Fire-fighting equipment whole life cycle predictive operation and maintenance method and system

Abstract

The application provides a fire-fighting facility whole life cycle predictive operation and maintenance method and system, the method comprises collecting multi-modal operation and maintenance data such as sensor time series data, maintenance text log and structured event alarm, inputting the unified coding into the autoregressive prediction model, and storing the context key-value pair at each time step in KVCache during the inference process. In view of the problem that the cache capacity continues to grow with long sequence inference, the importance score of the key-value pair is calculated by fusing the modal weight, the entity centrality of the operation and maintenance knowledge graph and the cumulative attention value at each inference step; when the cache capacity reaches the threshold, based on the similar neighborhood clustering of the key vector, the redundant information cluster and the isolated information point with low importance are identified, the redundant cluster is compressed and the isolated point is expelled by using the abstract generation network, the cache optimization is completed, and then the failure probability, the remaining life and / or the maintenance level of the fire-fighting facility in the future period are predicted.

Description

Technical Field

[0001] This application belongs to the field of fire protection facilities, and in particular relates to a predictive operation and maintenance method and system for the entire life cycle of fire protection facilities. Background Technology

[0002] With the continuous advancement of smart city construction, fire protection facilities, as crucial infrastructure for ensuring public safety, play a vital role in building safety management through their stable operation throughout their entire lifecycle. Traditional passive, reactive maintenance and periodic inspection methods, relying primarily on manual experience and fixed cycles, struggle to identify early degradation trends in equipment in a timely manner, failing to meet the demands of modern buildings for high reliability, continuity, and refined operation and maintenance. In actual operation and maintenance scenarios, fire protection systems continuously generate massive amounts of multimodal data, including long-term sensor data from temperature, water pressure, and smoke detectors, unstructured text logs filled out by maintenance personnel, and structured event alarms automatically generated by the system. This data is characterized by diverse sources, heterogeneous structures, high dimensionality, and long time spans. Existing operation and maintenance methods struggle to fully extract and retain the deep spatiotemporal evolution characteristics within this massive, long-term data spanning the entire equipment lifecycle. This limits the accuracy of predicting equipment failure probability, remaining lifespan, and maintenance level, impacting the level of intelligent operation and maintenance of fire protection facilities.

[0003] To enhance the processing capabilities of multimodal long-sequence data, sequence prediction models based on autoregressive architectures are increasingly being applied to operational prediction scenarios. These models utilize key-value caching mechanisms during the inference phase, constructing key-value pairs from the feature sequences corresponding to historical multimodal data and storing them in a KVCache. This reduces redundant contextual computation and improves the model's efficiency in fusing historical degradation information and its inference speed. However, as the service life of fire-fighting equipment extends and input sequences accumulate, the demand for KVCache capacity increases, easily leading to increased memory or GPU memory usage and inference latency. Existing KVCache management methods typically employ fixed window truncation, first-in-first-out (FIFO), or coarse-grained eviction strategies based on a single attention threshold. When the cache reaches the threshold, some historical key-value pairs are often directly discarded. This makes it difficult to differentiate redundant information by combining feature similarity and contextual importance, and also makes it difficult to perform low-information-loss semantic compression of redundant information clusters using summarization networks. This may result in the loss of early degradation clues or long-term evolutionary features of the equipment, making it difficult to balance contextual information preservation, cache capacity release, and prediction stability with limited computing resources. Summary of the Invention

[0004] To address the problems of insufficient retention of multimodal features, KVCache bloat, easy loss of key degradation clues due to coarse-grained eviction, and limited prediction stability in the long-term operation and maintenance prediction of fire protection facilities, this invention proposes a predictive operation and maintenance method and system for the entire life cycle of fire protection facilities.

[0005] In a first aspect, the present invention proposes a predictive operation and maintenance method for the entire life cycle of fire protection facilities, comprising the following steps: Acquire multimodal operation and maintenance data, including time series sensor data, unstructured text logs, and structured event alarms; encode the multimodal operation and maintenance data into feature sequences and input them into the autoregressive sequence prediction model; during model inference, construct key-value pairs for the context of each generated time step and store them in the KVCache; In each inference step, a context importance score is calculated for each key-value pair encoded by historical multimodal operation and maintenance data in the cache. The score is obtained by weighting the modality weight to which the data belongs, the centrality of the associated entity in the operation and maintenance event ontology knowledge graph, and the cumulative attention value of the key-value pair query vector access. The key-value pairs of newly generated time steps are assigned the highest score and retained. When the KVCache capacity reaches a threshold, the key vectors in the cache are clustered based on the similar neighborhood of the key vectors to identify information clusters. Combining the score and the clustering results, redundant information clusters and isolated information points are identified from the cache. The redundant information clusters belong to information clusters with low average importance, and the isolated information points do not belong to any cluster and have low importance themselves. The summary generation network, which is pre-trained with the goal of minimizing KL divergence loss, is invoked to compress multiple key-value pairs in redundant information clusters into single summary key-value pairs with low information loss to replace the original clusters, thereby expelling isolated information points and freeing up cache capacity. The prediction model managed by KVCache is used to generate prediction results of the failure probability, remaining life and / or maintenance level of fire protection facilities within a specific future time window.

[0006] Optionally, the acquisition of multimodal operation and maintenance data includes time-series sensor data, unstructured text logs, and structured event alarms; the multimodal operation and maintenance data is uniformly encoded into feature sequences and input into an autoregressive sequence prediction model, including: Time-series sensor data is extracted using a fixed-size sliding window, and a one-dimensional time feature vector is generated by using a one-dimensional convolutional neural network layer. Extract the word segmentation content of unstructured text logs, map each word into a numerical tensor through a word embedding matrix, and then extract and generate text feature vectors after processing by a bidirectional long short-term memory network. The discrete category fields in the structured event alarm are converted by one-hot encoding and multiplied with the set alarm feature mapping matrix to obtain the alarm feature vector; The one-dimensional time feature vector, text feature vector, and alarm feature vector are mapped to a unified feature dimension through a fully connected layer, and then concatenated along the sequence time step dimension to form a feature sequence input autoregressive sequence prediction model.

[0007] Optionally, each inference step involves calculating the context importance score for each key-value pair encoded by historical multimodal operation and maintenance data within the cache. This score is achieved by weighting the modality weight to which the data belongs, the centrality of the associated entities in the operation and maintenance event ontology knowledge graph, and the cumulative attention value of the key-value pair query vector access using a preset weighted fusion method. The score includes: Retrieve the set modality mapping relationship dictionary from the table and extract the first value; The out-degree and in-degree edges of the corresponding original data associated entities on the knowledge graph of the operation and maintenance event are summed and calculated, and then the second value is obtained after the maximum and minimum normalization process. At each time step generated by the sequence prediction model inference, the specific attention score value assigned to the key-value pair by the multi-head attention mechanism is calculated by exponential moving average and then processed by max-min normalization to obtain a third value. The first value, the normalized second value, and the normalized third value are multiplied by a fixed preset weight and summed. The summed value is then assigned to the key-value pair as a context importance score.

[0008] Optionally, the step of clustering and identifying information clusters based on key vector similarity neighborhoods in the cache when the KVCache capacity reaches a threshold includes: The cosine distance calculation function is used to calculate the pairwise cosine distance between all key vectors in KVCache and construct a complete metric matrix; Based on the above metric matrix, each key vector is used as the core query node. The entire data range is traversed and screened to identify all adjacent key vectors whose distance from the core query node is less than the judgment threshold as set elements. When the total number of elements of the query core node and its adjacent key vectors is greater than the boundary count value, the query core node and its adjacent key vectors are bound and integrated into an information cluster. Perform a union fusion operation on any different information clusters that share at least one internal element, and iteratively repeat the fusion process until there are no more information clusters that share elements. Output a set of mutually disjoint information clusters after fusion.

[0009] Optionally, the step of combining scoring and clustering results to identify redundant information clusters and isolated information points from the cache, wherein the redundant information clusters belong to information clusters with low average importance, and the isolated information points do not belong to any cluster and have low importance themselves, including: Traverse each output information cluster in sequence, obtain the context importance score of all elements in a single cluster that has been pre-assigned independently, and calculate the overall mean index within a single information cluster. All original key-value pairs in the corresponding information clusters whose size exceeds the lower limit of the number of elements covered and whose total average index within a single information cluster is lower than the preset importance threshold are assigned and mapped to members within redundant information clusters. For individual key-value pairs that are not incorporated into any output information cluster and whose independently calculated context importance scores are lower than the limit constant, the individual key-value pairs are divided and recorded in the isolated information point archive queue within the underlying data mapping list.

[0010] Optionally, the invocation of a summary generation network pre-trained with the goal of minimizing KL divergence loss to compress multiple key-value pairs within redundant information clusters into a single summary key-value pair with low information loss to replace the original cluster includes: Extract a set of internal data elements belonging to a single redundant information cluster, and stack all the extracted original key vectors and the bound value vectors in order based on the time evolution axis dimension to obtain a combined input tensor matrix; The combined input tensor matrix is ​​input into a summary generation network containing a self-attention layer and a feedforward neural network processing layer to extract the compact feature layer array in the hidden state. Global average pooling is performed on the compressed feature layer array along the time step dimension to obtain a unique summary hidden state vector; the summary hidden state vector is input into the key vector mapping head and the value vector mapping head respectively to generate a summary key vector and a summary value vector with a set feature vector dimension respectively; The digest key vector and digest value vector are stored as a single digest key-value pair in KVCache, replacing the initial position index of the redundant information cluster at the beginning of the column. The remaining key-value pairs within the scope of the information cluster are cleared, and the key-value pair tensors in KVCache are compacted and rearranged along the time series dimension to maintain dimensional continuity.

[0011] Optionally, evicting isolated information points to release cache capacity includes: Based on the upper-level isolated information point archiving queue, the logical position index of the corresponding data element body is retrieved in the tensor management data structure of KVCache; A Boolean mask is generated using logical position indexes, and the key-value pairs corresponding to the isolated information points are filtered out from the KVCache tensor during cache updates. The filtered valid key-value pair tensor is compacted and rearranged along the time series dimension to reduce the sequence length, and the original cache logic capacity is released during subsequent overwrites.

[0012] On the other hand, the present invention also proposes a predictive operation and maintenance system for the entire life cycle of fire protection facilities, comprising the following modules: The input module is used to acquire multimodal operation and maintenance data, including time series sensor data, unstructured text logs, and structured event alarms; the multimodal operation and maintenance data is uniformly encoded into feature sequences and input into the autoregressive sequence prediction model; during model inference, key-value pairs are constructed for the context of each generated time step and stored in KVCache. The calculation module, used for each inference step, calculates the context importance score for each key-value pair encoded by historical multimodal operation and maintenance data in the cache. The score is obtained by weighting the modality weight to which the data belongs, the centrality of the associated entity in the operation and maintenance event ontology knowledge graph, and the cumulative attention value of the key-value pair query vector access by preset weights. The key-value pairs of newly generated time steps are assigned the highest score and retained. When the KVCache capacity reaches a threshold, the key vectors in the cache are clustered based on the similar neighborhood of the key vectors to identify information clusters. Combining the score and the clustering results, redundant information clusters and isolated information points are identified from the cache. The redundant information clusters belong to information clusters with low average importance, and the isolated information points do not belong to any cluster and have low importance themselves. The generation module is used to call a summary generation network pre-trained with the training objective of minimizing KL divergence loss to compress multiple key-value pairs in redundant information clusters into single summary key-value pairs with low information loss to replace the original clusters, and to expel isolated information points to release cache capacity; and to use the prediction model managed by KVCache to generate prediction results of the failure probability, remaining life and / or maintenance level of fire protection facilities in a specific future time window.

[0013] Preferably, the acquisition of multimodal operation and maintenance data includes time-series sensor data, unstructured text logs, and structured event alarms; the multimodal operation and maintenance data is uniformly encoded into feature sequences and input into an autoregressive sequence prediction model, including: Time-series sensor data is extracted using a fixed-size sliding window, and a one-dimensional time feature vector is generated by using a one-dimensional convolutional neural network layer. Extract the word segmentation content of unstructured text logs, map each word into a numerical tensor through a word embedding matrix, and then extract and generate text feature vectors after processing by a bidirectional long short-term memory network. The discrete category fields in the structured event alarm are converted by one-hot encoding and multiplied with the set alarm feature mapping matrix to obtain the alarm feature vector; The one-dimensional time feature vector, text feature vector, and alarm feature vector are mapped to a unified feature dimension through a fully connected layer, and then concatenated along the sequence time step dimension to form a feature sequence input autoregressive sequence prediction model.

[0014] Preferably, the inference step is characterized in that each inference step involves calculating the context importance score for each key-value pair encoded by historical multimodal operation and maintenance data in the cache. The score is achieved by weighting the modality weight to which the data belongs, the centrality of the associated entity in the operation and maintenance event ontology knowledge graph, and the cumulative attention value of the key-value pair query vector access using a preset weighted fusion method. This includes: Retrieve the set modality mapping relationship dictionary from the table and extract the first value; The out-degree and in-degree edges of the corresponding original data associated entities on the knowledge graph of the operation and maintenance event are summed and calculated, and then the second value is obtained after the maximum and minimum normalization process. At each time step generated by the sequence prediction model inference, the specific attention score value assigned to the key-value pair by the multi-head attention mechanism is calculated by exponential moving average and then processed by max-min normalization to obtain a third value. The first value, the normalized second value, and the normalized third value are multiplied by a fixed preset weight and summed. The summed value is then assigned to the key-value pair as a context importance score.

[0015] Preferably, the step of clustering and identifying information clusters based on key vector similarity neighborhoods in the cache when the KVCache capacity reaches a threshold includes: The cosine distance calculation function is used to calculate the pairwise cosine distance between all key vectors in KVCache and construct a complete metric matrix; Based on the above metric matrix, each key vector is used as the core query node. The entire data range is traversed and screened to identify all adjacent key vectors whose distance from the core query node is less than the judgment threshold as set elements. When the total number of elements of the query core node and its adjacent key vectors is greater than the boundary count value, the query core node and its adjacent key vectors are bound and integrated into an information cluster. Perform a union fusion operation on any different information clusters that share at least one internal element, and iteratively repeat the fusion process until there are no more information clusters that share elements. Output a set of mutually disjoint information clusters after fusion.

[0016] Preferably, the step of identifying redundant information clusters and isolated information points from the cache by combining scoring and clustering results, wherein the redundant information clusters belong to information clusters with low average importance, and the isolated information points do not belong to any cluster and have low importance themselves, including: Traverse each output information cluster in sequence, obtain the context importance score of all elements in a single cluster that has been pre-assigned independently, and calculate the overall mean index within a single information cluster. All original key-value pairs in the corresponding information clusters whose size exceeds the lower limit of the number of elements covered and whose total average index within a single information cluster is lower than the preset importance threshold are assigned and mapped to members within redundant information clusters. For individual key-value pairs that are not incorporated into any output information cluster and whose independently calculated context importance scores are lower than the limit constant, the individual key-value pairs are divided and recorded in the isolated information point archive queue within the underlying data mapping list.

[0017] Preferably, the step of calling a summary generation network pre-trained with the goal of minimizing KL divergence loss to compress multiple key-value pairs within redundant information clusters into a single summary key-value pair with low information loss to replace the original cluster includes: Extract a set of internal data elements belonging to a single redundant information cluster, and stack all the extracted original key vectors and the bound value vectors in order based on the time evolution axis dimension to obtain a combined input tensor matrix; The combined input tensor matrix is ​​input into a summary generation network containing a self-attention layer and a feedforward neural network processing layer to extract the compact feature layer array in the hidden state. Global average pooling is performed on the compressed feature layer array along the time step dimension to obtain a unique summary hidden state vector; the summary hidden state vector is input into the key vector mapping head and the value vector mapping head respectively to generate a summary key vector and a summary value vector with a set feature vector dimension respectively; The digest key vector and digest value vector are stored as a single digest key-value pair in KVCache, replacing the initial position index of the redundant information cluster at the beginning of the column. The remaining key-value pairs within the scope of the information cluster are cleared, and the key-value pair tensors in KVCache are compacted and rearranged along the time series dimension to maintain dimensional continuity.

[0018] Preferably, the step of evicting isolated information points to release cache capacity includes: Based on the upper-level isolated information point archiving queue, the logical position index of the corresponding data element body is retrieved in the tensor management data structure of KVCache; A Boolean mask is generated using logical position indexes, and the key-value pairs corresponding to the isolated information points are filtered out from the KVCache tensor during cache updates. The filtered valid key-value pair tensor is compacted and rearranged along the time series dimension to reduce the sequence length, and the original cache logic capacity is released during subsequent overwrites.

[0019] This invention proposes a predictive operation and maintenance (O&M) method for fire protection facilities based on an improved multimodal data KVCache management mechanism. By integrating time-series sensor data, text logs, and structured alarm data, and combining this with an O&M event ontology knowledge graph to extract related features, the stability and accuracy of predictions regarding facility failure probability, remaining lifespan, and maintenance level are improved. To address the issue of continuously increasing KVCache capacity during long-sequence inference, a contextual importance assessment mechanism is constructed that integrates modality weights, knowledge graph entity centrality, and cumulative attention values. This is combined with a clustering method based on key vector similarity neighborhoods to analyze the information distribution within the cache. For redundant information clusters with high similarity but low importance, a summary generation network is used for low-information-loss semantic compression. For isolated information points with low importance that are not incorporated into information clusters, eviction processing is performed, thereby reducing cache usage and inference latency, and improving the operational efficiency and computational stability of long-cycle O&M prediction tasks. Attached Figure Description

[0020] Figure 1 A flowchart of the first embodiment; Figure 2 This is a schematic diagram illustrating feature dimension alignment. Figure 3 This is a diagram illustrating the composition of the context importance score. Figure 4 This is a diagram illustrating the trend of cache usage over time. Detailed Implementation

[0021] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0022] In the first embodiment, the present invention proposes a predictive operation and maintenance method for the entire life cycle of fire protection facilities, such as... Figure 1 As shown, it includes the following steps: S1. Acquire multimodal operation and maintenance data, including time series sensor data, unstructured text logs, and structured event alarms; encode the multimodal operation and maintenance data into feature sequences and input them into the autoregressive sequence prediction model; during model inference, construct key-value pairs for the context of each generated time step and store them in the KVCache.

[0023] The initial data parsing and preprocessing parameters were verified through the standard data loading module. The multilayer perceptron was used to map and transform the data to the model backbone network, and the key-value pair tensor structure required for the underlying autoregressive prediction was established and stored in the video memory queue.

[0024] In an optional embodiment, the acquisition of multimodal operation and maintenance data includes time-series sensor data, unstructured text logs, and structured event alarms; the multimodal operation and maintenance data is uniformly encoded into feature sequences and input into an autoregressive sequence prediction model, including: Time-series sensor data is extracted using a fixed-size sliding window, and a one-dimensional time feature vector is generated by using a one-dimensional convolutional neural network layer. Extract the word segmentation content of unstructured text logs, map each word into a numerical tensor through a word embedding matrix, and then extract and generate text feature vectors after processing by a bidirectional long short-term memory network. The discrete category fields in the structured event alarm are converted by one-hot encoding and multiplied with the set alarm feature mapping matrix to obtain the alarm feature vector; The one-dimensional time feature vector, text feature vector, and alarm feature vector are mapped to a unified feature dimension through a fully connected layer, and then concatenated along the sequence time step dimension to form a feature sequence input autoregressive sequence prediction model.

[0025] The sliding window length L and sliding step S are set to segment real-time numerical sequences of time-series sensor data, such as ambient temperature and water pressure. For example, the sliding window length L=128 corresponds to a data acquisition duration of 128 minutes, and the step S=16. The segmented time segments are input into a one-dimensional convolutional neural network layer for local feature extraction. The input to this one-dimensional convolutional neural network layer is the time series segment, and the structure contains three convolutional blocks with kernel sizes of 5, 3, and 3, and output channels of 32, 64, and 128 respectively. After flattening through max pooling, the output is a one-dimensional time feature vector of length 256. For unstructured text logs, a word segmentation tool is used for cleaning and word segmentation. A vocabulary is built based on historical maintenance text logs, alarm descriptions, and inspection records, and a unique index is assigned to each word in the vocabulary. Using a word embedding matrix obtained by pre-training on historical text corpora or joint training in maintenance prediction tasks, each segmented word is mapped to a word embedding numerical tensor, with the word embedding dimension set to [missing value]. =128. The tensor sequence is fed into a bidirectional long short-term memory network step by step. The input of the bidirectional long short-term memory network is the above word embedding numerical tensor sequence. The structure is a loop structure with a two-layer stacked structure and 128 hidden state nodes in each layer. The output is a text feature vector with a dimension of 256 generated by concatenating the positive and negative hidden layer states of the last time step.

[0026] Structured event alarm data typically records discrete state variables such as specific components and error codes, which are then one-hot encoded. For example, a certain type of alarm may cover 10 states, including low voltage and communication disconnection. After one-hot encoding, this forms a 10-dimensional Boolean sparse vector. This sparse vector is then multiplied by an alarm feature mapping matrix trained on historical records with a size of 10×64 to reduce its dimensionality, outputting a 64-dimensional alarm feature vector. To achieve feature alignment and fusion, the obtained 256-dimensional time features, 256-dimensional text features, and 64-dimensional alarm features are fed into a fully connected network. The input to this fully connected network is each unimodal feature vector, and the network structure consists of independently configured neural connection layers, each containing 256 neurons. The output is a vector mapped to the same unified feature dimension space, meaning the unimodal vector dimensions converge to a single dimension. Feature vectors with a length of 256. Feature dimension alignment is as follows. Figure 2 As shown, the mapped three feature matrices are concatenated and aggregated on their respective time-step dimensions to form a multimodal input sequence, which is then passed sequentially to the autoregressive prediction model. The input of the autoregressive prediction model is the multimodal feature input sequence obtained by the above concatenation and aggregation. The structure is a decoder architecture that includes a combination of multi-layer self-attention mechanism and feedforward network, and the output is the temporal prediction result of the next time step.

[0027] Optionally, the alarm feature mapping matrix is ​​obtained through training on historical alarm records, including collecting historical structured event alarm data and their corresponding subsequent fault, maintenance, or operation status labels, and performing one-hot encoding on the alarm category field to obtain a 10-dimensional alarm state vector; using the alarm state vector as input, inputting an alarm feature extraction network containing a linear mapping layer, with the weight matrix size of the linear mapping layer set to 10×64; constructing a supervised loss function based on subsequent fault categories, fault probabilities, or maintenance level labels, performing backpropagation training on the linear mapping layer, and using the weight matrix of the linear mapping layer as the alarm feature mapping matrix after training convergence.

[0028] S2, at each inference step, calculate the context importance score for each key-value pair encoded by historical multimodal operation and maintenance data in the cache. The score is obtained by weighting the modality weight to which the data belongs, the centrality of the associated entity in the operation and maintenance event ontology knowledge graph, and the cumulative attention value of the key-value pair query vector access by a preset weight. The key-value pair of the newly generated time step is assigned the highest score and retained. When the KVCache capacity reaches the threshold, clustering is performed on the key vectors in the cache based on the similar neighborhood of the key vectors to identify information clusters. Combining the score and the clustering results, redundant information clusters and isolated information points are identified from the cache. The redundant information clusters belong to information clusters with low average importance, and the isolated information points do not belong to any cluster and have low importance themselves.

[0029] The centrality constant is obtained by routinely calling the graph structure attribute evaluation algorithm. The standardized feature scoring constant is obtained by combining the preset multimodal weight matrix and historical cumulative query records. The global extreme value weight of the latest time step is assigned to lock and retain the state.

[0030] After the cache threshold is triggered, the basic high-dimensional space metric algorithm is used to divide the feature space point set, and the vector set that meets the redundancy and edge features is extracted based on the point distribution density and average constant value.

[0031] In an optional embodiment, the inference step is characterized in that each inference step calculates a context importance score for each key-value pair encoded by historical multimodal operation and maintenance data in the cache. The score is achieved by weighting the modality weight to which the data belongs, the centrality of the associated entities in the operation and maintenance event ontology knowledge graph, and the cumulative attention value of the key-value pair query vector access using a preset weighted fusion method. This includes: Retrieve the set modality mapping relationship dictionary from the table and extract the first value; The out-degree and in-degree edges of the corresponding original data associated entities on the knowledge graph of the operation and maintenance event are summed and calculated, and then the second value is obtained after the maximum and minimum normalization process. At each time step generated by the sequence prediction model inference, the specific attention score value assigned to the key-value pair by the multi-head attention mechanism is calculated by exponential moving average and then processed by max-min normalization to obtain a third value. The first value, the normalized second value, and the normalized third value are multiplied by a fixed preset weight and summed. The summed value is then assigned to the key-value pair as a context importance score.

[0032] Considering the prioritization of operational data modalities, by querying a pre-built modal mapping dictionary, if the data source attribute of the currently inspected key-value pair is resolved to a sensor reading sequence, the first value is extracted as a constant of 1.0; if it is a specific alarm signal code, the first value is extracted as 0.8; if it is a system text log, the first value is assigned as 0.5. The system searches for the device entity node to which the data belongs in the pre-built operational event ontology knowledge graph, obtaining the topological edge information of the entity node on the graph. Assuming the query finds that the device entity node has 4 in-degree edges and 6 out-degree edges, the original degree is summed to obtain 10. Using pre-recorded global maximum degree set to 50 and minimum degree set to 0, the normalized association centrality variable is calculated using the maximum-minimum extremum formula, i.e., the second value is 0.20.

[0033] In the interactive considerations, the local attention scalar score of the query vector at each generation step of the autoregression and each key in the current KVCache pool is set to 0.15. Combined with the average sliding score retained at the previous time step (0.12), and updated using an exponential moving average function with a smoothing coefficient of α=0.2, the current smoothing score is obtained as 0.126. A third value of 0.126 is obtained by normalizing within the boundary value range of 0.0 to 1.0. The score is calculated using a set weighted formula, multiplying the three sets of numerical indicators by set empirical weights and summing them. According to the formula Score=0.3×1.0+0.3×0.20+0.4×0.126, the context importance score is calculated as 0.4104. This specific score is written into the metadata management table corresponding to the KVCache tensor as the basis for later determining whether to retain, compress, or evict the segment at this position. The context importance score is composed as follows... Figure 3 As shown.

[0034] In an optional embodiment, the step of clustering and identifying information clusters based on key vector similarity neighborhoods in the cache when the KVCache capacity reaches a threshold includes: The cosine distance calculation function is used to calculate the pairwise cosine distance between all key vectors in KVCache and construct a complete metric matrix; Based on the above metric matrix, each key vector is used as the core query node. The entire data range is traversed and screened to identify all adjacent key vectors whose distance from the core query node is less than the judgment threshold as set elements. When the total number of elements of the query core node and its adjacent key vectors is greater than the boundary count value, the query core node and its adjacent key vectors are bound and integrated into an information cluster. Perform a union fusion operation on any different information clusters that share at least one internal element, and iteratively repeat the fusion process until there are no more information clusters that share elements. Output a set of mutually disjoint information clusters after fusion.

[0035] Assume that when the eviction trigger is reached, the current KVCache contains N=2048 key-value tensors with a feature dimension of D=512. Extract the entire set of key vector matrices from these cached tensors, and then use dot product operations combined with norms to uniformly convert them into pairwise cosine distances. A symmetric distance metric matrix of size 2048×2048 is obtained through parallel computation. A row-by-row traversal search is performed on this two-dimensional matrix. Taking any row as the index i of the core query node, each column in the same row is scanned and filtered to find the column index j containing distance values ​​less than the threshold of 0.1. This forms a locally connected micro-set of neighborhoods. Taking node number 156 as an example, if the search finds neighboring vector indices with distance values ​​less than 0.1 including 157, 189, 212, and 334, then the core query node 156 and these four neighboring nodes are included in the basic neighborhood subset. These five nodes will initially form the basic neighborhood subset.

[0036] The total number of elements included in a micro-set is counted. If the count exceeds the set boundary count value of 3 (e.g., in the example above, it contains 5 elements greater than 3), then the micro-set is treated as an independent information cluster. During the merging and reconstruction phase, the problem of overlapping intersections among multiple neighboring micro-sets needs to be addressed. This is done by performing cross-cluster boundary checks using a disjoint-set data structure. If it is determined that two independently labeled sets have any overlapping elements (e.g., cluster A contains 156, 157, and 189, and cluster B contains 189, 190, and 205), since they share the number 189, a union operation is initiated to merge them into an expanded group containing 156, 157, 189, 190, and 205. This merging process is iterated until all remaining information clusters are mutually exclusive and non-overlapping, outputting this series of independent discrete information cluster queues.

[0037] In an optional embodiment, the step of combining scoring and clustering results to identify redundant information clusters and isolated information points from the cache, wherein the redundant information clusters belong to information clusters with low average importance, and the isolated information points do not belong to any cluster and have low importance themselves, including: Traverse each output information cluster in sequence, obtain the context importance score of all elements in a single cluster that has been pre-assigned independently, and calculate the overall mean index within a single information cluster. All original key-value pairs in the corresponding information clusters whose size exceeds the lower limit of the number of elements covered and whose total average index within a single information cluster is lower than the preset importance threshold are assigned and mapped to members within redundant information clusters. For individual key-value pairs that are not incorporated into any output information cluster and whose independently calculated context importance scores are lower than the limit constant, the individual key-value pairs are divided and recorded in the isolated information point archive queue within the underlying data mapping list.

[0038] Initiate a sequence traversal check, sequentially extracting the node code information from each information cluster entity output by the previous process. Assume that a specific number is retrieved at this point. The mutually exclusive information cluster contains key-value pair element indices for eight specific time steps. The control layer retrieves the historical individual context importance scores previously generated at these eight indices from the main cache dictionary; for example, the records are 0.25, 0.31, 0.28, 0.40, 0.33, 0.29, 0.24, and 0.30 respectively. The summation and averaging yields an internal overall mean index parameter of 0.30 for this information cluster. Comparing this to the set threshold, it is found that the actual number of members in this cluster (8) is greater than the system's lower limit, meaning it covers more than five elements; however, the internal overall mean index of this cluster is only 0.30, which is lower than the preset importance threshold of 0.40. Meeting these two conditions indicates that the cluster contains a segment of records with high similarity and relatively limited contribution to the current prediction. Therefore, all eight elements of this cluster are assigned to the redundant information cluster candidate pool.

[0039] For isolated key-value feature units that cannot be incorporated into any information cluster in topology analysis, since there is a lack of similar data within 0.1 and no more than three local neighbors, the background independently retrieves the context importance score of these isolated single objects. When verification reveals that the independent calculated importance evaluation index of an individual isolated node is lower than the set threshold constant of 0.20, for example, an extremely low score of 0.11, the decision-maker sends a removal notification to the corresponding underlying pointer and classifies it into the isolated information point archive elimination list queue in the underlying data mapping list. Using a two-tiered filtering rule, low-importance information groups with high similarity are distinguished and processed separately from low-importance isolated information nodes.

[0040] S3, invoke the summary generation network pre-trained with the training objective of minimizing KL divergence loss to compress multiple key-value pairs in redundant information clusters into single summary key-value pairs with low information loss to replace the original clusters, and evict isolated information points to release cache capacity; use the prediction model managed by KVCache to generate prediction results of the failure probability, remaining life and / or maintenance level of fire protection facilities in a specific future time window.

[0041] The pre-trained Transformer autoencoder network is invoked to compress and map redundant cluster tensors into single summary key-value pairs. During the training phase, the summary generation network constrains the difference in information distribution before and after summarization using the KL divergence loss function. During the inference phase, the trained summary generation network is invoked to cover and replace the corresponding key-value pair positions in the KVCache tensor. Boolean mask filtering and tensor compaction rearrangement are used to expel the key-value pair data corresponding to outliers. The shared fully connected layer, fault probability prediction head, remaining lifetime regression head, and / or maintenance level classification head in the prediction output layer are invoked to output the fault probability, remaining lifetime prediction value, and / or maintenance level prediction result for the future time window.

[0042] In an optional embodiment, the prediction output layer includes at least one of a shared fully connected layer, a fault probability prediction head, a remaining lifetime regression head, and a maintenance level classification head. The autoregressive sequence prediction model performs attention calculations based on the updated KVCache and recursively outputs the time-series prediction feature sequence within the future time window in an autoregressive manner. First, the time-series prediction feature sequence is input into the shared fully connected layer for feature fusion to obtain the maintenance prediction feature vector. When a fault probability needs to be output, the maintenance prediction feature vector is input into the fault probability prediction head, and the probability of fault occurrence within the future time window is obtained through the Sigmoid function. When the remaining lifetime needs to be output, the maintenance prediction feature vector is input into the remaining lifetime regression head, and the remaining lifetime prediction value is output through a linear layer combined with a non-negative activation function. When the maintenance level needs to be output, the maintenance prediction feature vector is input into the maintenance level classification head, and the probability distribution of each maintenance level is output through the Softmax function. The category with the highest probability is determined as the maintenance level prediction result. During the training phase, training samples are constructed using historical multimodal operation and maintenance data and their corresponding fault markers, remaining life labels, and maintenance level labels. Binary cross-entropy loss, mean squared error loss, and cross-entropy loss are applied according to the output task, and the weighted summation according to preset weights forms a joint training loss. The autoregressive sequence prediction model, the summary generation network, and the prediction output layer are jointly trained or trained in stages. Based on the validation set, the modal weights, importance thresholds, neighborhood determination thresholds, and boundary counts used in the KVCache management process are adjusted until the joint training loss converges.

[0043] In an optional embodiment, the invocation of a summary generation network pre-trained with the goal of minimizing KL divergence loss to compress multiple key-value pairs within a redundant information cluster into a single summary key-value pair with low information loss to replace the original cluster includes: Extract a set of internal data elements belonging to a single redundant information cluster, and stack all the extracted original key vectors and the bound value vectors in order based on the time evolution axis dimension to obtain a combined input tensor matrix; The combined input tensor matrix is ​​input into a summary generation network containing a self-attention layer and a feedforward neural network processing layer to extract the compact feature layer array in the hidden state. Global average pooling is performed on the compressed feature layer array along the time step dimension to obtain a unique summary hidden state vector; the summary hidden state vector is input into the key vector mapping head and the value vector mapping head respectively to generate a summary key vector and a summary value vector with a set feature vector dimension respectively; The digest key vector and digest value vector are stored as a single digest key-value pair in KVCache, replacing the initial position index of the redundant information cluster at the beginning of the column. The remaining key-value pairs within the scope of the information cluster are cleared, and the key-value pair tensors in KVCache are compacted and rearranged along the time series dimension to maintain dimensional continuity.

[0044] When the scheduler extracts members from a specific redundant cluster, it assumes that the cluster contains a set of associated original data pairs with eight consecutive time points and each feature dimension reaching D=128. These eight original keys are removed from the original memory structure and stacked sequentially to form a local temporal combination input key matrix of shape 8×128. The value tensor is also constructed as a tensor combination of the same size. The local temporal combination input key matrix and the value matrix are concatenated according to the feature dimension, or they are added and fused after being mapped by a linear layer, to obtain the combined input tensor matrix of the summarization generation network. This type of input matrix is ​​then input to the pre-trained summarization generation network. The input of the pre-trained summarization generation network is the local temporal combination input tensor matrix. The network structure of this matrix includes a single self-attention layer with a head set to 2 for feature fusion, and a feedforward neural network layer with an activation function for feature extraction. The output is a hidden state feature array that has been aggregated after the interchange mapping. After entering the network model structure, the matrix sequence with a size of 8×128 undergoes the aforementioned attention and feedforward layer structure, and outputs a compact feature array that retains the original 8×128 size.

[0045] The time-axis dimension index is used to sum the eight cross-sectional columns containing data of each dimension, and the arithmetic mean is calculated by multiplying by a coefficient of one-eighth. After pooling compression, the array of the above eight sequence depths is compressed and output as a 1×128-sized summary hidden state vector. The summary hidden state vector is input into the key vector mapping head and value vector mapping head set at the end of the summary generation network, respectively, to generate a 1×128-sized summary key vector and a 1×128-sized summary value vector, which together constitute a single summary key-value pair. Under the constraint of minimizing KL divergence loss as the training objective, the summary key-value pair can retain key information in the original redundant information cluster to a certain extent. During the compression stage, an overwrite instruction is issued to KVCache to overwrite the summary key vector and summary value vector to the starting position of the original interval, such as the 135th position of the time sequence number. The remaining seven original data positions occupied by the latter part are cleared and discarded. The data in the latter part of KVCache is sequentially slid and joined to the front end of the time sequence to complete the data, reducing the time dimension gaps caused by discarding and maintaining the continuity and compactness of the tensor dimension.

[0046] In an optional embodiment, evicting isolated information points to release cache capacity includes: Based on the upper-level isolated information point archiving queue, the logical position index of the corresponding data element body is retrieved in the tensor management data structure of KVCache; A Boolean mask is generated using logical position indexes, and the key-value pairs corresponding to the isolated information points are filtered out from the KVCache tensor during cache updates. The filtered valid key-value pair tensor is compacted and rearranged along the time series dimension to reduce the sequence length, and the original cache logic capacity is released during subsequent overwrites.

[0047] Data entering the orphaned information point archiving queue undergoes cache cleanup. First, the logical position indices in the orphaned information point archiving queue are read, and the corresponding key-value pairs are located within the KVCache tensor management structure of length N=2048 and feature dimension D=512. For example, if the time series positions corresponding to the orphaned information points to be evicted are 15, 66, and 128, these positions are recorded as the set of indices to be deleted. Then, a one-dimensional Boolean mask array of length 2048 is created, and all initial values ​​of the mask array are set to logical true. Next, the mask values ​​corresponding to positions 15, 66, and 128 are modified to logical false, resulting in the Boolean mask used for cache filtering.

[0048] Based on the Boolean mask, a filtering operation is synchronously performed on the key tensors, value tensors, and corresponding metadata tables in the KVCache. Key vectors, value vectors, and their associated metadata at positions where the mask value is logically false are deleted, retaining only valid key-value pairs with a mask value of logically true. After filtering, the remaining valid key-value pairs are compacted and rearranged along the time series dimension in their original chronological order, reducing the KVCache length from 2048 to 2045 while maintaining the continuous arrangement of cache tensors. The three logical cache positions released at the tail of the queue serve as reusable cache capacity, cyclically used when generating new key-value pairs during subsequent inference. The cache usage trend over time is as follows: Figure 4 As shown.

[0049] In the second embodiment, the present invention also proposes a predictive operation and maintenance system for the entire life cycle of fire protection facilities, comprising the following modules: The input module is used to acquire multimodal operation and maintenance data, including time series sensor data, unstructured text logs, and structured event alarms; the multimodal operation and maintenance data is uniformly encoded into feature sequences and input into the autoregressive sequence prediction model; during model inference, key-value pairs are constructed for the context of each generated time step and stored in KVCache. The calculation module, used for each inference step, calculates the context importance score for each key-value pair encoded by historical multimodal operation and maintenance data in the cache. The score is obtained by weighting the modality weight to which the data belongs, the centrality of the associated entity in the operation and maintenance event ontology knowledge graph, and the cumulative attention value of the key-value pair query vector access by preset weights. The key-value pairs of newly generated time steps are assigned the highest score and retained. When the KVCache capacity reaches a threshold, the key vectors in the cache are clustered based on the similar neighborhood of the key vectors to identify information clusters. Combining the score and the clustering results, redundant information clusters and isolated information points are identified from the cache. The redundant information clusters belong to information clusters with low average importance, and the isolated information points do not belong to any cluster and have low importance themselves. The generation module is used to call a summary generation network pre-trained with the training objective of minimizing KL divergence loss to compress multiple key-value pairs in redundant information clusters into single summary key-value pairs with low information loss to replace the original clusters, and to expel isolated information points to release cache capacity; and to use the prediction model managed by KVCache to generate prediction results of the failure probability, remaining life and / or maintenance level of fire protection facilities in a specific future time window.

[0050] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0051] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A predictive operation and maintenance method for the entire life cycle of fire protection facilities, characterized in that, Includes the following steps: Acquire multimodal operation and maintenance data, including time series sensor data, unstructured text logs, and structured event alarms; encode the multimodal operation and maintenance data into feature sequences and input them into the autoregressive sequence prediction model; during model inference, construct key-value pairs for the context of each generated time step and store them in the KVCache; For each inference step, a context importance score is calculated for each key-value pair encoded by historical multimodal operation and maintenance data in the cache. The score is obtained by weighting the modality weight to which the data belongs, the centrality of the associated entity in the operation and maintenance event ontology knowledge graph, and the cumulative attention value of the key-value pair query vector access through preset weights. The key-value pairs of the newly generated time step are assigned the highest score and retained. When the KVCache capacity reaches the threshold, the key vectors in the cache are clustered based on the similar neighborhood of the key vectors to identify information clusters. Combining the score and the clustering results, redundant information clusters and isolated information points are identified from the cache. The redundant information clusters belong to information clusters with low average importance, and the isolated information points do not belong to any cluster and have low importance themselves. The summary generation network, which is pre-trained with the goal of minimizing KL divergence loss, is invoked to compress multiple key-value pairs in redundant information clusters into single summary key-value pairs with low information loss to replace the original clusters, thereby expelling isolated information points and freeing up cache capacity. The prediction model managed by KVCache is used to generate prediction results of the failure probability, remaining life and / or maintenance level of fire protection facilities within a specific future time window.

2. The method according to claim 1, characterized in that, The acquisition of multimodal operation and maintenance data includes time-series sensor data, unstructured text logs, and structured event alarms; the multimodal operation and maintenance data is uniformly encoded into feature sequences and input into an autoregressive sequence prediction model, including: Time-series sensor data is extracted using a fixed-size sliding window, and a one-dimensional time feature vector is generated by using a one-dimensional convolutional neural network layer. Extract the word segmentation content of unstructured text logs, map each word into a numerical tensor through a word embedding matrix, and then extract and generate text feature vectors after processing by a bidirectional long short-term memory network. The discrete category fields in the structured event alarm are converted by one-hot encoding and multiplied with the set alarm feature mapping matrix to obtain the alarm feature vector; The one-dimensional time feature vector, text feature vector, and alarm feature vector are mapped to a unified feature dimension through a fully connected layer, and then concatenated along the sequence time step dimension to form a feature sequence input autoregressive sequence prediction model.

3. The method according to claim 1, characterized in that, Each inference step calculates the context importance score for each key-value pair encoded by historical multimodal operation and maintenance data in the cache. This score is achieved by weighting the modality weight of the data, the centrality of the associated entities in the operation and maintenance event ontology knowledge graph, and the cumulative attention value of the key-value pair query vector access, using preset weights. Retrieve the set modality mapping relationship dictionary from the table and extract the first value; The out-degree and in-degree edges of the corresponding original data associated entities on the knowledge graph of the operation and maintenance event are summed and calculated, and then the second value is obtained after the maximum and minimum normalization process. At each time step generated by the sequence prediction model inference, the specific attention score value assigned to the key-value pair by the multi-head attention mechanism is calculated by exponential moving average and then processed by max-min normalization to obtain a third value. The first value, the normalized second value, and the normalized third value are multiplied by a fixed preset weight and summed. The summed value is then assigned to the key-value pair as a context importance score.

4. The method according to claim 2, characterized in that, The step of clustering and identifying information clusters based on key vector similarity neighborhoods in the cache when the KVCache capacity reaches a threshold includes: The cosine distance calculation function is used to calculate the pairwise cosine distance between all key vectors in KVCache and construct a complete metric matrix; Based on the above metric matrix, each key vector is used as the core query node. The entire data range is traversed and screened to identify all adjacent key vectors whose distance from the core query node is less than the judgment threshold as set elements. When the total number of elements of the query core node and its adjacent key vectors is greater than the boundary count value, the query core node and its adjacent key vectors are bound and integrated into an information cluster. Perform a union fusion operation on any different information clusters that share at least one internal element, and iteratively repeat the fusion process until there are no more information clusters that share elements. Output a set of mutually disjoint information clusters after fusion.

5. The method according to claim 1, characterized in that, The method combines scoring and clustering results to identify redundant information clusters and isolated information points from the cache. Redundant information clusters belong to clusters with low average importance, while isolated information points do not belong to any cluster and have low individual importance. These include: Traverse each output information cluster in sequence, obtain the context importance score of all elements in a single cluster that has been pre-assigned independently, and calculate the overall mean index within a single information cluster. All original key-value pairs in the corresponding information clusters whose size exceeds the lower limit of the number of elements covered and whose total average index within a single information cluster is lower than the preset importance threshold are assigned and mapped to members within redundant information clusters. For individual key-value pairs that are not incorporated into any output information cluster and whose independently calculated context importance scores are lower than the limit constant, the individual key-value pairs are divided and recorded in the isolated information point archive queue within the underlying data mapping list.

6. The method according to claim 1, characterized in that, The invocation of a summary generation network pre-trained with the goal of minimizing KL divergence loss compresses multiple key-value pairs within redundant information clusters into a single summary key-value pair with low information loss, replacing the original clusters, including: Extract a set of internal data elements belonging to a single redundant information cluster, and stack all the extracted original key vectors and the bound value vectors in order based on the time evolution axis dimension to obtain a combined input tensor matrix; The combined input tensor matrix is ​​input into a summary generation network containing a self-attention layer and a feedforward neural network processing layer to extract the compact feature layer array in the hidden state. Global average pooling is performed on the compressed feature layer array along the time step dimension to obtain a unique summary hidden state vector; the summary hidden state vector is input into the key vector mapping head and the value vector mapping head respectively to generate a summary key vector and a summary value vector with a set feature vector dimension respectively; The digest key vector and digest value vector are stored as a single digest key-value pair in KVCache, replacing the initial position index of the redundant information cluster at the beginning of the column. The remaining key-value pairs within the scope of the information cluster are cleared, and the key-value pair tensors in KVCache are compacted and rearranged along the time series dimension to maintain dimensional continuity.

7. The method according to claim 1, characterized in that, The process of evicting isolated information points to free up cache capacity includes: Based on the upper-level isolated information point archiving queue, the logical position index of the corresponding data element body is retrieved in the tensor management data structure of KVCache; A Boolean mask is generated using logical position indexes, and the key-value pairs corresponding to the isolated information points are filtered out from the KVCache tensor during cache updates. The filtered valid key-value pair tensors are compacted and rearranged along the time series dimension to reduce the sequence length, and the original cache logic capacity is released during subsequent overwrites.

8. A predictive operation and maintenance system for the entire life cycle of fire protection facilities, characterized in that, Includes the following modules: The input module is used to acquire multimodal operation and maintenance data, including time series sensor data, unstructured text logs, and structured event alarms; the multimodal operation and maintenance data is uniformly encoded into feature sequences and input into the autoregressive sequence prediction model; during model inference, key-value pairs are constructed for the context of each generated time step and stored in KVCache. The calculation module is used for each inference step to calculate the context importance score for each key-value pair corresponding to the historical multimodal operation and maintenance data encoded in the cache. The score is obtained by weighting the modality weight to which the data belongs, the centrality of the associated entity in the operation and maintenance event ontology knowledge graph, and the cumulative attention value of the key-value pair query vector access through preset weights. The key-value pairs of newly generated time steps are assigned the highest score and retained. When the KVCache capacity reaches the threshold, the key vectors in the cache are clustered based on the similar neighborhood of the key vectors to identify information clusters. Combining the score and the clustering results, redundant information clusters and isolated information points are identified from the cache. The redundant information clusters belong to information clusters with low average importance, and the isolated information points do not belong to any cluster and have low importance themselves. The generation module is used to call a summary generation network pre-trained with the training objective of minimizing KL divergence loss to compress multiple key-value pairs in redundant information clusters into single summary key-value pairs with low information loss to replace the original clusters, and to expel isolated information points to release cache capacity; and to use the prediction model managed by KVCache to generate prediction results of the failure probability, remaining life and / or maintenance level of fire protection facilities in a specific future time window.

9. The system according to claim 8, characterized in that, The acquisition of multimodal operation and maintenance data includes time-series sensor data, unstructured text logs, and structured event alarms; the multimodal operation and maintenance data is uniformly encoded into feature sequences and input into an autoregressive sequence prediction model, including: Time-series sensor data is extracted using a fixed-size sliding window, and a one-dimensional time feature vector is generated by using a one-dimensional convolutional neural network layer. Extract the word segmentation content of unstructured text logs, map each word into a numerical tensor through a word embedding matrix, and then extract and generate text feature vectors after processing by a bidirectional long short-term memory network. The discrete category fields in the structured event alarm are converted by one-hot encoding and multiplied with the set alarm feature mapping matrix to obtain the alarm feature vector; The one-dimensional time feature vector, text feature vector, and alarm feature vector are mapped to a unified feature dimension through a fully connected layer, and then concatenated along the sequence time step dimension to form a feature sequence input autoregressive sequence prediction model.

10. The system according to claim 8, characterized in that, Each inference step calculates the context importance score for each key-value pair encoded by historical multimodal operation and maintenance data in the cache. This score is achieved by weighting the modality weight of the data, the centrality of the associated entities in the operation and maintenance event ontology knowledge graph, and the cumulative attention value of the key-value pair query vector access, using preset weights. Retrieve the set modality mapping relationship dictionary from the table and extract the first value; The out-degree and in-degree edges of the corresponding original data associated entities on the knowledge graph of the operation and maintenance event are summed and calculated, and then the second value is obtained after the maximum and minimum normalization process. At each time step generated by the sequence prediction model inference, the specific attention score value assigned to the key-value pair by the multi-head attention mechanism is calculated by exponential moving average and then processed by max-min normalization to obtain a third value. The first value, the normalized second value, and the normalized third value are multiplied by a fixed preset weight and summed. The summed value is then assigned to the key-value pair as a context importance score.

Images