Multi-source heterogeneous data fusion asset intelligent management method and device

Abstract

The invention discloses an asset intelligent management method and device based on multi-source heterogeneous data fusion, and relates to the related technical field of asset management, and the method comprises the steps: scanning an asset through a handheld device to generate an identification code, and collecting a multi-source heterogeneous data set; constructing a depreciation semantic mapping space, extracting a first depreciation sensitive feature vector set, calculating a preliminary prediction depreciation rate, and comparing the preliminary prediction depreciation rate with historical data to obtain a depreciation prediction error; reversely mapping to a multi-source heterogeneous data set to learn a multi-source fusion weight, updating fusion data and recalculating a second prediction depreciation rate; and inputting the target assets into the corresponding asset management sub-regions according to the second predicted depreciation rate. The technical problems that in the prior art, due to the fact that asset depreciation prediction multi-source heterogeneous data fusion is insufficient, the real state of assets cannot be reflected comprehensively, error feedback closed-loop optimization is lacked, the deviation between a prediction result and the actual state is large, and management precision is insufficient are solved. The technical effect of improving asset depreciation prediction accuracy, reliability and management precision is achieved.

Description

Intelligent asset management method and device based on multi-source heterogeneous data fusion Technical Field

[0001] This invention relates to the field of asset management technology, specifically to an intelligent asset management method and apparatus that integrates multi-source heterogeneous data. Background Technology

[0002] The accuracy and efficiency of intelligent asset management directly impact a company's operational cost control, financial planning, and resource allocation. Traditional asset management typically relies on manual inventory and data entry, which is not only time-consuming and labor-intensive but also prone to data errors, omissions, and delayed information updates, making it difficult to meet the needs of large-scale, dynamic asset management. While the automation of asset information collection has improved with the development of the Internet of Things and automatic identification, it still cannot effectively integrate and extract key information for precise management when faced with diverse, structurally varied, and temporally interwoven asset-related data. Currently, asset depreciation forecasting and management rely on accounting rules and historical experience. Commonly used models such as the straight-line method and accelerated depreciation methods often ignore multi-dimensional dynamic information such as the actual use of assets, environmental factors, and maintenance records, leading to discrepancies between forecast results and actual depreciation. Especially when a company's asset structure is complex and its usage scenarios are diverse, a single data source or a simple weighted fusion method cannot fully reflect the true state of assets, affecting the accuracy of depreciation accrual and the reasonable assessment of asset value.

[0003] Therefore, current technologies suffer from several technical problems: insufficient fusion of multi-source heterogeneous data in asset depreciation forecasting, difficulty in fully reflecting the true state of assets, and lack of error feedback closed-loop optimization, resulting in significant deviations between forecast results and actual conditions and insufficient management accuracy. Summary of the Invention

[0004] This application provides an intelligent asset management method and device that integrates multi-source heterogeneous data. This solves the technical problems in the prior art, such as insufficient integration of multi-source heterogeneous data in asset depreciation prediction, difficulty in fully reflecting the true state of assets, and lack of error feedback closed-loop optimization, which leads to large deviations between prediction results and actual conditions and insufficient management precision. This achieves the technical effect of improving the accuracy, reliability, and management precision of asset depreciation prediction.

[0005] This application provides an intelligent asset management method based on multi-source heterogeneous data fusion. The method includes: scanning a target asset using a handheld external input device, collecting asset identification information and generating an asset identification code; collecting a multi-source heterogeneous dataset corresponding to the target asset using the asset identification code as an index; constructing a depreciation semantic mapping space; inputting the multi-source heterogeneous dataset into the depreciation semantic mapping space to obtain a corresponding first set of depreciation-sensitive feature vectors; calculating a first predicted depreciation rate for the target asset based on the depreciation-sensitive feature vectors; obtaining a depreciation prediction error based on the first predicted depreciation rate and historical actual depreciation rates; back-mapping the depreciation prediction error to the multi-source heterogeneous dataset to obtain multi-source fusion weights; updating and fusion the multi-source heterogeneous dataset based on the multi-source fusion weights to re-obtain a second predicted depreciation rate; and recording the target asset into the corresponding asset management sub-area according to the second predicted depreciation rate.

[0006] In a possible implementation, the method for collecting asset identification information and generating an asset identification code includes: collecting asset identification information, wherein the asset identification information includes asset appearance image data, asset physical tag data, asset scanning pose and spatial information, and scanning timestamp; and encoding and fusing the collected asset identification information through a hash mapping function to generate an asset identification code.

[0007] In a possible implementation, the multi-source heterogeneous dataset is input into the depreciation semantic mapping space to obtain a corresponding first set of depreciation-sensitive feature vectors. The method includes: the depreciation semantic mapping space includes a predefined set of first-level depreciation semantic labels, which includes usage intensity semantics, environmental erosion semantics, structural stability semantics, maintenance adequacy semantics, and abnormal exposure semantics; for the input multi-source heterogeneous dataset, semantic parsing is performed on each type of data source to extract semantic feature description vectors; a correlation calculation function is used to perform correlation matching calculations between each type of semantic feature description vector and the first-level depreciation semantic label set to obtain a set of semantic relevance corresponding to the first-level depreciation semantic label set; and the set of semantic relevance corresponding to each type of data source is weighted and fused to obtain the first set of depreciation-sensitive feature vectors.

[0008] In a possible implementation, the set of correlation degrees corresponding to each type of data source is weighted and fused to obtain a first set of depreciation-sensitive feature vectors. The method includes: obtaining multiple correlation degrees under each depreciation semantic label in the first-level depreciation semantic label set based on the set of correlation degrees corresponding to each type of data source; performing semantic consistency analysis on the multiple correlation degrees under the same depreciation semantic label to obtain multiple semantic consistency weights; and performing weighted fusion calculation on the multiple correlation degrees according to the multiple semantic consistency weights to obtain a first set of depreciation-sensitive feature vectors corresponding to the first-level depreciation semantic label set.

[0009] In a possible implementation, the method for calculating the first predicted depreciation rate of the target asset based on the depreciation-sensitive feature vector includes: introducing a pre-trained model; obtaining an initial depreciation prediction weight vector according to the asset category of the target asset; wherein the training samples of the pre-trained model include historical asset samples corresponding to different asset sample category labels, actual depreciation rate samples corresponding to the historical asset samples, and depreciation-sensitive feature vector samples of the historical asset samples; training until convergence to obtain the pre-trained model; and calculating the first predicted depreciation rate of the target asset based on the initial depreciation prediction weight vector and the depreciation-sensitive feature vector.

[0010] In possible implementations, the method for calculating the historical actual depreciation rate includes: obtaining the starting asset value and ending asset value of the target asset by calling the historical asset management records of the handheld external input device; and obtaining the historical actual depreciation rate based on the ratio of the difference between the starting asset value and the ending asset value to the starting asset value.

[0011] In a possible implementation, the depreciation prediction error is back-mapped to the multi-source heterogeneous dataset to obtain multi-source fusion weights. The method includes: constructing a depreciation semantic association matrix based on the multi-source semantic association set of the multi-source heterogeneous dataset; decomposing the depreciation prediction error according to the first-level depreciation semantic label set to obtain the error components corresponding to the first-level depreciation semantic label set; calculating the data source error contribution according to the error components and the depreciation semantic association matrix; and back-updating the fusion weights according to the data source error contribution to obtain the multi-source fusion weights.

[0012] In a possible implementation, the asset intelligent management method based on the fusion of multi-source heterogeneous data further includes: the depreciation semantic mapping space further includes a predefined set of secondary depreciation semantic tags, wherein the set of secondary depreciation semantic tags is a refined tag set of the set of primary depreciation semantic tags.

[0013] In a possible implementation, the asset intelligent management method based on multi-source heterogeneous data fusion further includes: the handheld external input device includes multiple asset management sub-regions, and the multiple asset management sub-regions correspond to multiple depreciation rate thresholds; the target asset is input into the corresponding asset management sub-region by comparing the second predicted depreciation rate with the multiple depreciation rate thresholds.

[0014] This application also provides an intelligent asset management device for multi-source heterogeneous data fusion. The device includes: an asset information acquisition module, used to scan a target asset using a handheld external input device, collect asset identification information and generate an asset identification code, and collect a multi-source heterogeneous dataset corresponding to the target asset using the asset identification code as an index; a feature vector acquisition module, used to construct a depreciation semantic mapping space, input the multi-source heterogeneous dataset into the depreciation semantic mapping space to obtain a corresponding first depreciation sensitive feature vector set; a depreciation prediction error acquisition module, used to calculate a first predicted depreciation rate of the target asset based on the depreciation sensitive feature vectors, and obtain a depreciation prediction error based on the first predicted depreciation rate and the historical actual depreciation rate; a data update and fusion module, used to back-map the depreciation prediction error to the multi-source heterogeneous dataset to obtain a multi-source fusion weight, update and fuse the multi-source heterogeneous dataset based on the multi-source fusion weight, and re-obtain a second predicted depreciation rate; and a target asset input module, used to input the target asset into the corresponding asset management sub-area according to the second predicted depreciation rate.

[0015] This application proposes a method and apparatus for intelligent asset management based on multi-source heterogeneous data fusion. The method involves using a handheld device to scan assets and generate identification codes, collecting multi-source heterogeneous datasets. A depreciation semantic mapping space is constructed, a first set of depreciation-sensitive feature vectors is extracted, and a preliminary predicted depreciation rate is calculated and compared with historical data to obtain the depreciation prediction error. The data is then back-mapped to the multi-source heterogeneous dataset to learn multi-source fusion weights, update the fused data, and recalculate a second predicted depreciation rate. The target asset is then entered into the corresponding asset management sub-region according to the second predicted depreciation rate. This method solves the technical problems in existing technologies, such as insufficient multi-source heterogeneous data fusion in asset depreciation prediction, difficulty in comprehensively reflecting the true state of assets, and lack of error feedback closed-loop optimization, leading to significant deviations between prediction results and actual conditions and insufficient management accuracy. This method achieves the technical effect of improving the accuracy, reliability, and management precision of asset depreciation prediction. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings of the embodiments of this disclosure will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the apparatus according to the embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0017] Figure 1 is a schematic diagram of the asset intelligent management method based on multi-source heterogeneous data fusion provided in the embodiments of this application.

[0018] Figure 2 is a schematic diagram of the structure of the asset intelligent management device for multi-source heterogeneous data fusion provided in the embodiment of this application.

[0019] Figure labeling: Digital twin network construction module 10, reconstruction mode triggering module 20, reconstruction action intention generation module 30, joint simulation module 40, and collaborative action strategy generation module 50. Detailed Implementation

[0020] To further illustrate the technical means and effects adopted by the present invention in order to achieve the intended purpose, the following detailed description is provided in conjunction with the accompanying drawings and preferred embodiments, based on the specific implementation methods, structures, features and effects of the present invention.

[0021] This application provides an intelligent asset management method based on the fusion of multi-source heterogeneous data, as shown in Figure 1. The method includes: step S100, scanning the target asset using a handheld external input device, collecting asset identification information and generating an asset identification code, and collecting the multi-source heterogeneous dataset corresponding to the target asset using the asset identification code as an index.

[0022] Preferably, a portable external data entry device such as an RFID reader / writer, QR code scanner, or mobile terminal integrating sensors and cameras with data acquisition capabilities is used to scan and read the physical identification information of the target asset through close-range contact or non-contact methods. This collects asset identification information, including but not limited to unique asset coding information such as nameplate number, serial number, and asset number; physical characteristic data such as size, weight, and material specifications; identification characteristic data such as barcode, QR code, and RFID electronic tag storage information; spatial location information of the asset obtained through the built-in positioning component of the data entry device; and asset appearance image data and corresponding scan timestamps. Then, the collected raw asset identification information is processed using a hash function or Base64 encoding to convert it into a machine-recognizable unique identifier. A unique identifier string is used to obtain an asset identification code, such as AST-7F3A9B2C4D5E, which serves as the unique index key for the asset. Then, using the generated asset identification code as an index, relevant data corresponding to the target asset is collected from multiple systems. This includes at least business data such as purchase date, purchase cost, and supplier information from the ERP system; maintenance data such as repair records, maintenance cycles, and fault history from the EAM / CMMS system; environmental monitoring data such as temperature, humidity, vibration, and corrosive gas concentration from the IoT sensor network; usage record data such as runtime, workload, and number of start-ups and shutdowns collected from the production management system or log server; and visual inspection data such as surface wear, deformation, and corrosion. These data are then standardized to determine a multi-source heterogeneous dataset.

[0023] Furthermore, step S100 also includes collecting asset identification information, wherein the asset identification information includes asset appearance image data, asset physical tag data, asset scanning pose and spatial information, and scanning timestamp; the collected asset identification information is encoded and fused through a hash mapping function to generate an asset identification code.

[0024] Preferably, the asset identification information includes asset appearance image data, asset physical tag data, asset scanning pose and spatial information, and scanning timestamp. Among them, asset appearance image data refers to photos or video streams of the target asset taken by the built-in or external camera of a handheld external recording device, which includes visual information such as the overall shape of the asset, key components, and surface condition, and is used for appearance feature extraction and condition recognition. Asset physical tag data refers to the digital information carried by various physical identifiers attached to the asset entity, which are read by a handheld external recording device. This includes at least optical tags such as barcodes, QR codes, and DPM codes, as well as radio frequency tags such as RFID tags and NFC tags, and nameplate text and digital information indirectly obtained through OCR.

[0025] Preferably, the asset scanning pose and spatial information refers to the attitude data such as pitch and roll angles acquired by the handheld external input device's integrated inertial measurement unit (IMU), gyroscope, accelerometer, and other pose sensors at the moment of scanning, combined with spatial coordinate data acquired by positioning components such as GPS and UWB, to jointly determine the precise physical location and orientation of the scanning action; the scanning timestamp refers to the precise date and time when the handheld external input device performs the scan. All the collected asset identification information is concatenated into complete data blocks or strings according to a predetermined order and format. For example, first, key feature descriptors of the image data are extracted or hash values ​​are generated; text / numerical information is converted to a standard string format; pose and spatial coordinates are converted to standard format numerical values; and timestamps are converted to standard time strings. Then, a one-way hash function, such as SHA-256 or SHA-3, is selected, and the concatenated data block is used as input to calculate its hash value to ensure the uniqueness of each asset identifier and to prevent the asset identification information from being deduced from the generated hash value. Finally, the calculated hash value is used as the unique asset identifier for the target asset.

[0026] Step S200: Construct a depreciation semantic mapping space, and input the multi-source heterogeneous dataset into the depreciation semantic mapping space to obtain the corresponding first set of depreciation-sensitive feature vectors.

[0027] Step S200 further includes: the depreciation semantic mapping space includes a predefined set of first-level depreciation semantic labels, which includes usage intensity semantics, environmental erosion semantics, structural stability semantics, maintenance adequacy semantics, and abnormal exposure semantics; for the input multi-source heterogeneous dataset, perform semantic parsing for each type of data source to extract semantic feature description vectors; perform correlation matching calculation between each type of semantic feature description vector and the first-level depreciation semantic label set using a correlation calculation function to obtain a semantic correlation set corresponding to the first-level depreciation semantic label set; and perform weighted fusion of the semantic correlation sets corresponding to each type of data source to obtain a first set of depreciation sensitive feature vectors.

[0028] Preferably, the depreciation semantic mapping space is a structured semantic knowledge framework used to describe the depreciation of asset value. It maps multi-source heterogeneous data to key dimensions influencing asset depreciation. Specifically, based on knowledge from accounting standards, equipment engineering, materials science, and other fields, and through extensive historical asset data mining, quantifiable depreciation semantic labels are abstractly defined as the basic dimensions of the space. These include, but are not limited to, labels such as usage intensity semantics, environmental erosion semantics, structural stability semantics, maintenance adequacy semantics, and abnormal exposure semantics. Each label clearly defines its physical connotation, quantitative indicators, and data representation form. Then, data source parsing rules are established for each semantic label, for example, for environmental erosion semantics... The system employs a tagging mechanism to construct an analytical process for extracting corrosive feature vectors from temperature and humidity sensor data, corrosion detection images, and chemical gas concentration readings. For the use of intensity semantic tags, a analytical process is defined to calculate cumulative wear from operation logs, load current data, and operation count records. Next, using the cosine similarity formula, the feature vectors parsed from each data source are compared with the standard vectors of the semantic tags to calculate similarity or correlation, outputting a quantified semantic correlation score. Finally, historical data is used to validate and iteratively optimize the definition, parsing rules, and correlation calculation network of the semantic tags, ensuring stable and accurate extraction of semantic features reflecting depreciation from the original data, thereby generating a depreciation semantic mapping space.

[0029] Preferably, the core of the depreciation semantic mapping space is a predefined set of multiple first-level depreciation semantic labels representing different dimensions of depreciation impact. These include usage intensity semantics, environmental erosion semantics, structural stability semantics, maintenance adequacy semantics, and abnormal exposure semantics. Usage intensity semantics quantifies the dimensions of asset usage frequency and load, such as the wear and tear reflected by machine operating hours, vehicle mileage, and equipment cycle count. Environmental erosion semantics quantifies the dimensions of damage caused by the environment in which the asset is located, such as the chemical corrosion or physical aging caused by environmental parameters like temperature, humidity, corrosive gas / liquid concentration, dust, and vibration. Structural stability semantics quantifies the dimensions of changes in the physical structure or functional integrity of the asset, such as the gap dimensions of key components, deformation data of mechanical structures, and material fatigue indicators. Maintenance adequacy semantics quantifies the dimensions of whether the maintenance received by the asset is timely and effective, such as the execution rate of preventative maintenance plans, historical repair response time, and the frequency and quality of replacing key spare parts. Abnormal exposure semantics quantifies the dimensions and extent of abnormal events experienced by the asset, such as failures, accidents, and overload operation, such as the number of historical failures, overload alarm records, and unexpected impact events.

[0030] Preferably, for each type of data source in the input multi-source heterogeneous dataset, such as maintenance records, sensor data, and image reports, semantic parsing is performed and the data is transformed and extracted into semantic feature description vectors corresponding to the first-level depreciation semantic labels. For example, for vibration sensor time-series data, feature vectors reflecting the intensity of mechanical vibration are extracted through signal processing such as calculating the root mean square value (RMS), peak value, and frequency spectrum features; for maintenance work order text records, feature vectors such as "maintenance operation type," "replaced parts," and "fault description" are extracted through natural language processing. A cosine similarity calculation formula is defined as the relevance calculation function. The semantic feature description vectors corresponding to each type of data source are matched with the idealized feature vectors corresponding to each first-level depreciation semantic label to obtain the relevance score of each type of data source for each depreciation semantic label, indicating the extent to which the data source reflects the depreciation semantic label, and obtaining the semantic relevance set corresponding to all first-level depreciation semantic label sets. For each data source of the same asset, a semantic correlation set for five primary depreciation semantic tags is output. Then, the semantic correlation sets corresponding to each type of data source are weighted and fused to generate a comprehensive correlation value for each primary depreciation semantic tag, ultimately obtaining a first depreciation sensitive feature vector set, including a first depreciation sensitive feature vector with five dimensions. Further, step S200 also includes obtaining multiple subordinate correlations under each depreciation semantic tag in the primary depreciation semantic tag set based on the correlation set corresponding to each type of data source; performing semantic consistency analysis on the multiple subordinate correlations under the same depreciation semantic tag to obtain multiple semantic consistency weights; and performing weighted fusion calculation on the multiple subordinate correlations according to the multiple semantic consistency weights to obtain the first depreciation sensitive feature vector set corresponding to the primary depreciation semantic tag set.

[0031] Preferably, for each primary depreciation semantic tag, a correlation score from multiple different data sources is calculated. Then, based on the correlation set corresponding to each type of data source, multiple correlations under each depreciation semantic tag in the primary depreciation semantic tag set are obtained. Next, semantic consistency analysis is performed on multiple correlations under the same depreciation semantic tag. That is, the consistency or consensus among multiple correlations is evaluated by calculating the statistical confidence or inverse dispersion of the scores of this group to measure the consensus of discrete data. If multiple correlations are close to each other, it indicates that they corroborate each other and have high credibility, and a higher semantic consistency weight is assigned. If a certain degree of association differs significantly from other degrees of association, it indicates that it may reflect noise, anomalies, or specific local phenomena, and its credibility is relatively low. Therefore, a lower semantic consistency weight is assigned to it. Then, according to multiple semantic consistency weights, multiple degrees of association under the same depreciation semantic label are weighted and fused. Semantic consistency weights are repeatedly generated and weighted and fused for each first-level depreciation semantic label. Finally, the first depreciation sensitive feature vector set corresponding to the first-level depreciation semantic label set is determined. Each first depreciation sensitive feature vector set represents the quantitative status of the asset in all key depreciation dimensions.

[0032] Furthermore, step S200 also includes the fact that the depreciation semantic mapping space further includes a predefined set of secondary depreciation semantic tags, wherein the set of secondary depreciation semantic tags is a refined tag set of the set of primary depreciation semantic tags.

[0033] Preferably, the depreciation semantic mapping space adopts a two-layer hierarchical semantic structure. The first layer is a general first-level depreciation semantic label, and the second layer is a more granular second-level depreciation semantic label under the first-level depreciation semantic label. That is, the set of second-level depreciation semantic labels is a refined label set of the set of first-level depreciation semantic labels. This is used to decompose the macroscopic depreciation causes into multiple more operable and measurable sub-causes. Among them, for the usage intensity semantic, it may include mechanical wear, thermal fatigue, overload stress, start-stop cycle count, etc., to distinguish different forms of wear caused by use; for the environmental erosion semantic, it may include chemical corrosion, electrochemical corrosion, oxidative aging, ultraviolet degradation, biological erosion, etc., to indicate the specific physicochemical mechanisms of environmental erosion; for the structural stability semantic, it may include specific modes of structural failure or deterioration such as material performance degradation, geometric deformation displacement, connection fastening failure, and functional attenuation; for the maintenance adequacy semantic, it may include maintenance behaviors such as preventive maintenance compliance rate, lubrication adequacy, and calibration timeliness; for the abnormal exposure semantic, it may include abnormal event types such as sudden failure events, recordable accidents, extreme operating condition exposure, and external impact conditions.

[0034] Step S300: Calculate the first predicted depreciation rate of the target asset based on the depreciation-sensitive feature vector, and obtain the depreciation prediction error based on the first predicted depreciation rate and the historical actual depreciation rate.

[0035] Step S300 further includes introducing a pre-trained model, wherein the training samples of the pre-trained model include historical asset samples corresponding to different asset sample category labels, actual depreciation rate samples corresponding to the historical asset samples, and depreciation-sensitive feature vector samples of the historical asset samples, and training until convergence to obtain the pre-trained model; and calculating the first predicted depreciation rate of the target asset based on the initial depreciation prediction weight vector and the depreciation-sensitive feature vector.

[0036] Preferably, training samples for the pre-trained model are extracted from historical asset data for different assets. These samples include historical asset samples corresponding to different asset sample category labels, actual depreciation rate samples corresponding to historical asset samples, and depreciation-sensitive feature vector samples of historical asset samples. Training is performed based on a deep neural network. The input layer has a fixed number of 5 neurons, corresponding to the five first-level depreciation-sensitive features of the input vector. The hidden layer may have one or more layers for nonlinear feature transformation and interaction, learning the common impact of the five first-level depreciation-sensitive features on depreciation. For example, two hidden layers can be used: the first layer has 128 neurons, and the second layer has 64 neurons. The weight matrix is ​​the parameter connecting the neurons in each layer, with shapes [5, 128] and [128, 64], respectively. The bias vector is the bias parameter of each neuron, and the ReLU activation function is used to introduce nonlinearity. The output layer is a single neuron. The high-level representation learned by the hidden layer is mapped to the final depreciation rate prediction. The activation function uses a sigmoid or linear function to limit the output to a reasonable depreciation rate range of 0-1. Specifically, the historical asset samples and depreciation-sensitive feature vector samples corresponding to each asset sample category label are input into the network for forward propagation to obtain the initial predicted depreciation rate. The mean squared error is used as the loss function to calculate the error between the initial predicted depreciation rate and the actual depreciation rate samples corresponding to the historical asset samples. Gradient descent is used to optimize the calculation of the gradient of the loss function with respect to all weights and bias parameters in the network. Based on the calculated gradient and learning rate, all parameters in the network are updated. The training is repeated for all training samples until the loss converges to a low value and stabilizes, thus obtaining a pre-trained model that can predict the first predicted depreciation rate of the target asset according to the initial depreciation-sensitive feature vector samples of the target asset.

[0037] Preferably, the depreciation pattern learned from historical data is obtained according to the asset category of the target asset and used as the initial depreciation prediction weight vector, representing the weights and biases of each layer of the network for the target asset category. The depreciation-sensitive feature vector is input into the pre-trained model to perform forward prediction calculation, that is, the depreciation-sensitive feature vector is passed layer by layer in the network, and each layer performs weighted summation and nonlinear activation operations. The weights and biases used are all from the initial depreciation prediction weight vector loaded for the target asset category. Finally, a single neuron in the output layer produces a numerical output, which, after being adjusted by the final activation function, serves as the first predicted depreciation rate for the target asset.

[0038] Furthermore, step 300 also includes obtaining the starting value and ending value of the target asset by calling the historical asset management records of the handheld external input device; and obtaining the historical actual depreciation rate based on the ratio of the difference between the starting value and the ending value to the starting value.

[0039] Preferably, the historical asset management records of the target asset stored in the handheld external input device are retrieved to obtain the starting and ending asset values ​​of the target asset. The starting asset value is the value recorded on the books at the beginning of the depreciation period to be calculated, typically the net value at the end of the previous period. The ending asset value is the value recorded on the books at the end of the same depreciation period, calculated by deducting the depreciation already accrued during that period from the starting asset value and considering all impairment or appreciation adjustments. The historical actual depreciation rate is calculated using the formula: (Starting asset value - Ending asset value) / Starting asset value, representing the true value loss during that period. Then, the difference between the first predicted depreciation rate and the historical actual depreciation rate is calculated to obtain the depreciation prediction error, quantifying the accuracy of the pre-trained model's prediction.

[0040] Step S400: The depreciation prediction error is back-mapped to the multi-source heterogeneous dataset to obtain multi-source fusion weights. The multi-source heterogeneous dataset is updated and fused based on the multi-source fusion weights to obtain the second predicted depreciation rate again.

[0041] Step S400 further includes: constructing a depreciation semantic association matrix based on the multi-source semantic association set of the multi-source heterogeneous dataset; decomposing the depreciation prediction error according to the first-level depreciation semantic label set to obtain the error component corresponding to the first-level depreciation semantic label set; calculating the data source error contribution according to the error component and the depreciation semantic association matrix; and updating the fusion weights in reverse according to the data source error contribution to obtain the multi-source fusion weights.

[0042] Preferably, the depreciation prediction error is learned through reverse mapping, that is, it is backtracked and quantified and allocated to multi-source heterogeneous data. This allows for intelligent adjustment of the trust in each data source during future data fusion. Specifically, a depreciation semantic association matrix is ​​constructed to quantify the contribution of each original heterogeneous data source to each depreciation semantic dimension. The rows of the matrix represent different data sources, such as vibration sensor data, maintenance work order text, and environmental temperature and humidity logs. The columns represent multiple first-level depreciation semantic labels, and the value of each element is a set of multi-source semantic association degrees, representing the association strength of each heterogeneous data source across different depreciation semantic labels. Next, the depreciation prediction error is decomposed into five first-level depreciation semantic labels according to the numerical proportions of each first-level depreciation semantic label dimension in the current feature vector, obtaining five error components corresponding to the first-level depreciation semantic label set. Then, the depreciation semantic association matrix is ​​used as a converter, with the error component of that semantic dimension as the weight, to calculate the weighted sum of the contributions of each data source in each semantic dimension, determining the data source error contribution. A positive value indicates that the data source tends to overestimate the prediction, while a negative value indicates that the data source tends to underestimate the prediction. The absolute value represents the magnitude of its error contribution.

[0043] Preferably, the fusion weights are updated in reverse according to the error contribution of the data sources. That is, for data sources with large absolute error contribution values, their weights in the next fusion are reduced, while for data sources with error contribution values ​​close to zero or negative, their weights are maintained or increased. All fusion weights are normalized to ensure that the sum is 1, resulting in updated multi-source fusion weights. Then, the updated multi-source fusion weights are used to re-weight and fuse to generate a depreciation-sensitive feature vector, thereby generating a more accurate second depreciation-sensitive feature vector that relies more on data sources that have been verified as more reliable in the past and suppresses the influence of unreliable data sources. Similar to the process of calculating the first predicted depreciation rate, the second depreciation-sensitive feature vector is input into the pre-trained model to perform forward calculation to obtain a more accurate second predicted depreciation rate.

[0044] Step S500: Enter the target asset into the corresponding asset management sub-area according to the second predicted depreciation rate.

[0045] Step S500 further includes the following: the handheld external input device includes multiple asset management sub-areas, and the multiple asset management sub-areas correspond to multiple depreciation rate thresholds; the target asset is input into the corresponding asset management sub-area by comparing the second predicted depreciation rate with the multiple depreciation rate thresholds.

[0046] Preferably, the handheld external data entry device includes multiple asset management sub-areas, which may refer to different indicator light areas on the device or different category tabs on the display screen. Multiple depreciation rate threshold ranges are predefined, each corresponding to a different asset management sub-area and representing different management strategies. For example, sub-area A (green / normal zone) corresponds to a depreciation rate of <5%, sub-area B (yellow / concern zone) corresponds to a depreciation rate of 5% ≤ <15%, and sub-area C (red / key maintenance / disposal assessment zone) corresponds to a depreciation rate of ≥15%. The second predicted depreciation rate is compared with multiple depreciation rate thresholds to determine which depreciation rate threshold range the second predicted depreciation rate falls within. Then, the target asset is entered into the corresponding asset management sub-area. For example, the handheld external data entry device lights up a yellow indicator light and marks it as "Class B - Concern Asset" in the asset data management system, greatly improving the efficiency and accuracy of inventory, inspection, or disposal management.

[0047] In the preceding text, a method for intelligent asset management based on multi-source heterogeneous data fusion according to an embodiment of the present invention was described in detail with reference to FIG1. ​​Next, an apparatus for intelligent asset management based on multi-source heterogeneous data fusion according to an embodiment of the present invention will be described with reference to FIG2.

[0048] The asset intelligent management device based on multi-source heterogeneous data fusion according to embodiments of the present invention addresses the technical problems in existing technologies, such as insufficient fusion of multi-source heterogeneous data in asset depreciation prediction, difficulty in comprehensively reflecting the true state of assets, and lack of error feedback closed-loop optimization, leading to significant deviations between prediction results and actual conditions and insufficient management accuracy. It achieves the technical effect of improving the accuracy, reliability, and management precision of asset depreciation prediction. As shown in Figure 2, the asset intelligent management device based on multi-source heterogeneous data fusion includes: an asset information acquisition module 10, a feature vector acquisition module 20, a depreciation prediction error acquisition module 30, a data update and fusion module 40, and a target asset input module 50.

[0049] The asset information acquisition module 10 is used to scan the target asset using a handheld external input device, collect asset identification information and generate an asset identification code, and collect the multi-source heterogeneous dataset corresponding to the target asset using the asset identification code as an index; the feature vector acquisition module 20 is used to construct a depreciation semantic mapping space, input the multi-source heterogeneous dataset into the depreciation semantic mapping space to obtain the corresponding first depreciation sensitive feature vector set; the depreciation prediction error acquisition module 30 is used to calculate the first predicted depreciation rate of the target asset based on the depreciation sensitive feature vector, and obtain the depreciation prediction error based on the first predicted depreciation rate and the historical actual depreciation rate; the data update and fusion module 40 is used to back-map the depreciation prediction error to the multi-source heterogeneous dataset to obtain multi-source fusion weights, update and fuse the multi-source heterogeneous dataset based on the multi-source fusion weights, and re-obtain the second predicted depreciation rate; the target asset input module 50 is used to input the target asset into the corresponding asset management sub-area according to the second predicted depreciation rate.

[0050] The specific configuration of the asset information acquisition module 10 will be described in detail below. The asset information acquisition module 10 further includes: acquiring asset identification information, wherein the asset identification information includes asset appearance image data, asset physical tag data, asset scanning pose and spatial information, and scanning timestamp; and encoding and fusing the acquired asset identification information through a hash mapping function to generate an asset identification code.

[0051] The specific configuration of the feature vector acquisition module 20 will be described in detail below. The feature vector acquisition module 20 further includes: the depreciation semantic mapping space includes a predefined set of first-level depreciation semantic labels, which includes usage intensity semantics, environmental erosion semantics, structural stability semantics, maintenance adequacy semantics, and anomaly exposure semantics; for the input multi-source heterogeneous dataset, semantic parsing is performed on each type of data source to extract semantic feature description vectors; the semantic feature description vectors of each type are matched with the first-level depreciation semantic label set using a correlation calculation function to obtain a semantic correlation set corresponding to the first-level depreciation semantic label set; the semantic correlation sets corresponding to each type of data source are weighted and fused to obtain a first depreciation-sensitive feature vector set.

[0052] The specific configuration of the feature vector acquisition module 20 will be described in detail below. The feature vector acquisition module 20 further includes: obtaining multiple related degrees under each depreciation semantic label in the first-level depreciation semantic label set based on the related degree set corresponding to each type of data source; performing semantic consistency analysis on the multiple related degrees under the same depreciation semantic label to obtain multiple semantic consistency weights; and performing weighted fusion calculation on the multiple related degrees according to the multiple semantic consistency weights to obtain a first depreciation-sensitive feature vector set corresponding to the first-level depreciation semantic label set.

[0053] The specific configuration of the depreciation prediction error acquisition module 30 will be described in detail below. The depreciation prediction error acquisition module 30 further includes: introducing a pre-trained model; obtaining an initial depreciation prediction weight vector according to the asset category of the target asset; wherein the training samples of the pre-trained model include historical asset samples corresponding to different asset sample category labels, actual depreciation rate samples corresponding to the historical asset samples, and depreciation-sensitive feature vector samples of the historical asset samples; training until convergence to obtain the pre-trained model; and calculating a first predicted depreciation rate of the target asset based on the initial depreciation prediction weight vector and the depreciation-sensitive feature vector.

[0054] The specific configuration of the depreciation prediction error acquisition module 30 will be described in detail below. The depreciation prediction error acquisition module 30 further includes: obtaining the starting asset value and ending asset value of the target asset by calling the historical asset management records of the handheld external input device; and obtaining the historical actual depreciation rate based on the ratio of the difference between the starting asset value and the ending asset value to the starting asset value.

[0055] The specific configuration of the data update and fusion module 40 will be described in detail below. The data update and fusion module 40 further includes: constructing a depreciation semantic association matrix based on the multi-source semantic association set of the multi-source heterogeneous dataset; decomposing the depreciation prediction error according to the first-level depreciation semantic label set to obtain the error components corresponding to the first-level depreciation semantic label set; calculating the data source error contribution degree according to the error components and the depreciation semantic association matrix; and updating the fusion weights in reverse according to the data source error contribution degree to obtain the multi-source fusion weights.

[0056] The specific configuration of the feature vector acquisition module 20 will be described in detail below. The feature vector acquisition module 20 further includes: the depreciation semantic mapping space also includes a predefined set of secondary depreciation semantic labels, which is a refined label set of the primary depreciation semantic label set.

[0057] The specific configuration of the target asset entry module 50 will be described in detail below. The target asset entry module 50 further includes: the handheld external entry device includes multiple asset management sub-areas, and the multiple asset management sub-areas correspond to multiple depreciation rate thresholds; the target asset is entered into the corresponding asset management sub-area by comparing the second predicted depreciation rate with the multiple depreciation rate thresholds.

[0058] The asset intelligent management device for multi-source heterogeneous data fusion provided in the embodiments of the present invention can execute the asset intelligent management method for multi-source heterogeneous data fusion provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the method execution.

[0059] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An intelligent asset management method based on the fusion of multi-source heterogeneous data, characterized in that, The method includes: scanning the target asset using a handheld external input device, collecting asset identification information and generating an asset identification code; collecting a multi-source heterogeneous dataset corresponding to the target asset using the asset identification code as an index; constructing a depreciation semantic mapping space; inputting the multi-source heterogeneous dataset into the depreciation semantic mapping space to obtain a corresponding first set of depreciation-sensitive feature vectors; calculating a first predicted depreciation rate for the target asset based on the depreciation-sensitive feature vectors; obtaining a depreciation prediction error based on the first predicted depreciation rate and the historical actual depreciation rate; back-mapping the depreciation prediction error to the multi-source heterogeneous dataset to obtain multi-source fusion weights; updating and fusion the multi-source heterogeneous dataset based on the multi-source fusion weights to re-obtain a second predicted depreciation rate; and recording the target asset into the corresponding asset management sub-area according to the second predicted depreciation rate.

2. The asset intelligent management method based on multi-source heterogeneous data fusion as described in claim 1, characterized in that, The method for collecting asset identification information and generating asset identification codes includes: collecting asset identification information, wherein the asset identification information includes asset appearance image data, asset physical tag data, asset scanning pose and spatial information, and scanning timestamp; and encoding and fusing the collected asset identification information through a hash mapping function to generate asset identification codes.

3. The asset intelligent management method based on multi-source heterogeneous data fusion as described in claim 1, characterized in that, The method for inputting the multi-source heterogeneous dataset into the depreciation semantic mapping space to obtain the corresponding first set of depreciation-sensitive feature vectors includes: the depreciation semantic mapping space includes a predefined set of first-level depreciation semantic labels, which includes usage intensity semantics, environmental erosion semantics, structural stability semantics, maintenance adequacy semantics, and abnormal exposure semantics; for the input multi-source heterogeneous dataset, performing semantic parsing for each type of data source to extract semantic feature description vectors; performing correlation matching calculation between each type of semantic feature description vector and the first-level depreciation semantic label set using a correlation calculation function to obtain a set of semantic relevance corresponding to the first-level depreciation semantic label set; and weighted fusion of the semantic relevance sets corresponding to each type of data source to obtain the first set of depreciation-sensitive feature vectors.

4. The asset intelligent management method based on multi-source heterogeneous data fusion as described in claim 3, characterized in that, The method involves weighted fusion of the correlation sets corresponding to each type of data source to obtain a first depreciation-sensitive feature vector set. This includes: obtaining multiple correlations under each depreciation semantic label in the first-level depreciation semantic label set based on the correlation sets corresponding to each type of data source; performing semantic consistency analysis on the multiple correlations under the same depreciation semantic label to obtain multiple semantic consistency weights; and weighted fusion calculation of the multiple correlations according to the multiple semantic consistency weights to obtain the first depreciation-sensitive feature vector set corresponding to the first-level depreciation semantic label set.

5. The asset intelligent management method based on multi-source heterogeneous data fusion as described in claim 1, characterized in that, The method for calculating the first predicted depreciation rate of the target asset based on the depreciation-sensitive feature vector includes: introducing a pre-trained model; obtaining an initial depreciation prediction weight vector according to the asset category of the target asset; wherein the training samples of the pre-trained model include historical asset samples corresponding to different asset sample category labels, actual depreciation rate samples corresponding to the historical asset samples, and depreciation-sensitive feature vector samples of the historical asset samples; training until convergence to obtain the pre-trained model; and calculating the first predicted depreciation rate of the target asset based on the initial depreciation prediction weight vector and the depreciation-sensitive feature vector.

6. The asset intelligent management method based on multi-source heterogeneous data fusion as described in claim 5, characterized in that, The method for calculating the historical actual depreciation rate includes: obtaining the starting asset value and ending asset value of the target asset by calling the historical asset management records of the handheld external input device; and obtaining the historical actual depreciation rate based on the ratio of the difference between the starting asset value and the ending asset value to the starting asset value.

7. The asset intelligent management method based on multi-source heterogeneous data fusion as described in claim 3, characterized in that, The method for back-mapping the depreciation prediction error to the multi-source heterogeneous dataset to obtain multi-source fusion weights includes: constructing a depreciation semantic association matrix based on the multi-source semantic association set of the multi-source heterogeneous dataset; decomposing the depreciation prediction error according to the first-level depreciation semantic label set to obtain the error components corresponding to the first-level depreciation semantic label set; calculating the data source error contribution according to the error components and the depreciation semantic association matrix; and back-updating the fusion weights according to the data source error contribution to obtain the multi-source fusion weights.

8. The asset intelligent management method based on multi-source heterogeneous data fusion as described in claim 3, characterized in that, The depreciation semantic mapping space also includes a predefined set of secondary depreciation semantic tags, which is a refined set of tags of the primary depreciation semantic tag set.

9. The asset intelligent management method based on multi-source heterogeneous data fusion as described in claim 1, characterized in that, The handheld external input device includes multiple asset management sub-areas, each corresponding to a multiple depreciation rate threshold. The target asset is input into the corresponding asset management sub-area by comparing the second predicted depreciation rate with the multiple depreciation rate thresholds.

10. An intelligent asset management device that integrates multi-source heterogeneous data, characterized in that, The apparatus is used to implement the asset intelligent management method of multi-source heterogeneous data fusion according to any one of claims 1 to 9. The apparatus includes: an asset information acquisition module, used to scan a target asset using a handheld external input device, acquire asset identification information and generate an asset identification code, and acquire a multi-source heterogeneous dataset corresponding to the target asset using the asset identification code as an index; a feature vector acquisition module, used to construct a depreciation semantic mapping space, input the multi-source heterogeneous dataset into the depreciation semantic mapping space to obtain a corresponding first depreciation sensitive feature vector set; a depreciation prediction error acquisition module, used to calculate a first predicted depreciation rate of the target asset based on the depreciation sensitive feature vectors, and obtain a depreciation prediction error based on the first predicted depreciation rate and the historical actual depreciation rate; a data update and fusion module, used to back-map the depreciation prediction error to the multi-source heterogeneous dataset to obtain a multi-source fusion weight, update and fuse the multi-source heterogeneous dataset based on the multi-source fusion weight, and re-acquire a second predicted depreciation rate; and a target asset input module, used to input the target asset into the corresponding asset management sub-area according to the second predicted depreciation rate.

Images