obtain the final temperature prediction interval sequence. The linear interpolation formula is as follows: ; in, and are the predicted temperature values ​​at two adjacent time points, and is the corresponding time point. Through this step, a smooth temperature prediction interval sequence is obtained.

[0032] In a specific embodiment, the process of executing step S6 may specifically include the following steps: The interval width of the temperature prediction interval sequence is calculated, and the relative interval width value is obtained by dividing the upper and lower limit difference by the median to obtain the interval width sequence. A dynamic threshold calculation matrix is ​​constructed based on the interval width sequence to obtain a smoothed interval sequence. Outlier detection is performed on the smoothed interval sequence. The dynamic fluctuation range is established by calculating the mean and standard deviation to obtain the fluctuation threshold boundary. The fluctuation threshold boundary is numerically mapped with the cold chain logistics temperature control standard to obtain the temperature control boundary. The temperature control boundary is processed in a graded manner, and a multi-level warning threshold is set. Based on the multi-level warning threshold, anomaly detection is performed on the temperature prediction interval. The degree of anomaly is obtained by calculating the overlap between the prediction interval and the warning threshold. The abnormality degree value is input into the warning level classifier, and the probability of different warning levels is calculated by weighted voting method to obtain the warning level sequence. The warning level sequence is prioritized and a comprehensive score is given based on the abnormality degree value and duration to obtain a multi-level temperature warning trigger signal.

[0033] Specifically, the interval width of the temperature prediction interval sequence is calculated, and the upper and lower limits of the interval are defined. Assume that and ,in Indicates Time points. Interval width Calculated by the following formula: ; Divide the bin width by the median of the bin ; in For the The relative interval width of each time point, Then through and Find the median after sorting. Based on the relative interval width sequence , construct dynamic threshold calculation matrix , which is used to smooth the interval sequence. Dynamic threshold calculation is achieved by performing sliding average calculation on historical data, for example: ; in, is the sliding window size, For the The dynamic threshold at each moment. Perform outlier detection on the smoothed interval sequence. By calculating the mean and standard deviation, establish the dynamic fluctuation range, and obtain the fluctuation threshold boundary: ; ; in is a constant used to adjust the sensitivity of the fluctuation range. The fluctuation threshold boundary is numerically mapped with the cold chain logistics temperature control standard to obtain the temperature control boundary For example, define a standard temperature range , mapped by linear interpolation or direct comparison: ; Grading temperature control boundaries and setting multi-level warning thresholds , which is achieved by setting threshold intervals, such as dividing the temperature interval into low temperature warning, normal range and high temperature warning. Based on the multi-level warning threshold, anomaly detection is performed on the temperature prediction interval, and the overlap between the prediction interval and the warning threshold is calculated. , the overlap is calculated by the following formula: ; if If it is greater than zero, it means there is overlap, that is, there is a possibility of abnormality. The abnormality value is input into the warning level classifier, which uses the weighted voting method to calculate the probability of different frequency warning levels and obtain the warning level sequence. The weight of each warning level The final warning level is associated with its historical abnormality level. The final warning level is scored by combining the abnormality level value and duration to form a multi-level temperature warning trigger signal. : ; in Indicates The moment corresponding to The abnormality level of each warning level, is the number of warning levels.

[0034] The above describes the temperature prediction method of the multi-modal AIGC cold chain logistics monitoring platform in the embodiment of the present invention. The following describes the temperature prediction system of the multi-modal AIGC cold chain logistics monitoring platform in the embodiment of the present invention. Figure 2 In one embodiment of the present invention, a temperature prediction system of a multi-modal AIGC cold chain logistics monitoring platform includes: The acquisition module is used to collect and fuse multimodal data of the cold chain logistics transportation process based on the cold chain logistics monitoring platform to obtain a multimodal fusion data matrix; The calculation module is used to input the multimodal fusion data matrix into the AIGC model for feature extraction and inter-modal feature weight calculation to obtain the multimodal feature importance vector; A construction module is used to perform cross-modal weighted fusion calculation and associated feature graph construction on the multimodal feature importance vector and the multimodal fusion data matrix to obtain a temperature state feature sequence; A modeling module is used to input the temperature state feature sequence into a multimodal deep belief network for probability distribution modeling to obtain a temperature prediction probability distribution sequence; An analysis module is used to perform time series feature analysis on the temperature prediction probability distribution sequence to obtain a temperature prediction interval sequence; The determination module is used to perform dynamic threshold determination on the temperature prediction interval sequence to obtain a multi-level temperature warning trigger signal.

[0035] Through the collaborative cooperation of the above-mentioned components, the accuracy of temperature prediction is enhanced through multimodal data fusion and deep feature extraction of the AIGC model. The cross-modal attention mechanism and particle swarm optimization algorithm are used to realize the adaptive weight calculation of inter-modal features, which effectively solves the heterogeneity problem of multi-source data. The inter-modal correlation feature map is constructed based on the adaptive density peak clustering algorithm, which enhances the model's ability to recognize temperature change patterns. The probability distribution modeling of the multimodal deep belief network is used to realize the quantification of the uncertainty of the temperature prediction results. The combination of quantile regression and long short-term memory units improves the model's ability to capture temperature time series characteristics. The dynamic threshold judgment and multi-level early warning mechanism are used to reduce the false alarm rate of the system. The present invention has strong adaptability and scalability, and can support the temperature monitoring needs of different types of cold chain logistics scenarios.

[0036] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a display screen, an input system, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0037] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0038] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0039] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0040] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0041] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0042] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A temperature prediction method for a multi-modal AIGC cold chain logistics monitoring platform, characterized in that: The method comprises: Based on the cold chain logistics monitoring platform, multimodal data collection and data fusion are carried out on the cold chain logistics transportation process to obtain a multimodal fusion data matrix; Inputting the multimodal fusion data matrix into the AIGC model to perform feature extraction and inter-modal feature weight calculation to obtain a multimodal feature importance vector; Performing cross-modal weighted fusion calculation and associated feature graph construction on the multimodal feature importance vector and the multimodal fusion data matrix to obtain a temperature state feature sequence; Inputting the temperature state feature sequence into a multimodal deep belief network for probability distribution modeling to obtain a temperature prediction probability distribution sequence; Performing time series feature analysis on the temperature prediction probability distribution sequence to obtain a temperature prediction interval sequence; Dynamic threshold determination is performed on the temperature prediction interval sequence to obtain a multi-level temperature warning trigger signal.

2. The temperature prediction method of the multimodal AIGC cold chain logistics monitoring platform according to claim 1 is characterized in that: The cold chain logistics monitoring platform is used to collect and fuse multimodal data in the cold chain logistics transportation process to obtain a multimodal fusion data matrix, including: Based on the cold chain logistics monitoring platform, the temperature data, humidity data, door status data and image data of the cold chain logistics transportation process are collected; The temperature data and the humidity data are segmented, data normalization is performed according to the cross-modal feature distribution in the time dimension to obtain temperature and humidity feature data, and cross-modal normalization calculation is performed on the temperature and humidity feature data to obtain a normalized temperature and humidity matrix; Performing time series encoding on the door state data, performing feature mapping in combination with the timestamp of the normalized temperature and humidity matrix to obtain door state time series data, and performing time series segmentation on the image data, establishing a time index according to the sampling interval of the normalized temperature and humidity matrix, and obtaining image feature time series data; Aligning the door state time series data with the normalized temperature and humidity matrix in the time dimension to obtain aligned door state data, and resampling the image feature time series data to obtain aligned image feature data; A time reference sequence is established based on the normalized temperature and humidity matrix, the alignment gate state data and the alignment image feature data are synchronized in time series to obtain multimodal time series features, and cross-modal feature fusion is performed on the multimodal time series features. Data integration is performed through inter-modal correlation analysis to obtain a multimodal fusion data matrix.

3. The temperature prediction method of the multimodal AIGC cold chain logistics monitoring platform according to claim 2 is characterized in that: The multimodal fusion data matrix is ​​input into the AIGC model for feature extraction and inter-modal feature weight calculation to obtain a multimodal feature importance vector, including: Input the multimodal fusion data matrix into the encoding layer of the AIGC model, wherein the encoding layer includes a three-layer Transformer structure, each layer is provided with 8 attention heads, and the hidden layer dimension is 512, to obtain an initial feature representation; The initial feature representation is processed through a cross-modal attention layer, wherein the cross-modal attention layer includes a query matrix, a key matrix, and a value matrix, and the matrix dimensions are all 512×512 to obtain a cross-modal attention feature; The cross-modal attention feature is input into the decoding layer of the AIGC model, where the decoding layer adopts a bidirectional attention network structure and includes 4 decoding blocks, each of which has an output dimension of 256, to obtain a decoding feature sequence; Based on the decoded feature sequence, a feature weight calculation unit is constructed, feature dimension information is mapped to the initial position of a particle swarm, an initial weight distribution is obtained, and the initial weight distribution is iteratively calculated by a particle swarm optimization algorithm to obtain an optimized weight vector; Input the optimized weight vector into the feature fusion layer of the AIGC model, wherein the feature fusion layer includes two fully connected layers, the output dimension of the first layer is 128, and the output dimension of the second layer is 64, to obtain a fused feature vector; The fused feature vector is processed by the attention pooling layer of the AIGC model. The attention pooling layer includes a self-attention mechanism and a cross-attention mechanism. The dimension after pooling is 32 to obtain a multimodal feature vector. The multimodal feature vector is input into the importance evaluation layer of the AIGC model and normalized by the sigmoid activation function to obtain a multimodal feature importance vector.

4. The temperature prediction method of the multimodal AIGC cold chain logistics monitoring platform according to claim 3 is characterized in that: The multimodal feature importance vector and the multimodal fusion data matrix are subjected to cross-modal weighted fusion calculation and associated feature graph construction to obtain a temperature state feature sequence, including: Performing cross-modal feature decomposition on the multimodal feature importance vector and the multimodal fusion data matrix to obtain inter-modal shared features and modality-specific features; A cross-modal weighted matrix is ​​constructed based on the inter-modal shared features and the modality-specific features to obtain inter-modal weight coefficients, and a weighted operation is performed on the multimodal fusion data matrix using the inter-modal weight coefficients, and modal interaction features are calculated in combination with an additive attention mechanism to obtain weighted fusion features; Calculating the local density of data points based on the weighted fusion features to obtain local density distribution, calculating the minimum distance of high-density points based on the local density distribution, calculating the shortest path between points of different densities through the distance matrix, and obtaining a density distance matrix; Performing peak detection on the density distance matrix, determining the number of cluster centers by multiplying the local density by the minimum distance, obtaining a density peak point set, constructing a feature association graph based on the density peak point set, calculating the data point attribution relationship by relative density ratio and Euclidean distance, and obtaining an association feature graph; The temperature state information in the associated feature graph is arranged in time sequence, and is serialized and reconstructed through a preset time window to obtain a temperature state feature sequence.

5. The temperature prediction method of the multimodal AIGC cold chain logistics monitoring platform according to claim 4 is characterized in that: The step of inputting the temperature state feature sequence into a multimodal deep belief network for probability distribution modeling to obtain a temperature prediction probability distribution sequence includes: Inputting the temperature state feature sequence into a multimodal deep belief network, the multimodal deep belief network includes a feature encoder, a cross-modal fusion device and a probability generator, the feature encoder adopts a stacked restricted Boltzmann machine structure, the cross-modal fusion device adopts a contrastive learning structure, and the probability generator adopts a generative adversarial network structure; Inputting the temperature state feature sequence into the feature encoder, the feature encoder comprises three layers of restricted Boltzmann machines, the first layer has 256 visible units and 128 hidden units, the second layer has 64 hidden units, and the third layer has 32 hidden units, and each layer uses a sigmoid activation function to obtain a deep feature representation; The deep feature representation is subjected to contrastive learning by the cross-modal fuser, wherein the cross-modal fuser comprises feature extraction units for positive sample pairs and negative sample pairs, each unit comprises two fully connected layers, the output dimension of the first layer is 64, the output dimension of the second layer is 32, and a ReLU activation function is used to obtain a contrast feature vector; Input the contrast feature vector into the generation network of the probability generator, the generation network comprises four fully connected layers, the output dimensions are 128, 256, 512, 256 respectively, each layer uses a LeakyReLU activation function and a batch normalization layer, and obtains a generated sample sequence; The generated sample sequence is input into the discriminant network of the probability generator, wherein the discriminant network comprises three convolutional layers, the convolution kernel size is 3×3, the number of channels is 64, 128, and 256 respectively, each layer is followed by a maximum pooling layer, and finally the discrimination result is output through a fully connected layer to obtain an adversarial training loss, and the generation network and the discriminant network are alternately optimized based on the adversarial training loss to obtain an optimized probability generation model; The output of the optimized probability generation model is converted into a probability distribution through a kernel density function to obtain a temperature probability distribution, and the temperature probability distribution is arranged in chronological order to obtain a temperature prediction probability distribution sequence.

6. The temperature prediction method of the multimodal AIGC cold chain logistics monitoring platform according to claim 5 is characterized in that: The step of performing time series feature analysis on the temperature prediction probability distribution sequence to obtain a temperature prediction interval sequence includes: The temperature prediction probability distribution sequence is decomposed in time series by a quantile calculation unit to obtain time series quantile data, and the time series quantile data is input into a long short-term memory unit, the input gate dimension of the long short-term memory unit is 128, the forget gate dimension is 128, and the output gate dimension is 128, so as to obtain a time series state sequence; Performing feature fusion on the temporal state sequence through a cross-modal attention unit, wherein the cross-modal attention unit includes 4 attention heads, each with an output dimension of 32, to obtain an attention feature sequence; Performing feature mapping dimension reduction on the attention feature sequence to obtain a mapping feature vector, and performing numerical transformation on the mapping feature vector to calculate temperature values ​​corresponding to different quantiles to obtain multi-quantile temperature data; The temperature interval boundary prediction is performed on the multi-quantile temperature data to obtain an interval prediction result, and the interval prediction result is tested to obtain a valid temperature interval, the valid temperature interval is reconstructed in chronological order, and discrete prediction points are connected by linear interpolation to obtain a temperature prediction interval sequence.

7. The temperature prediction method of the multimodal AIGC cold chain logistics monitoring platform according to claim 6 is characterized in that: The step of performing dynamic threshold determination on the temperature prediction interval sequence to obtain a multi-level temperature warning trigger signal includes: Performing interval width calculation on the temperature prediction interval sequence, obtaining a relative interval width value by dividing the upper and lower limit difference by the median, obtaining an interval width sequence, and constructing a dynamic threshold calculation matrix based on the interval width sequence to obtain a smoothed interval sequence; Perform outlier detection on the smoothed interval sequence, establish a dynamic fluctuation range by calculating the mean and standard deviation, obtain a fluctuation threshold boundary, and numerically map the fluctuation threshold boundary to the cold chain logistics temperature control standard to obtain a temperature control boundary; The temperature control boundary is processed in a graded manner to obtain a multi-level warning threshold, and an abnormality detection is performed on the temperature prediction interval based on the multi-level warning threshold, and an abnormality degree value is obtained by calculating the overlap between the prediction interval and the warning threshold; The abnormality degree value is input into the warning level classifier, and the probability of different warning levels is calculated by weighted voting method to obtain a warning level sequence, and the warning level sequence is prioritized, and a comprehensive score is performed in combination with the abnormality degree value and duration to obtain a multi-level temperature warning trigger signal.

8. A temperature prediction system for a multi-modal AIGC cold chain logistics monitoring platform, characterized in that: Used to execute the temperature prediction method of the multimodal AIGC cold chain logistics monitoring platform according to any one of claims 1 to 7, the temperature prediction system of the multimodal AIGC cold chain logistics monitoring platform includes: The acquisition module is used to collect and fuse multimodal data of the cold chain logistics transportation process based on the cold chain logistics monitoring platform to obtain a multimodal fusion data matrix; A calculation module, used for inputting the multimodal fusion data matrix into the AIGC model to perform feature extraction and inter-modal feature weight calculation to obtain a multimodal feature importance vector; A construction module is used to perform cross-modal weighted fusion calculation and associated feature graph construction on the multimodal feature importance vector and the multimodal fusion data matrix to obtain a temperature state feature sequence; A modeling module, used for inputting the temperature state feature sequence into a multimodal deep belief network for probability distribution modeling to obtain a temperature prediction probability distribution sequence; An analysis module, used for performing time series feature analysis on the temperature prediction probability distribution sequence to obtain a temperature prediction interval sequence; The determination module is used to perform dynamic threshold determination on the temperature prediction interval sequence to obtain a multi-level temperature warning trigger signal.

9. A computer device, characterized in that: It includes a memory and a processor, the memory stores a computer program that can be run on the processor, and is characterized in that when the processor executes the computer program, it implements the temperature prediction method of the multimodal AIGC cold chain logistics monitoring platform described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the processor executes the temperature prediction method of the multimodal AIGC cold chain logistics monitoring platform as described in any one of claims 1 to 7.

Images