Warehousing equipment fault pre-judgment method based on digital twinning and multi-modal feature migration

Abstract

A storage equipment fault pre-judgment method based on digital twinning and multi-modal feature migration comprises the following steps: carrying out multi-source heterogeneous acquisition and adaptive noise reduction purification on full life cycle data, environmental perception data and physical structure data of storage equipment to generate an equipment twinning modeling basic data set; based on the equipment twinning modeling basic data set, constructing a storage equipment digital twinning pre-judgment model through modal feature embedded mapping and virtual-real dynamic calibration; equipment operation multi-modal features and trend evolution data are extracted from the storage equipment digital twinning pre-judgment model, and a fault pre-judgment core feature set is generated through feature migration fusion and kernel space dimension reduction processing; and based on the fault pre-judgment core feature set, through failure mode Bayesian reasoning and probability evolution modeling, a storage equipment fault pre-judgment report is generated, and the storage equipment fault pre-judgment report comprises a potential fault type, an evolution rate and a pre-judgment confidence coefficient.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent operation and maintenance of warehouse equipment, in particular to a warehouse equipment fault prediction method based on digital twinning and multi-modal feature migration. BACKGROUND

[0002] In the modern warehouse logistics system, warehouse equipment, as the core carrier of cargo loading, unloading, handling and storage, directly affects the efficiency of warehouse operation and the smoothness of the logistics link. With the expansion of warehouse scale and the improvement of equipment automation, the complexity of equipment operation state has increased significantly. The traditional fault post-maintenance and regular preventive maintenance mode has been difficult to adapt to the high-efficiency, low-loss warehouse operation and maintenance requirements.

[0003] There is a warehouse equipment fault diagnosis scheme in the prior art. The scheme collects single type data of equipment operation by deploying sensing equipment, judges whether the equipment has a fault based on a preset threshold range, retrieves a historical fault case library for matching when the data is detected to be out of the threshold, determines the fault type, and finally outputs a corresponding maintenance suggestion according to the matched historical case.

[0004] However, the fault diagnosis scheme has obvious technical defects. It only relies on single type operation data and fixed threshold for fault judgment, does not realize fusion analysis of multi-source data, and cannot construct a dynamic evolution model of equipment operation state, resulting in weak prediction ability of potential equipment faults, and often can only intervene after the fault occurs, which is difficult to avoid the risk of interruption of warehouse operation caused by equipment downtime in advance. SUMMARY

[0005] To solve the above technical problems, the present application provides a warehouse equipment fault prediction method based on digital twinning and multi-modal feature migration to at least alleviate the above technical problems.

[0006] The technical scheme provided by the embodiments of the present application is as follows:

[0007] A warehouse equipment fault prediction method based on digital twinning and multi-modal feature migration, comprising the following steps:

[0008] Step 1, multi-source heterogeneous collection and adaptive denoising purification are performed on the whole life cycle data, environment perception data and physical structure data of the warehouse equipment to generate equipment twinning modeling basic data set;

[0009] Step 2, based on the equipment twinning modeling basic data set, a warehouse equipment digital twinning prediction model is constructed through modal feature embedded mapping and virtual-real dynamic calibration;

[0010] Step 3, extract equipment operation multi-modal feature and trend evolution data from the warehouse equipment digital twin prediction model, generate fault prediction core feature set through feature migration fusion and kernel space dimension reduction processing;

[0011] Step 4, based on the fault prediction core feature set, generate a warehouse equipment fault prediction report through failure mode Bayesian inference and probability evolution modeling, which contains potential fault type, evolution rate and prediction confidence.

[0012] The technical scheme provided by the present application has the following technical benefits

[0013] In view of the technical defects in the background art that the existing scheme only relies on a single data type and fixed threshold, and lacks dynamic prediction capability, the technical scheme of the present application forms a complete technical link through four core steps, realizes the early prediction of warehouse equipment failure, and the specific technical benefits are as follows:

[0014] First, the present scheme collects and adaptively denoises multi-source heterogeneous data of warehouse equipment full life cycle data, environment perception data and physical structure data, and generates equipment twin modeling basic data set. Compared with the existing scheme which only collects single type operation data, the collection of multi-source heterogeneous data covers the full range of information of the equipment itself and the running environment, and the adaptive denoising purification reduces the interference of abnormal data and redundant data on subsequent analysis, providing a high-quality data basis for building an accurate equipment model, and solving the problem of one-sided fault judgment caused by single data dimension in the existing scheme.

[0015] Second, based on the equipment twin modeling basic data set, the present scheme constructs a warehouse equipment digital twin prediction model through modal feature embedded mapping and virtual-real dynamic calibration. The digital twin model realizes the linkage of the equipment physical entity and the virtual model, and the virtual-real dynamic calibration can correct the model parameters in real time, so that the running state of the virtual model and the physical entity maintain high consistency. Compared with the defect of the existing scheme without dynamic model support, the model can dynamically reflect the change process of the equipment running state, and provides model support for potential fault mining.

[0016] Third, the present scheme extracts multi-modal features and trend evolution data from the digital twin prediction model, generates a fault prediction core feature set through feature migration fusion and kernel space dimension reduction. The migration fusion of multi-modal features breaks the information barrier of different types of features, and the kernel space dimension reduction focuses on key fault correlation features. Compared with the existing scheme without feature fusion analysis, this step can mine the correlation between different features, identify the abnormal evolution trend of the equipment running state, and improve the capture ability of potential fault features.

[0017] Fourthly, based on the fault prediction core feature set, the failure mode Bayesian inference and probability evolution modeling are used to generate a fault prediction report. Bayesian inference can combine historical fault information and current feature data for probability analysis, and probability evolution modeling can predict the evolution rate of the fault. Compared with the existing scheme which can only diagnose after the fault occurs, this step can identify potential fault types in advance and predict their development trend, output a report containing prediction confidence, help operation and maintenance personnel to take preventive measures in advance, and reduce the risk of equipment downtime.

[0018] In summary, the technical scheme of the present application solves the technical defects of weak prediction ability of existing schemes through the whole-link design of multi-source data acquisition, digital twin modeling, multi-modal feature fusion, and probability reasoning prediction, and has significant application value in the intelligent operation and maintenance scene of modern warehouse equipment. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 A warehouse equipment fault prediction method based on digital twinning and multi-modal feature migration.

[0020] Figure 2 A warehouse equipment fault prediction device based on digital twinning and multi-modal feature migration.

[0021] Figure 3 An electronic device. DETAILED DESCRIPTION

[0022] As shown in Figure 1 A warehouse equipment fault prediction method based on digital twinning and multi-modal feature migration, comprising the following steps: step 1, multi-source heterogeneous collection and adaptive denoising purification of warehouse equipment full life cycle data, environment perception data and physical structure data are performed to generate equipment twin modeling basic data set; step 2, based on the equipment twin modeling basic data set, through modal feature embedded mapping and virtual-real dynamic calibration, a warehouse equipment digital twin prediction model is constructed; step 3, equipment operation multi-modal features and trend evolution data are extracted from the warehouse equipment digital twin prediction model, and after feature migration fusion and kernel space dimension reduction processing, a fault prediction core feature set is generated; step 4, based on the fault prediction core feature set, through failure mode Bayesian inference and probability evolution modeling, a warehouse equipment fault prediction report is generated, which contains potential fault type, evolution rate and prediction confidence.

[0023] Optionally, step 1 comprises the following sub-steps: step 11, collecting design parameter data, operation history data, maintenance record data, environmental sensing data and structural stress data of the warehouse equipment to form an equipment multi-source original data set; step 12, performing adaptive noise filtering processing on the equipment multi-source original data set to remove abnormal fluctuation data and collection error data to generate an equipment denoising data set; step 13, performing feature fault correlation mining analysis on the equipment denoising data set, retaining data fields related to equipment faults, and forming an equipment twin modeling basis data set through data field screening and aggregation.

[0024] Optionally, step 11 comprises the following sub-steps: step 111, performing structural analysis on the equipment design document to obtain structural dimensions, material parameters, rated load, rated speed and rated power of the equipment to form design parameter data; step 112, sampling the operation parameters, start-stop times, operation time, mean time between failures and mean time to repair of the equipment collected by the equipment control system bus interface to form operation history data; step 113, capturing the maintenance records, component replacement information, fault handling process and inspection project detection values of the equipment retrieved by the maintenance management system interface to form maintenance record data; step 114, grabbing the temperature, humidity, dust concentration, PM2.5 content, CO concentration and SO2 concentration of the equipment operation area collected by the distributed environmental sensor array to form environmental sensing data; step 115, analyzing the stress change, deformation data, vibration frequency and amplitude of the key components of the equipment collected by the micro-electro-mechanical system stress sensing module to form structural stress data; step 116, performing heterogeneous data fusion on the design parameter data, operation history data, maintenance record data, environmental sensing data and structural stress data to form an equipment multi-source original data set.

[0025] Preferably, the specific implementation process of step 111 is as follows: first, determine the type of warehouse equipment design document, including but not limited to two-dimensional engineering drawings, three-dimensional model files, technical specifications, select the corresponding analysis protocol for different types of design documents, for example, adopt the industry standard graphics interchange protocol for two-dimensional engineering drawings, adopt the three-dimensional model data exchange protocol for three-dimensional model files, and adopt the text analysis protocol for technical specifications; extract the original structured information in the design document based on the selected analysis protocol, the original structured information contains the geometric parameter description of each component of the equipment, the material composition description, and the performance index limitation; select the target information category with high correlation with equipment failure from the original structured information, the target information category is specifically structure size information, material parameter information, rated load information, rated speed information, and rated power information; standardize the structure size information, convert size data in different units to a unified unit (for example, convert millimeters and centimeters to meters), and at the same time, eliminate redundant explanatory words in the size label to obtain standardized structure size data; classify and arrange the material parameter information, divide according to the physical properties (such as hardness, tensile strength, and heat resistance temperature) and chemical properties (such as corrosion resistance and oxidation resistance) of the material, and form classified material parameter data; perform numerical verification on the rated load information, rated speed information, and rated power information, eliminate abnormal values (for example, if the rated load of a small warehouse handling equipment is labeled as several thousand tons, it is determined to be abnormal) that obviously exceed the conventional range of equipment types, and obtain verified performance index data; integrate the standardized structure size data, classified material parameter data, and verified performance index data, establish data mapping according to the equipment component attribution relationship, and finally form the equipment design parameter data set.

[0026] Preferably, in the specific technical implementation of step 112, the type of the bus interface of the warehouse equipment control system is first identified, and common bus interface types include a Controller Area Network (CAN) bus interface, an EtherNet / IP bus interface, and a Process Field Bus (PROFIBUS) interface. Corresponding communication parameters (such as a baud rate set to hundreds of kilobits per second to several megabits per second, which can be specifically set to 500 kilobits per second), data bit length (usually 8 bits), and check bit type (such as odd check, even check, or no check) are configured according to the interface type. A stable data connection with the control system bus interface is established based on the configured communication parameters, ensuring that there is no packet loss or error during data transmission. Dynamic data during equipment operation is collected in real time through the bus interface, including operating parameters of each actuator of the equipment (such as motor operating current and voltage), equipment start and stop trigger signals, equipment continuous operation time records, equipment time intervals of previous failures, and equipment operating time after each failure repair. The collected operating parameters are subjected to unit normalization processing, and current and voltage data of different ranges are converted into a unified percentage form (for example, voltage of 0 to 220 volts is converted into a relative value of 0 to 100%). Normalized operating parameter data is obtained. The start and stop trigger signals are counted and statistically analyzed, and the time of each start and stop of the equipment is recorded in chronological order to form start and stop timing record data. The work duration record is accumulated and calculated, and the total work duration of the equipment in a unit of time (such as per day or per week) is statistically analyzed to obtain cumulative work duration data. The fault interval time and the operating time after repair are sorted, and the average fault interval time and the average repair time of the equipment in a continuous operation cycle are calculated. The average fault interval time is obtained by statistically analyzing the time intervals of the last several (for example, 30) failures and taking the arithmetic mean value, and the average repair time is obtained by statistically analyzing the time consumed for repairing the last several (for example, 30) failures and taking the arithmetic mean value, to form fault statistical data. The normalized operating parameter data, start and stop timing record data, cumulative work duration data, and fault statistical data are time-axis aligned, and an association index is established according to the data collection time stamp, and then a device operation history data set is formed.

[0027] Preferably, in a scenario, when step 113 is implemented, first determine the interface access mode of the maintenance management system, the interface access mode includes Application Programming Interface (API) call, database direct access, File Transfer Protocol (FTP) download, according to the access mode to obtain the interface access permission, configure the corresponding access parameters (such as API call key, database access user name and password, FTP server address and port); Connect to the maintenance management system through the configured access mode, call the relevant record data of the target storage device according to the preset query condition (such as device number, time range), the record data includes the detailed record of the device maintenance, the model specification and replacement time of the device replacement parts, the processing flow description when the device fault occurs, the detection value in the device regular inspection process; Structured extraction of maintenance records, extract fault occurrence time, fault location, fault phenomenon description, maintenance measures, maintenance personnel information from record text to form structured maintenance record data; Classified statistics of component replacement information, record the model, manufacturer, replacement frequency and single replacement service life of the replaced components according to the function category of the components (such as transmission components, sensing components, control components), form classified component replacement data; Step disassembly of fault handling process description, comb the process nodes according to the logical order of fault diagnosis, fault positioning, fault repair and fault verification, form process fault handling data; Effectiveness judgment of inspection detection value, compare the detection value with the device inspection standard range (such as the standard range of device operating temperature is 0 to 60 degrees Celsius), eliminate the abnormal value which exceeds the standard range and has no reasonable explanation, get the effective inspection detection data; Correlation binding of structured maintenance record data, classified component replacement data, process fault handling data and effective inspection detection data, establish the corresponding relationship of various data through device fault number, finally form device maintenance record data set.

[0028] Preferably, the specific implementation process of step 114 is as follows: first, confirm the deployment position of the distributed environmental sensor array, including the top, middle and bottom of the equipment operation area and the periphery of the key components of the equipment (such as near the motor and beside the transmission mechanism), and determine the monitoring range and direction of each sensor; identify the type of sensor, including temperature sensor, humidity sensor, dust concentration sensor, PM2.5 sensor, carbon monoxide (CO) sensor, sulfur dioxide (SO2) sensor, and configure the corresponding signal acquisition parameters according to the sensor type (such as setting the sampling frequency to several times per second to several times per minute, which can be set to 2 times per second); collect the environmental data of the equipment operation area in real time through the sensor array, including the temperature value, humidity value, dust concentration value, PM2.5 particle number, CO gas concentration and SO2 gas concentration of each position; calibrate the temperature value, correct the influence of environmental temperature on sensor measurement according to the factory calibration parameters of the sensor (for example, correct the measurement value according to the sensor temperature compensation curve in high temperature environment), and obtain the calibrated temperature data; convert the humidity value, convert the relative humidity value to absolute humidity value (unit: grams per cubic meter), and eliminate the abnormally high humidity value caused by sensor dewing, and obtain the converted humidity data; unify the units of dust concentration value and PM2.5 particle number, convert the dust concentration to milligrams per cubic meter, and convert the PM2.5 particle number to the number of particles per cubic meter, to form standardized particulate matter data; perform threshold judgment on CO gas concentration and SO2 gas concentration, eliminate abnormal concentration values exceeding the range of environmental safety standards (such as CO gas concentration exceeding 50 milligrams per cubic meter is determined as abnormal), and obtain safe range gas concentration data; integrate the calibrated temperature data, converted humidity data, standardized particulate matter data and safe range gas concentration data according to the collection position and collection time, establish the space-time index of environmental data, and further form the equipment operation environment sensing data set.

[0029] Preferably, in the specific technical implementation of step 115, first determine the installation position of the Micro-Electro-Mechanical System (MEMS) stress sensing module, the installation position includes the transmission shaft, bearing, rack, connecting rod and other key stress components of the warehouse equipment, at least one stress sensing module is deployed for each key component; configure the working parameters of the MEMS stress sensing module, set the measurement range of the module (such as the stress measurement range is 0 to several hundred megapascals, which can be set to 0 to 200 megapascals), the sampling frequency (such as several tens to several hundred times per second, which can be set to 100 times per second), the data output format (such as binary format, decimal format); collect the stress data and motion data of the key components in real time through the stress sensing module, the collected data includes the tensile stress value, compressive stress value, shear stress value borne by the component, deformation data of the component, displacement data generated by vibration; perform synthesis processing on the collected stress values, combine the tensile stress, compressive stress and shear stress to obtain the comprehensive stress value according to the stress direction of the component, and eliminate the abnormal stress peak value caused by the installation deviation of the sensor, to obtain the synthesized comprehensive stress data; perform unit conversion on the deformation data, convert the micron-level deformation data into millimeter-level data, and calculate the ratio of the deformation value to the original size of the component (i.e. strain value), to obtain strain analysis data; perform frequency domain analysis on the vibration displacement data, convert the time domain displacement data into frequency domain frequency data through Fourier transform, extract the fundamental frequency and each harmonic frequency of the vibration, and calculate the amplitude value of the vibration, to obtain frequency domain vibration characteristic data; associate the synthesized comprehensive stress data, strain analysis data and frequency domain vibration characteristic data according to the component category and collection time, establish the stress vibration correlation data of the key components, and finally form the equipment structure stress data set.

[0030] Preferably, in a scenario, step 116 is specifically implemented, first, data format identification is performed on the device design parameter data set, the device operation history data set, the device maintenance record data set, the device operation environment sensing data set, and the device structure stress data set. The data format of each data set includes table format, text format, binary format, and XML format. Format conversion is performed on data sets of different formats, and all data sets are uniformly converted into a standardized table format, wherein each row of the table represents a data record, and each column represents a data field. The data field includes data identification, collection time, data category, data value, and associated device components. Field alignment is performed on the converted standardized table data, and the public field names of each data set are unified (such as “collection time” and “data collection time” are unified as “collection time”). Missing public field data is supplemented (such as the “associated device components” field is missing in a data set, and is supplemented according to the data set source and data content). A multi-source data association index is established, taking the device component number and the collection time as the joint index key, and the data records of the same component and the same time corresponding to different data sets are associated. After association, the data is subjected to redundancy elimination, and completely repeated data records (i.e., data records with completely consistent data identification, collection time, data category, data value, and associated device components) are deleted. Partially repeated data records (i.e., records with consistent core data values but different auxiliary description fields) are merged. The data is subjected to integrity checking, and the data missing condition of each device component at each time node is counted. If the missing data amount accounts for less than a preset proportion (such as 5%), interpolation is used to supplement the data of adjacent time nodes. If the missing data amount accounts for more than the preset proportion, the time period is marked as a data missing period and is recorded. After format conversion, field alignment, association index, redundancy elimination, and integrity checking, the data is integrated, classified and stored according to the data category and the device component, and finally a device multi-source original data set is formed.

[0031] Optionally, step 12 includes the following sub-steps: step 121, using a kernel density estimation statistical filtering model to detect abnormal values in continuous data in the device multi-source original data set, and marking data outside the reasonable distribution range; step 122, using an interpolation completion model to perform adaptive replacement processing on the marked abnormal values, and outputting time-continuous preprocessed data; and step 123, using a repeated data elimination model based on hash mapping to perform redundancy record cleaning on the preprocessed data to generate a device denoising data set.

[0032] Preferably, the specific implementation process of step 121 is as follows: first, continuous data is screened out from the device multi-source original data set, and the continuous data in the warehouse scenario includes current, voltage time series data in the device running parameters, temperature, humidity continuous monitoring data in the environmental sensing data, stress, deformation time series data in the structural stress data, etc., and the corresponding device monitoring object of each type of continuous data is determined (such as motor current corresponding to motor components, bearing stress corresponding to bearing components); for each type of continuous data, based on the historical normal running samples of the warehouse equipment of this type of data, a designed kernel density estimation statistical filtering model is constructed, the kernel function of the model is selected as a Gaussian kernel function, and the bandwidth parameter of the kernel function is adaptively determined according to the discrete degree of the data, the discrete degree is obtained by calculating the standard deviation of the data, and the bandwidth parameter is a certain proportion (for example, 0.8 to 1.2 times, and specifically can be set to 1.0 times) of the standard deviation. The probability density distribution fitting of the continuous data is carried out by using the constructed kernel density estimation statistical filtering model, and the probability density distribution curve of the data under the normal running state is obtained, the abscissa of the probability density distribution curve is the data value, and the ordinate is the occurrence probability of the corresponding value; the reasonable distribution range of the data is determined based on the probability density distribution curve, the reasonable distribution range is the data interval with a probability density greater than a preset probability threshold (for example, 5%), and the preset probability threshold is set according to the data type and the requirement of the equipment fault sensitivity, for the key data (such as bearing stress data) with greater fault influence, the preset probability threshold can be appropriately reduced (for example, 3%), and for the non-key data (such as environmental humidity data), the preset probability threshold can be appropriately increased (for example, 7%); each data point in the current continuous data is compared with the reasonable distribution range, if the value of the data point exceeds the reasonable distribution range, and the adjacent multiple (for example, 3 to 5) data points before and after the data point are within the reasonable distribution range, it is determined that the data point is abnormal fluctuation data or acquisition error data, the data point is marked as abnormal, and finally the continuous data set after the abnormal value is marked is obtained.

[0033] Preferably, in the specific technical implementation of step 122, first, the continuous data set after the marked outliers is obtained, the position information of the data points with abnormal marks is extracted, the position information includes the acquisition time stamp corresponding to the data point, the data type it belongs to and the associated equipment component; for each abnormal data point, analyze the time sequence characteristics of the data type it belongs to, the time sequence characteristics of the warehouse equipment data include data change trend (such as uniform change, periodic fluctuation, random fluctuation), correlation strength of adjacent data points (judged by calculating the Pearson correlation coefficient of adjacent data points, and the absolute value of the correlation coefficient is greater than a preset threshold (for example, 0.7) is strong correlation); according to the time sequence characteristics, select the corresponding interpolation completion model, if the data type is uniform change trend and the adjacent data is strongly correlated (such as energy consumption data when the equipment runs at a uniform speed), select the linear interpolation model; if the data type is periodic fluctuation trend (such as vibration data when the equipment operates periodically), select the spline interpolation model; if the data type is random fluctuation but has certain rules (such as environmental temperature data), select the K-nearest neighbor interpolation model; take each multiple (for example, 5 to 10) normal data points before and after the abnormal data point as reference data, input the selected interpolation completion model, and the model calculates the reasonable value of the abnormal data point according to the change rule of the reference data, for example, the linear interpolation model calculates the interpolation result of the abnormal data point through the values and time interval of the two adjacent normal data points before and after the abnormal data point, and the spline interpolation model obtains the curve value corresponding to the abnormal data point by fitting the smooth curve of the reference data; replace the original abnormal data point value with the calculated reasonable value, and record the related information of the replacement operation (including the original abnormal value, the interpolation value, the selected interpolation model, and the reference data range); perform time sequence continuity verification on the replaced continuous data, calculate the change rate of adjacent data points, if the change rate exceeds the normal change rate range of this type of data (for example, the normal change rate range is ±5% per second, which is set according to the data type), readjust the parameters of the interpolation completion model (such as increasing the number of reference data points, changing the type of interpolation model), and perform interpolation calculation again until the change rates of all adjacent data points are within the normal change rate range; repeat the above process to complete the interpolation replacement of all abnormal data points, and finally output the time sequence continuous preprocessed data set.

[0034] Preferably, in one scenario, step 123 is implemented as follows: first, data structuring is performed on the time-continuous preprocessed dataset, and data of different types and from different device components are arranged in a unified format, with each data record containing a data identifier, a collection timestamp, a data type, a device component identifier, a data value, and associated parameters (such as a collection location in environmental data or a job status in operation data); then, a hash mapping model suitable for the storage scenario is designed for the structured preprocessed data, with a hash function of the model being constructed based on core features of the data records, the core features including the data type, the device component identifier, a time granularity of the collection timestamp (such as division by minutes or hours), and an approximate value of the data value (with a certain number of decimal places, for example, 2 decimal places); the hash function generates a unique hash value for each data record by performing a weighted operation on the core features (the weights of the features are set according to their importance in data deduplication in the storage, with the data type and the device component identifier having higher weights (for example, each accounting for 0.3), the time granularity of the collection timestamp having a lower weight (0.2), and the approximate value of the data value having a lower weight (0.2)); all data records in the preprocessed dataset are traversed to calculate the hash value of each record and to construct a mapping relationship table of the hash values and the data records; during the traversal, if the hash value of the current data record already exists in the mapping relationship table, the current data record is determined to be redundant and repetitive data, which includes completely repetitive data (all fields have the same values) and approximately repetitive data (the core features are the same, and there are slight differences in non-core fields); the data records determined to be redundant and repetitive are removed, and only the data record corresponding to the hash value first appearing in the mapping relationship table is retained; after the removal, integrity checks are performed on the remaining data records to ensure that there is no missing data for each type of data and each device component; if the number of records of a certain type of data is reduced too much (for example, by more than 20%) due to the removal of redundant data, the core feature selection and weight setting of the hash mapping model are rechecked, and the redundant removal operation is performed again after adjustment; finally, a device denoised dataset with redundant records removed is obtained, which ensures time continuity and removes abnormal data and redundant data.

[0035] Optionally, step 13 includes the following sub-steps: step 131, constructing a feature-fault correlation degree matrix, with rows of the matrix representing potential fault types of the device, columns representing data fields in the device denoised dataset, and matrix elements representing correlation strengths of the fields and the fault types; step 132, based on the feature-fault correlation degree matrix, filtering data fields with correlation strengths satisfying a set condition through an association strength threshold filtering model; and step 133, performing associated integration on the dataset corresponding to the filtered data fields, removing irrelevant data items, to generate a device twin modeling basis dataset.

[0036] Preferably, the specific implementation process of step 131 is as follows: first, sort out the potential fault types of the warehousing equipment, the potential fault types are determined based on the structural composition, operating principle and historical fault statistics of the warehousing equipment, including transmission component failure (such as bearing wear, gear damage, conveyor belt deviation), power system failure (such as motor overheating, motor locked-rotor, power supply anomaly), control system failure (such as sensor failure, actuator response delay, bus communication failure), structural strength failure (such as rack deformation, connecting rod fracture, fastener loosening), environmental adaptation failure (such as high temperature leading to component aging, dust blocking heat dissipation channel) and the like, each potential fault type is associated as a row vector of the feature fault correlation matrix, and is sorted in descending order of fault influence degree (such as power system failure is placed in the front row, and environmental adaptation failure is placed in the back row); extract all data fields in the equipment denoising data set, the data fields include design parameter type fields (structural size, material parameters, rated load, etc.), operating parameter type fields (working current, voltage, start-stop times, etc.), maintenance record type fields (repair times, replacement component model, inspection detection value, etc.), environmental sensing type fields (temperature, humidity, dust concentration, etc.), and structural stress type fields (stress value, deformation amount, vibration frequency, etc.), each data field is associated as a column vector of the feature fault correlation matrix, and is sorted in descending order of data acquisition real-time (such as the structural stress type field is placed in the front row, and the design parameter type field is placed in the back row); construct an initial framework of the feature fault correlation matrix, the number of rows of the matrix is equal to the number of potential fault types, the number of columns of the matrix is equal to the number of data fields, and the physical meaning of each element in the matrix is the correlation strength between the data field of the corresponding column and the potential fault type of the corresponding row; design a correlation strength calculation model, which comprehensively considers three-dimensional influence factors: mechanism correlation degree of the data field and the fault type (determined based on physical and chemical principles of equipment failure, such as the mechanism correlation degree of the motor temperature field and the motor overheating fault is higher), statistical correlation degree in historical fault data (by analyzing a large number of historical fault cases, the co-occurrence frequency of data field anomaly and fault occurrence is calculated, and the higher the co-occurrence frequency, the higher the statistical correlation degree), and sensitivity of the data field (i.e. the early warning ability of the data field value change to the fault occurrence, the sensitivity is determined by calculating the ratio of the change rate of the data field before the fault occurs to the change rate during normal operation); each influence factor is quantitatively assigned, and the assignment range is 0 to 1, wherein the mechanism correlation degree is determined by equipment failure mechanism analysis (such as the mechanism correlation degree of the bearing stress field and the bearing wear fault is assigned as 0.9, and the mechanism correlation degree of the environmental humidity field and the bearing wear fault is assigned as 0.2), the statistical correlation degree is calculated by historical data statistics (such as the abnormal rate of a data field before the corresponding fault occurs is 80%, and the statistical correlation degree is assigned as 0.8), and the data field sensitivity is determined by comparing the data change before and after the fault (such as the data field change rate before the fault occurs is 5 times that during normal operation, and the sensitivity is assigned as 0.8); the three influence factors are weighted and summed according to preset weights (the mechanism correlation degree weight is 0.5, the statistical correlation degree weight is 0.3, and the data field sensitivity weight is 0.2) to obtain the correlation strength value of each matrix element, the correlation strength value ranges from 0 to 1, and the larger the value is, the closer the corresponding data field is associated with the potential fault type; the correlation strength value obtained by calculation is normalized to ensure that the values of all matrix elements are in the interval of 0 to 1, and finally a complete feature fault correlation degree matrix is formed.

[0037] Preferably, in the specific technical implementation of step 132, first, a feature fault correlation degree matrix is obtained, and all correlation strength values corresponding to each data field in the column of the matrix (i.e., the correlation strength of the data field and each type of potential fault) are extracted; statistical analysis is performed on the correlation strength value of each data field, and statistical indicators are calculated, including the maximum value (the strongest strength value of the data field associated with all fault types), the average value (the arithmetic average of the correlation strength of the data field and all fault types), and the minimum value (the weakest strength value of the data field associated with all fault types); a correlation strength threshold screening model is designed, which includes a first-level screening threshold and a second-level screening threshold, the first-level screening threshold is a correlation strength maximum threshold, and the second-level screening threshold is a correlation strength average threshold, the values of the thresholds are determined based on the sensitivity requirement of the warehouse equipment fault prediction and the data redundancy tolerance, and are dynamically adjusted by referring to the historical data screening effect; the first-level screening threshold is set to a higher value (for example, 0.6 to 0.8, and specifically 0.7), which is used to screen out data fields highly associated with at least one type of potential fault; if the correlation strength maximum value of a data field is greater than or equal to the first-level screening threshold, the data field directly passes the first-level screening; for data fields that do not pass the first-level screening, second-level screening is entered, the second-level screening threshold is set to a lower value (for example, 0.3 to 0.5, and specifically 0.4), if the correlation strength average value of a data field is greater than or equal to the second-level screening threshold, and the potential fault type corresponding to the data field is a key fault type (such as a power system fault or a transmission component fault), the data field passes the second-level screening; for data fields that do not pass the first-level screening or the second-level screening, it is determined that the data fields are irrelevant to or weakly associated with the equipment fault prediction, and are marked as to-be-removed fields; at the same time, a special retention rule is set, if a data field does not meet the above screening conditions, but the data field is a key parameter (such as motor rated power or controller communication rate) of a core component (such as a motor or a controller) of the equipment, and has been used as an indirect basis for fault judgment in historical fault cases, the data field is also retained; finally, a list of screened data fields is obtained, which includes the name of each data field, the correlation strength maximum value, the correlation strength average value, and the main fault type associated with the data field.

[0038] Preferably, in one scenario, when step 133 is implemented, first, according to the filtered data field list, the corresponding data set of each filtered data field is extracted from the device denoising data set to obtain a plurality of single-field data sets, each single-field data set containing all data records of the data field (including acquisition timestamp, data value, associated device component, data state, etc.); the plurality of single-field data sets are subjected to timestamp alignment processing to unify the time granularity (such as millisecond level, second level, set according to data acquisition frequency, for example, 1 second) as a reference, and the data records of the same timestamp in different single-field data sets are associated to form a time-aligned multi-field data set, ensuring that various types of data at the same time node can be matched, facilitating subsequent model analysis of the comprehensive state of the device at a certain time; a data set association index is constructed, taking the device component identifier and the acquisition timestamp as the joint index key, to establish the association between different single-field data sets, for example, the motor temperature data set, the motor current data set, and the motor vibration frequency data set are associated through the motor component identifier and the same timestamp to form a motor comprehensive state data set; the associated multi-field data set is subjected to data item checking, and irrelevant data items are removed, including data records lacking corresponding index keys (such as data records without explicit device component identifier), data records with timestamp beyond a reasonable range (such as acquisition timestamp earlier than device start time or later than device stop time), and data records with empty or invalid data values; the checked multi-field data set is subjected to integrity verification, and the data missing condition of each device component and each time node is counted, if the data missing amount of a certain device component in a certain time period accounts for less than a preset proportion (for example, 5%), the data records of the time period are retained, and if the missing amount accounts for more than the preset proportion, the time period is marked as a data missing period and recorded; the data set after timestamp alignment, association index construction, data item checking, and integrity verification is integrated, classified and stored according to device components and data types, and finally a device twin modeling basis data set is generated, which only contains data fields and valid data records related to device fault prediction, and can provide accurate and efficient data support for subsequent digital twin model construction.

[0039] Optionally, step 2 includes the following sub-steps: step 21, based on the device physical structure data in the device twin modeling basis data set, device structure modeling is performed through three-dimensional modal decomposition reconstruction model to generate a device structure twin model; step 22, the multi-source features in the device twin modeling basis data set are embedded into the device structure twin model through feature embedding mapping rules to establish a one-to-one mapping relationship between the features and the structure components, to generate a feature-embedded twin model; step 23, dynamic comparison and calibration are performed between real-time sensing data and virtual state data output by the feature-embedded twin model, and the parameter deviation of the feature-embedded twin model is corrected to generate a warehouse equipment digital twin prediction model.

[0040] Optionally, step 21 comprises the following sub-steps: step 211, structurally semantic parsing of the equipment physical structure data in the equipment twin modeling base data set, extracting component composition, assembly relationship, connection mode, structure size and material parameters to form equipment structure information; step 212, using three-dimensional modal decomposition reconstruction model, decomposing the equipment structure according to function modules based on the equipment structure information, respectively performing three-dimensional geometric modeling on each substructure, and generating a substructure three-dimensional model; step 213, topologically combining and integrating each substructure three-dimensional model according to the assembly relationship, and forming an equipment structure twin model.

[0041] Preferably, the specific implementation process of step 211 is as follows: first, extract the equipment physical structure data in the equipment twin modeling basic data set, which contains the structure size data of the design parameter class, the classified material parameter data, the performance index data after verification, and the synthesized comprehensive stress data of the structure stress class, the strain analysis data; structure semantic annotation is performed on these physical structure data, the annotation rules are based on the industry specifications for warehouse equipment structure description, and the annotation content includes the data belonging to the component name (such as shelf beam, robot walking wheel, conveyor roller), data type label (structure size label, material parameter label, stress characteristic label), data correlation relationship label (such as the correlation label of “roller-shaft diameter-stainless steel”); a warehouse equipment special semantic analysis dictionary is constructed, which contains an equipment component terminology library, a structure relationship terminology library, and a material attribute terminology library, wherein the component terminology library covers common core components and auxiliary components of warehouse equipment (such as power components, transmission components, support components), the structure relationship terminology library contains structure correlation descriptions such as assembly, connection, and nesting, and the material attribute terminology library contains characteristic descriptions of commonly used materials of warehouse equipment (such as carbon steel, stainless steel, and aluminum alloy); semantic analysis is performed on the annotated physical structure data based on the semantic analysis dictionary, component composition information (such as component name, component quantity, component hierarchical relationship, for example, “the shelf is composed of columns, beams, and layer plates, the columns support the beams, and the beams bear the layer plates”) in the data, assembly relationship information (such as assembly sequence and assembly position constraint between components, for example, “the robot walking wheel is assembled on the bottom of the frame on both sides and is coaxially connected with the drive shaft”) and connection mode information (such as bolt connection, welding connection, and buckle connection, for example, “the conveyor roller is bolted with the motor through a shaft coupling”) are identified; the analyzed structure size information (such as specific numerical values and units of component length, width, height, diameter, and wall thickness) and material parameter information (such as density, tensile strength, and elastic modulus of component materials) are synchronously extracted; the analysis results are subjected to consistency verification to check whether the component composition and assembly relationship are matched (such as whether the components connected by welding are marked with adaptive welding material parameters), and whether the structure size and material parameters meet the physical logic (such as whether the material tensile strength of a thin-walled component meets the bearing demand); the component composition information, assembly relationship information, connection mode information, structure size information, and material parameter information that pass the verification are associated and integrated, and are organized in a “component-attribute-association relationship” hierarchical structure, and finally structured equipment structure information is formed.

[0042] Preferably, in the specific technical implementation of step 212, first, structured device structure information is obtained, and functional modules are divided based on the job function requirements of the warehouse device. The functional module division rules are formulated in combination with the core job logic of the device, including power supply modules (such as motors, hydraulic systems, responsible for providing device operation power), motion execution modules (such as walking mechanisms, lifting mechanisms, responsible for realizing device position or attitude adjustment), bearing support modules (such as shelves, racks, bases, responsible for bearing the weight of materials or the device itself), transmission conversion modules (such as gear sets, conveyors, shaft couplings, responsible for power transmission and motion form conversion), connection and fixing modules (such as bolts, brackets, buckles, responsible for the fixation and positioning between components); for each functional module, the corresponding component composition information, structure size information, and material parameter information are extracted, and the boundary range of each component in the module is determined (such as the power supply module including the motor, the motor bracket, and the power supply interface, but not including the subsequent transmission gear); a three-dimensional modal decomposition and reconstruction model is designed, which includes a geometric feature extraction layer, a mechanical property mapping layer, and a modal parameter calculation layer. The geometric feature extraction layer is used to extract the three-dimensional geometric contour data of the component (such as the length, width, and height of a cuboid, and the diameter and height of a cylindrical body), and to construct a basic geometric model based on the structure size information. The mechanical property mapping layer associates the material parameter information (such as the elastic modulus and Poisson's ratio) with the basic geometric model, and gives the model material mechanical properties. For example, the mechanical parameters of "stainless steel" are mapped to the basic geometric model of the conveyor drum, so that it has corresponding stiffness and strength characteristics. The modal parameter calculation layer calculates the modal parameters such as the natural frequency and mode shape of the component based on the stress condition of the component (the stress distribution during normal operation extracted from the structural stress data), for example, calculates the natural frequency of the shelf beam under the rated load, to provide a model basis for vibration anomaly detection in subsequent fault prediction. The above three-dimensional modal decomposition and reconstruction process is performed on each component in each functional module to generate a substructure three-dimensional model of each component. The substructure three-dimensional model not only contains accurate geometric shapes, but also integrates material mechanical properties and modal characteristics, while labeling the relative position information of the component in the functional module.

[0043] Preferably, in a scenario, step 213 is specifically implemented, the assembly relationship information and connection mode information of each functional module and internal component are first extracted from the equipment structure information, an assembly relationship topology graph is constructed, the nodes of the topology graph represent the substructure three-dimensional models (single components or functional modules), the node attributes include component name, functional module attribution, geometric size range, the edges of the topology graph represent the assembly relationship between the nodes, and the edge attributes include connection mode and assembly constraint conditions (such as position constraint, angle constraint and motion constraint); the combination and integration order is determined based on the assembly relationship topology graph, the integration order follows the actual assembly logic of the warehouse equipment, the integration of the basic bearing module (such as the shelf base and the equipment rack) is first performed, then the power supply module, the transmission conversion module and the motion execution module are integrated in sequence, and finally the connection and fixing module is integrated; in the integration process, the docking of the substructure three-dimensional models is performed according to the connection mode information, for example, for the bolt-connected components, the bolt hole positions of the two substructure three-dimensional models are accurately aligned, the docking gap is determined based on the bolt specification (extracted from the design parameter data) (the gap value is a certain proportion of the bolt diameter, for example, 5%, and specifically, 0.5 mm); for the welded components, the welding surfaces of the substructure three-dimensional models are completely matched, and the material mechanics properties of the welding area are mapped (for example, the tensile strength at the welding position is set as a certain proportion of the base material, for example, 90%); the assembly constraint conditions are checked in real time during the integration process, if the relative position of the substructure three-dimensional models exceeds the constraint range (for example, the coaxiality deviation of the motor shaft and the transmission gear exceeds the preset threshold value, and the threshold value is set as a certain proportion of the shaft diameter, for example, 3%, and specifically, 0.3 mm), the posture or position of the substructure three-dimensional model is adjusted until the constraint condition is met; after the combination and integration of all the substructure three-dimensional models are completed, the integrity of the overall model is detected, and whether there are problems such as missing components, connection failure and geometric interference (for example, whether there is collision interference between the conveyor roller and the rack) is checked; the detected problems are corrected, for example, the redundant connection components are deleted, the positions of the interfering components are adjusted, and the connection surfaces that are not completely docked are repaired; and finally, a complete equipment structure twin model is formed, which not only restores the physical structure form of the warehouse equipment, but also accurately reproduces the assembly relationship, connection mode and mechanics transmission path between the components.

[0044] Optionally, step 22 includes the following sub-steps: step 221, extracting multi-source feature metadata from the equipment twin modeling basic data set, and dividing the structure features, operation features and environment features into structure feature objects according to the data types; step 222, constructing a feature embedding mapping rule table, and defining a one-to-one mapping relationship between the structure feature objects and the corresponding components in the equipment structure twin model; and step 223, embedding the structure feature objects into the corresponding components of the equipment structure twin model according to the feature embedding mapping rule table, to generate a feature-embedded twin model.

[0045] Preferably, the specific implementation process of step 221 is as follows: first, traverse the device twin modeling basic data set to extract multi-source feature metadata therein, which covers structural correlation data in design parameter data, state change data in operation history data, fault correlation data in maintenance record data, influence factor data in environmental sensing data, and mechanical characteristic data in structural stress data; design feature metadata parsing rules, which are formulated based on the feature requirements of warehouse equipment fault prediction, including feature name identification, feature data type determination, feature associated component labeling, and feature time dimension division, such as identifying “motor temperature” as a feature name, determining its data type as continuous, labeling the associated component as “drive motor”, and dividing it into real-time features; classify the multi-source feature metadata based on the parsing rules, structural features include features directly related to the physical structure of the equipment, such as component structural size deviation, material performance attenuation coefficient, and connection fastening degree parameter, corresponding to the verticality of the shelf column of the warehouse equipment, the wear amount of the gear material, and the bolt pre-tightening force, etc.; operation features include dynamic features generated during equipment operation, such as motor speed, operating current, job start-stop frequency, and energy consumption change amount, corresponding to the lifting speed of the stacker, the conveying rate of the conveyor, and the turning angular velocity of the robot, etc.; environmental features include external environment-related features that affect equipment operation, such as temperature gradient of the operating area, humidity fluctuation amplitude, dust accumulation amount, and corrosive gas concentration change, corresponding to the temperature distribution of the high-temperature warehouse area and the dust adhesion amount of the shelf area with more dust, etc.; supplement the attributes of each feature, including feature data unit, collection frequency, normal value range, and fault correlation weight (determined based on the feature fault correlation degree matrix of step 131), for example, the unit of “motor temperature” is degrees Celsius, the collection frequency is once per second, the normal value range is 30 to 60 degrees Celsius, and the fault correlation weight is 0.85; encapsulate the classified structural features, operation features, and environmental features in a hierarchical structure of “feature category- associated component-feature attribute-data value” to form structured structural feature objects, each structural feature object contains a unique identification code, which is composed of feature category abbreviation, associated component number, and feature serial number (for example, the identification code of structural feature-shelf column-verticality is “SF-COL-001”).

[0046] Preferably, in the specific technical implementation of step 222, first, the structured structure feature object and the equipment structure twin model are obtained, all component information in the equipment structure twin model is extracted, including component unique identification, component function description, component position coordinates, component associated component list, for example, "drive motor (identification M001), function is to provide power, position coordinates (X100, Y50, Z80), associated components are coupling (C001), power supply interface (P002)"; the framework of the feature embedding mapping rule table is constructed, the rule table includes feature object identification code, feature category, associated component identification, mapping relationship type, feature action weight, data update frequency, constraint condition and other fields; based on the fault conduction logic of the warehouse equipment and the component function demand, the mapping relationship type is defined, including direct mapping (the feature directly acts on the corresponding component, such as the motor temperature feature directly mapping to the drive motor component), indirect mapping (the feature acts on the target component through the associated component, such as the power supply voltage feature first mapping to the power supply interface component, and then indirectly associated to the drive motor component), combined mapping (multiple features are jointly mapped to the same component, such as the gear speed feature and the torque feature combined mapping to the gear box component); the feature action weight is determined, the weight value is set based on the influence degree of the feature on the component fault prediction, the influence degree refers to the correlation strength value in the feature fault correlation matrix, the weight value ranges from 0 to 1, for example, the action weight of the motor current feature on the drive motor component is 0.9, and the action weight of the environment humidity feature on the drive motor component is 0.3; the data update frequency is set, which is determined according to the dynamic change characteristics of the feature and the response sensitivity of the component, the update frequency of the real-time feature (such as vibration frequency) is consistent with the collection frequency (for example, once per second), and the update frequency of the slow-changing feature (such as material aging coefficient) is once per hour; the constraint condition field is added, which clearly defines the prerequisite of feature embedding, for example, "the constraint condition of the gear stress feature mapping to the gear component is that the gear speed is greater than 50% of the rated speed"; for each structure feature object, based on the associated component information and the mapping rule defined above, each field of the feature embedding mapping rule table is filled, for example, the rule table record corresponding to the structure feature object "SF-M001-TEMP (drive motor temperature)" is: feature object identification code "SF-M001-TEMP", feature category "running feature", associated component identification "M001", mapping relationship type "direct mapping", feature action weight "0.88", data update frequency "once per second", constraint condition "motor running state is start"; the rule table is subjected to consistency check, whether the same feature object is mapped to multiple components, and whether there are conflicting mapping rules for the same component (such as the constraint conditions of two features on the same component contradict each other) are checked, and after the check passes, a complete feature embedding mapping rule table is formed.

[0047] Preferably, in a scenario, when step 223 is specifically implemented, first load the device structure twin model and the feature embedding mapping rule table, group the structure feature objects according to the associated component identifiers in the rule table, and group the structure feature objects mapped to the same component into a group, for example, group the feature objects such as "drive motor temperature", "drive motor current", "drive motor speed" into the "M001 (drive motor)" feature group; design a feature embedding execution model, which includes a feature analysis module, a component positioning module, a data binding module, and a dynamic updating module, the feature analysis module is used to analyze the data format, value range, unit, and other information of the structure feature object, the component positioning module accurately locates the three-dimensional coordinates and geometric model nodes of the target component in the device structure twin model based on the associated component identifier, for example, locates the shell model node and the shaft model node of the drive motor component; the data binding module binds the data of the structure feature object with the corresponding model node of the target component, and the binding mode is determined according to the feature type, for geometric related features (such as structural size deviation), the geometric parameter node of the component is bound, and the geometric shape of the model is directly corrected (for example, when the verticality deviation of the shelf column exceeds the allowed range, the three-dimensional pose of the column in the model is adjusted synchronously); for physical property related features (such as material elastic modulus), the physical property node of the component is bound, and the mechanical calculation parameters of the model are updated; for dynamic running features (such as speed and current), the state parameter node of the component is bound, and the running state identifier of the model is refreshed in real time; the dynamic updating module periodically obtains the latest feature data from the device twin modeling basic data set according to the data update frequency in the rule table, synchronously updates the data bound to the model, and ensures that the features in the model are consistent with the actual device state; during embedding, whether the feature data meets the constraint conditions in the rule table is monitored in real time, if not, the embedding of the feature is suspended, and after the constraint conditions are met, the embedding operation is executed again, for example, when the drive motor is not started, the embedding of the speed feature of the drive motor is suspended; after the embedding of all structure feature objects is completed, the integrity of the device structure twin model is detected, whether there are key components without bound features, and whether there are parameter conflicts between the feature data and the model components are checked, and after the detection is passed, the feature embedding type twin model is generated, which not only has the three-dimensional structure shape of the device, but also integrates multi-dimensional dynamic features, and the running state and potential failure trend of the device can be reflected through feature changes.

[0048] Optionally, step 23 comprises the following sub-steps: step 231, acquiring the running state data of the equipment entity collected through the real-time sensing network, and dynamically comparing the virtual state data output by the feature-embedded twin model; step 232, calculating the deviation value of the two types of data using a deviation value calculation model, and if the deviation value exceeds the set threshold range, adjusting the feature weight and structure parameter of the feature-embedded twin model through a parameter self-adaptive adjustment model; step 233, repeating the dynamic comparison and parameter adjustment process until the deviation value is within a reasonable range, to generate the warehouse equipment digital twin prediction model.

[0049] Preferably, the specific implementation process of step 231 is as follows: first, acquire the equipment entity running state data collected through the distributed real-time sensing network, the sensing network includes micro-electromechanical system stress sensors, temperature sensors, vibration sensors deployed on key components of the warehouse equipment (such as motors, gearboxes, conveyor belts, columns), and environmental sensors deployed in the operating area, and the collected data covers real-time stress values, temperature changes, vibration frequencies and amplitudes, running speeds, energy consumption data of the components, and temperature, humidity, dust concentration of the environment; preprocess the collected equipment entity running state data, including data format standardization (unifying the data storage format to JSON format), timestamp alignment (aligning all sensor data according to a unified time granularity (e.g. 10 milliseconds)), data validity check (eliminating data without collection timestamp and data value exceeding sensor range), to generate a standardized entity running state data set; call the state output interface of the feature-embedded twin model to acquire the virtual state data output by the model, the virtual state data corresponds one-to-one with the feature dimensions of the standardized entity running state data set, including component virtual stress values, virtual temperatures, virtual vibration parameters, virtual running parameters and virtual environmental impact parameters calculated by the model, such as the virtual temperature of the driving motor and the virtual vibration frequency of the gearbox simulated and generated by the model based on the embedded feature objects; construct a data comparison index table, the index table takes feature name, associated component identifier, and timestamp as joint index keys to establish the correspondence between the standardized entity running state data and the virtual state data, for example, the index key "motor temperature - driving motor - 1690000000000" corresponds to the entity motor temperature data and the virtual motor temperature data; based on the comparison index table, dynamically compare the entity data and the virtual data corresponding to the same index key point by point to generate a preliminary comparison result, which includes the entity data value, the virtual data value, the data collection time, and the associated component information of each feature dimension.

[0050] Preferably, in the specific technical implementation of step 232, first, the preliminary comparison result is obtained, the entity data value and the virtual data value are extracted therefrom, and a deviation calculation matrix is constructed, the rows of the deviation calculation matrix represent different feature dimensions, the columns represent different time stamps, and the matrix elements are the difference between the entity data value and the virtual data value of the corresponding feature dimension at the corresponding time stamp; a deviation value calculation model is designed, which includes an absolute deviation calculation layer, a relative deviation calculation layer, and a weighted comprehensive deviation calculation layer, the absolute deviation calculation layer is used to calculate the absolute difference between the entity data value and the virtual data value, for example, the entity motor temperature is 55 degrees Celsius, and the virtual motor temperature is 50 degrees Celsius, the absolute deviation is 5 degrees Celsius; the relative deviation calculation layer is used to calculate the ratio of the absolute deviation to the entity data value (retaining two decimal places), for example, the relative deviation of the above-mentioned absolute deviation 5 degrees Celsius to the entity data value 55 degrees Celsius is 0.09; the weighted comprehensive deviation calculation layer determines the weight coefficient of each feature dimension based on the correlation strength in the feature fault correlation degree matrix, the weight coefficient value range is 0 to 1, the higher the correlation strength, the greater the weight coefficient, for example, the weight coefficient of the motor temperature is 0.8, and the weight coefficient of the environmental humidity is 0.3, the comprehensive deviation value at each time stamp is calculated by weighted summation, and the comprehensive deviation value is the sum of the product of each feature dimension relative deviation and the corresponding weight coefficient; a deviation value threshold range is set, the threshold range is determined based on the operation accuracy requirement and the fault sensitivity requirement of the warehouse equipment, including an absolute threshold and a relative threshold, the absolute threshold is the maximum allowed value of the comprehensive deviation value (for example, 0.15), and the relative threshold is the upper limit of the ratio of the comprehensive deviation value to the historical average comprehensive deviation value (for example, 1.5), the historical average comprehensive deviation value is calculated based on the fault-free operation data in the past period of time (for example, 24 hours); the currently calculated comprehensive deviation value is compared with the set deviation value threshold range, if the comprehensive deviation value is less than or equal to the absolute threshold, and less than or equal to the relative threshold, it is determined that the model parameters are unbiased; if the comprehensive deviation value exceeds any threshold, the parameter adaptive adjustment model is started; the parameter adaptive adjustment model includes a feature weight adjustment module and a structure parameter adjustment module, the feature weight adjustment module adjusts the weight of the corresponding feature based on the deviation contribution degree, the deviation contribution degree is the proportion of the product of the relative deviation of a single feature dimension and the weight coefficient in the comprehensive deviation value, the feature dimension with a higher deviation contribution degree (for example, more than 30%) adjusts its weight coefficient in the deviation direction, for example, the entity data is greater than the virtual data and the deviation contribution degree is high, then the weight coefficient of the feature is increased (the adjustment amplitude is 5% to 10% of the original weight); the structure parameter adjustment module adjusts the structure parameters of the associated components, the structure parameters include the geometric parameter correction coefficient of the component, the material performance parameter correction coefficient, and the mechanical model parameter, for example, when the gear box vibration deviation is large, the elastic modulus correction coefficient and the damping coefficient of the gear box substructure model are adjusted, and the adjustment amplitude is 3% to 8% of the original parameter; after the adjustment is completed, the updated feature embedded twin model is generated.

[0051] Preferably, in one scenario, when step 233 is implemented, the updated feature-embedded twin model is first taken as a new feature-embedded twin model, and the process of step 231 is repeated to obtain new virtual state data output by the model, and the new virtual state data is dynamically compared with the latest collected and pre-processed device entity running state data to generate a new preliminary comparison result; based on the new preliminary comparison result, a new comprehensive deviation value is calculated by the deviation value calculation model of step 232, and the new comprehensive deviation value is compared with the threshold range of deviation values again; if the new comprehensive deviation value still exceeds the threshold range, the reason for the failure to reduce the deviation is analyzed, if the reason is that the feature weight adjustment is insufficient, the adjustment amplitude of the corresponding feature weight is increased (for example, the adjustment amplitude is increased to 10% to 15% of the original weight), if the reason is that the adjustment direction of the structure parameter is wrong, the structure parameter is adjusted in the opposite direction (for example, if the elastic modulus correction coefficient is originally increased, it is changed to be decreased), and a feature-embedded twin model that is updated again is generated; the above dynamic comparison and parameter adjustment process is repeated, and the change trend of the comprehensive deviation value, the parameter name and the adjustment amplitude are recorded after each adjustment; a maximum adjustment times threshold (for example, 10 times) is set, if the comprehensive deviation value is reduced to the threshold range within the maximum adjustment times, the adjustment is stopped; if the comprehensive deviation value still does not meet the standard after the maximum adjustment times are reached, an abnormal alarm is triggered, and the current model state and deviation data are recorded for subsequent technical analysis; when the comprehensive deviation value is in a reasonable range (i.e., less than or equal to the absolute threshold and less than or equal to the relative threshold), and the comprehensive deviation value in a plurality of (for example, 5) time periods is stable in the reasonable range, it is determined that the model parameters have been calibrated, and the feature-embedded twin model in this state is determined as the warehouse equipment digital twin prediction model, which can accurately simulate the running state of the equipment entity and provide a reliable basis for subsequent fault feature extraction and prediction.

[0052] Optionally, step 3 includes the following sub-steps: step 31, extracting vibration features, temperature features, energy consumption features, action response features and environment-related features of the equipment running from the warehouse equipment digital twin prediction model to form equipment multi-modal original features; step 32, performing spatial mapping conversion on the equipment multi-modal original features using a transfer learning mapping model to map different modal features to a pre-set unified feature space to generate modal unified features; step 33, performing redundant dimension orthogonal elimination and key feature weight strengthening on the modal unified features, and fusing equipment trend evolution data extracted from the warehouse equipment digital twin prediction model to generate a fault prediction core feature set.

[0053] Optionally, step 31 comprises the following sub-steps: step 311, extracting the vibration frequency of the equipment operation from the warehouse equipment digital twin prediction model to form the vibration feature; step 312, extracting the real-time temperature and temperature change rate of the key components of the equipment to form the temperature feature; step 313, extracting the energy consumption value and energy consumption fluctuation of the equipment to form the energy consumption feature; step 314, extracting the action response time and execution accuracy of the equipment to form the action response feature; step 315, extracting the environmental parameters of the equipment operation area and the associated data of the equipment state to form the environmental association feature; and step 316, performing multi-modal feature cross-dimension coupling coding on the vibration feature, temperature feature, energy consumption feature, action response feature, and environmental association feature to form the equipment multi-modal original feature.

[0054] Preferably, the specific implementation process of step 311 is as follows: first, locate the vibration-related feature output nodes in the warehouse equipment digital twin prediction model, which correspond to the vibration state simulation data of the key transmission components of the equipment (such as gearboxes, motor shafts, conveyor belt rollers, and shaft couplings), and contain vibration time-domain data and frequency-domain data output by the model based on structural mechanics simulation and real-time sensor data fusion; design vibration feature extraction rules based on common warehouse equipment vibration fault mechanisms (such as unbalanced vibration, misalignment vibration, and wear vibration), including time-domain feature extraction items and frequency-domain feature extraction items, the time-domain feature extraction items being peak value, effective value, and peak factor of the vibration signal, and the frequency-domain feature extraction items being fundamental frequency, multiple frequency, and harmonic component; call the vibration data interface of the model, intercept continuous vibration time-domain data according to the set time window (e.g., 1 second), convert the time-domain data into frequency-domain data through fast Fourier transform to generate vibration time-frequency domain data set; based on the vibration feature extraction rules, extract the vibration frequency (including the fundamental frequency value, main multiple frequency value, and energy proportion of the corresponding frequency), and amplitude (including peak amplitude, effective value amplitude, and amplitude fluctuation rate) of each key component from the vibration time-frequency domain data set, for example, extracting the fundamental frequency 20Hz, 2 multiple frequency 40Hz of the gearbox, and the corresponding peak amplitude 0.1mm and effective value amplitude 0.05mm; classify and organize the extracted vibration frequency and amplitude data according to the component identifier, label the component name, feature type (fundamental frequency / multiple frequency, peak value / effective value), and extraction time window corresponding to each vibration feature, and form a structured vibration feature.

[0055] Preferably, in the specific technical implementation of step 312, first, the temperature simulation data of the key components, including power components (motors, hydraulic pumps), transmission components (gearboxes, bearings), and electronic control components (controllers, sensors), are called from the warehouse equipment digital twin prediction model. The temperature simulation data is a continuous temperature sequence output by the model combining heat conduction simulation and real-time temperature sensing data. A temperature feature extraction period (for example, 5 seconds) is set, and the temperature sequence data of each key component is intercepted according to the period. The real-time temperature value in each period is calculated (taking the average of the temperature data in the period). Based on the real-time temperature values of adjacent periods, the temperature change rate is calculated (the difference between the real-time temperature of the next period and the real-time temperature of the previous period divided by the period length). For example, the real-time temperature of the motor in the previous period is 50°C, and the real-time temperature in the next period is 52°C, with a period length of 5 seconds. The temperature change rate is 0.4°C / s. At the same time, the temperature fluctuation features are extracted, including the maximum value, the minimum value, the fluctuation range (the difference between the maximum value and the minimum value), and the fluctuation coefficient (the ratio of the fluctuation range to the real-time temperature value) of the temperature data in each period. The real-time temperature, temperature change rate, and temperature fluctuation features of each key component are associated and labeled. The labeling information includes component identification, temperature measurement location (such as motor stator, bearing outer ring), and feature calculation period. The structured temperature features are formed to ensure that the temperature features can reflect the trend of the thermal state change of the components and provide support for overheating fault prediction.

[0056] Preferably, the specific implementation process of step 313 is as follows: first, determine the energy consumption data collection range of the warehouse equipment, covering the main power consumption data of the equipment, the branch energy consumption data of each functional module (power module, control module, execution module), the data derived from the output results of the energy consumption simulation module in the warehouse equipment digital twin prediction model, containing real-time energy consumption values and energy consumption time sequence change data; design energy consumption feature extraction dimensions, based on the energy consumption law of warehouse equipment operation (such as no-load energy consumption, load energy consumption, start-stop energy consumption difference), the extraction dimensions include instantaneous energy consumption value, average energy consumption value, energy consumption fluctuation feature, energy consumption and load correlation feature; according to the set statistical period (for example, 1 minute) to intercept the energy consumption time sequence data from the model, calculate the instantaneous energy consumption value (energy consumption peak value at a certain time in the period) and the average energy consumption value (arithmetic mean of energy consumption data in the period) in each period; get the energy consumption fluctuation feature by calculating the standard deviation and coefficient of variation of the energy consumption data in the period, the standard deviation reflects the dispersion degree of the energy consumption data, and the coefficient of variation is the ratio of the standard deviation to the average energy consumption value; based on the equipment load data (such as material weight, operation speed) output by the model, establish the correlation between energy consumption and load, calculate the unit load energy consumption (the ratio of average energy consumption value to average load) as the energy consumption and load correlation feature; sort the extracted energy consumption values (instantaneous energy consumption, average energy consumption), energy consumption fluctuation features (standard deviation, coefficient of variation), and energy consumption and load correlation features (unit load energy consumption) according to the statistical period and functional module, and form structured energy consumption features.

[0057] Preferably, in the specific technical implementation of step 314, first extract the simulation data related to the action response of the warehouse equipment digital twin from the equipment action response prediction model, including the instruction execution delay data and the action completion accuracy data output by the model based on kinematics simulation, and the action types involved include start-stop action, speed regulation action, positioning action and reversing action (such as lifting and horizontal movement of the stacker, start-stop and speed regulation action of the conveyor); set the action response feature extraction trigger condition, when the model simulates the equipment executing a specific action instruction, record the instruction issuing time, action start execution time, action completion time and action actual execution result synchronously; calculate the action response time, which includes the instruction delay time (the difference between the action start execution time and the instruction issuing time) and the action execution time length (the difference between the action completion time and the action start execution time), for example, the instruction issuing time is 10:00:00.000, the action start execution time is 10:00:00.005, and the action completion time is 10:00:01.005, then the instruction delay time is 5 milliseconds and the action execution time length is 1 second; calculate the execution accuracy, which is the deviation value of the action actual execution result from the instruction set target, for example, the instruction set target position of the positioning action is X = 1000 mm, and the actual execution result is X = 1000.2 mm, then the execution accuracy deviation is 0.2 mm, and the deviation rate (the ratio of the deviation value to the instruction set target) is calculated at the same time; classify and record the action response time (instruction delay time, action execution time length), execution accuracy (deviation value, deviation rate) of different action types, label the action instruction type, execution component and feature calculation time, and form a structured action response feature.

[0058] Preferably, the specific implementation process of step 315 is as follows: first, extract the environmental parameter data and equipment state data from the warehouse equipment digital twin prediction model, the environmental parameter data includes the running area temperature, humidity, dust concentration, corrosive gas concentration output by model fusion real-time environmental sensing data, the equipment state data includes the running parameters (speed, load) of the equipment, the component health state parameters (stress, wear amount); construct environment-state correlation analysis rules, the rules are formulated based on the environmental adaptability failure mechanism of warehouse equipment (such as high temperature leading to insulation aging, high humidity leading to component corrosion, dust leading to poor heat dissipation), and the correlation dimension of different environmental parameters and equipment state parameters is clear, for example, the correlation between temperature and motor temperature rise, the correlation between dust concentration and bearing wear rate, and the correlation between humidity and electrical component insulation resistance; based on the correlation analysis rules, time stamp alignment is performed on the environmental parameter data and equipment state data, and the environmental parameter statistical value (average value, maximum value) and equipment state parameter statistical value (average value, change amount) in each window are calculated according to the set analysis window (for example, 10 minutes); calculate the correlation characteristic value, the correlation characteristic value is the correlation coefficient of the environmental parameter statistical value and the corresponding equipment state parameter statistical value, and the ratio of the change amount of the environmental parameter and the change amount of the equipment state parameter, for example, the average value of the environmental temperature in a certain analysis window is 35℃, the average value of the motor temperature rise is 20℃, the correlation coefficient is 0.8, the change amount of the environmental temperature is 5℃, and the change amount of the motor temperature rise is 3℃, then the change amount ratio is 0.6; classify and arrange the correlation characteristic values of different correlation dimensions according to the environmental parameter type, equipment state parameter type and analysis window, mark the fault correlation type corresponding to the correlation characteristic (such as high temperature-temperature rise correlation, dust-wear correlation), and form a structured environment correlation feature.

[0059] Preferably, in the specific technical implementation of step 316, first, the structured vibration features, temperature features, energy consumption features, action response features, and environment-related features are collected, and data standardization processing is performed on various features to convert feature values of different dimensions to the same numerical interval (for example, 0 to 1), and the conversion method adopts maximum and minimum normalization, and the maximum and minimum values are determined based on the historical value range of various features, for example, the historical maximum amplitude of the vibration feature is 1 mm and the minimum value is 0 mm, so that a certain amplitude value of 0.1 mm is standardized to 0.1; a multi-modal feature cross-dimension coupling coding model is designed, which includes a feature dimension mapping layer, an association weight distribution layer, and a coupling coding layer, the feature dimension mapping layer maps various features to a preset fault feature dimension (such as a thermal fault dimension, a mechanical wear fault dimension, and an electrical fault dimension) according to the fault association attribute, for example, the temperature feature is mapped to the thermal fault dimension, and the vibration feature is mapped to the mechanical wear fault dimension; the association weight distribution layer assigns a corresponding association weight to each feature based on the feature-fault association degree matrix of step 131, and the association weight reflects the influence degree of the feature on different fault dimensions, for example, the association weight of the motor temperature feature on the thermal fault dimension is 0.9, and the association weight on the mechanical wear fault dimension is 0.3; the coupling coding layer adopts an attention mechanism, and different modal features mapped to the same fault dimension are weighted and fused to generate a fault dimension feature vector, and then all fault dimension feature vectors are spliced to form a coupling coding vector of a unified dimension; the standardized various features are input into the coupling coding model, and the feature dimension mapping, association weight distribution, and coupling coding processing are performed to generate a device multi-modal original feature, which contains the fusion information of various fault dimensions and can comprehensively reflect the multi-dimensional running state and fault association trend of the device.

[0060] Optionally, step 32 includes the following sub-steps: step 321, constructing a multi-modal feature migration network model, which includes a feature coding layer, a modal mapping layer, and a feature fusion layer; step 322, inputting the device multi-modal original feature into the feature coding layer to perform feature standardization and dimension unification processing to generate a coded feature; step 323, mapping the coded feature to a preset unified feature space through the modal mapping layer to generate a mapped feature; and step 324, performing weighted fusion processing on the mapped feature through the feature fusion layer to generate a modal unified feature.

[0061] Preferably, the specific implementation process of step 321 is as follows: first, the core design goal of the multi-modal feature migration network model is determined, that is, to solve the dimensional heterogeneity, dimensional difference and semantic fragmentation of multi-modal data such as warehouse equipment vibration features and temperature features, and to realize unified expression of features across modalities; a three-layer architecture system of the model is constructed, the feature coding layer is responsible for standardization and dimension regularization of the features, the modality mapping layer is responsible for mapping the coded features to a unified space, and the feature fusion layer is responsible for weighted fusion of the mapped features; the core function modules and parameter configurations of each layer are designed, the feature coding layer includes data cleaning module, dimension unification module and dimension alignment module, the modality mapping layer includes modality adaptation module, space mapping module and feature calibration module, and the feature fusion layer includes weight distribution module, correlation enhancement module and fusion output module; based on the statistical law of warehouse equipment fault features, the key parameters of each layer module are set, such as the dimension alignment target dimension of the feature coding layer (e.g. 256 dimensions), the unified feature space dimension of the modality mapping layer (e.g. 128 dimensions), and the initial weight distribution ratio of the feature fusion layer, to ensure that the model meets the migration and fusion requirements of multi-modal features of warehouse equipment; through modular assembly and parameter initialization, a complete multi-modal feature migration network model is formed, which can automatically call the corresponding processing module according to the modality type of the input feature to realize end-to-end feature migration and fusion.

[0062] Preferably, in the specific technical implementation of step 322, the multi-modal original features of the equipment are first classified and input according to the modality type (vibration features, temperature features, energy consumption features, action response features, and environment-related features), and each modality feature corresponds to an independent processing channel of the feature coding layer; the data cleaning module is started to adaptively fill in the missing values in each modality feature (mean value filling for continuous features and mode filling for discrete features), and to smooth the outliers (based on the 3σ principle to identify outliers and replace them with the neighborhood mean value) to generate clean feature data; the dimension unification module is used to perform standardization processing on the clean feature data, and the Z-score standardization method is used to convert the feature values into dimensionless standard scores based on the historical statistical data (mean value, standard deviation) of each modality feature, for example, the amplitude value of the vibration feature is 0.1 mm, the historical mean value is 0.05 mm, and the standard deviation is 0.02 mm, and the standardized score is 2.5; the dimension alignment module is called to expand or compress the dimensions of each modality feature after standardization to unify the dimensions of different modalities to the target dimension (e.g. 256 dimensions), the dimension expansion uses zero padding, and the dimension compression uses principal component analysis method to retain core information; the consistency of each modality processed feature is checked to verify whether the feature dimension, data format and numerical range meet the coding requirements, and the coded features corresponding to each modality are output after the verification, which retain the fault correlation semantics of the original modality and have unified dimensions and dimensions.

[0063] Preferably, the specific implementation process of step 323 is as follows: first, based on the feature requirements of the warehouse equipment failure prediction, the definition and attributes of the unified feature space are preset, the unified feature space is a high-dimensional semantic space, each dimension in the space corresponds to a failure associated semantic factor (such as a thermal failure associated factor, a mechanical wear associated factor, and an electrical failure associated factor), and the space dimension is set to a target dimension (for example, 128 dimensions); a modal-space adaptation matrix is constructed in the modal mapping layer, the rows of the matrix represent the dimensions of the encoded features, the columns represent the dimensions of the unified feature space, and the matrix elements represent the association strength of the encoded feature dimensions and the space semantic factors. The association strength is determined based on the feature failure association matrix, for example, the association strength of the encoded dimension of the temperature feature and the thermal failure associated factor is 0.9; the modal-space adaptation matrix is called by the modal adaptation module to perform semantic matching on the encoded features of each modal, and the space semantic dimension corresponding to each encoded dimension is determined; the space mapping module is started, and a mapping algorithm based on an attention mechanism is used to map the encoded features to the unified feature space according to the semantic matching results. Higher mapping weights are given to the encoded dimensions with higher association strengths in the mapping process, and the expression of the failure associated semantics is strengthened; the feature calibration module is called to calibrate the mapped modal features in the space position, calculate the semantic similarity of different modal features in the unified space, adjust the space position of the feature vector based on the similarity, and make the semantic related feature vectors closer; the calibrated feature vectors of each modal are taken as the mapped features, and output to the feature fusion layer. The mapped features not only retain the core information of the original modal, but also realize the semantic unification across the modal.

[0064] Preferably, in the specific technical implementation of step 324, first, the mapped features of each modality are obtained, and the initial allocation proportion of the preset modality weight of the feature fusion layer is extracted, the weight proportion is determined based on the contribution of each modality feature to the fault prediction, and the contribution is obtained through historical fault case data statistical analysis, for example, the weight of the vibration feature is 0.35, the weight of the temperature feature is 0.3, the weight of the energy consumption feature is 0.15, the weight of the action response feature is 0.1, and the weight of the environment associated feature is 0.1; start the weight distribution module, dynamically adjust the initial weight proportion based on the current device running state data, if the device is currently in a high load running state, increase the weight of the energy consumption feature and the action response feature (the adjustment amplitude is 10%-20% of the original weight), if the device running area environment parameter is abnormal, increase the weight of the environment associated feature; calculate the cross correlation coefficient between the mapped features of each modality through the correlation reinforcement module, the coefficient value represents the complementarity of different modality features in the fault correlation semantics, and the weight is adjusted again based on the cross correlation coefficient, the modality pair with higher correlation coefficient is given a cooperative weight, and the complementary information between the features is strengthened; the mapped features of each modality are fused by using a weighted summation algorithm, the fusion formula is modality unified feature = Σ(modality mapped feature × modality weight), and a regularization term is introduced to avoid overfitting; the fused features are normalized to map the feature values to the [0, 1] interval, and the final modality unified feature is generated, which integrates the core fault information of each modality and has strong semantic correlation and high discriminability.

[0065] Optionally, step 33 includes the following sub-steps: step 331, using a principal component analysis dimension reduction model to reduce the dimension of the modality unified feature in the kernel space, eliminating redundant dimensions, and generating dimension-reduced features; step 332, based on the evaluation result of the feature importance evaluation model, performing weight reinforcement processing on the key features in the dimension-reduced features to generate weight-reinforced features; step 333, extracting trend evolution data of the device running state from the warehouse device digital twin prediction model, the trend evolution data representing the change law of the device parameters over time; step 334, performing feature level deep embedding processing on the weight-reinforced features and the trend evolution data to generate a fault prediction core feature set.

[0066] Preferably, the specific implementation process of step 331 is as follows: first, the dimensionality of the modal unified feature and the data distribution characteristics are determined. The modal unified feature is a high-dimensional vector (e.g., 128 dimensions) containing multi-modal fusion information such as vibration and temperature. Some dimensions have information overlap or weak correlation with faults, and are redundant. A principal component analysis dimension reduction model suitable for warehouse equipment fault features is designed. The model includes a kernel function mapping module, a covariance matrix calculation module, and a principal component screening module. The Gaussian kernel function is selected as the kernel function. The modal unified feature is mapped to the kernel space through nonlinear mapping, and the nonlinear expression of the fault correlation feature is strengthened. Based on the feature vector in the kernel space, the covariance matrix is calculated. The rows and columns of the matrix correspond to the feature dimensions in the kernel space. The matrix elements represent the linear correlation degree between different dimensions, such as the correlation degree between the vibration correlation dimension and the temperature correlation dimension. The eigenvalue and eigenvector of the covariance matrix are solved by the eigenvalue decomposition algorithm. The size of the eigenvalue represents the information contribution degree of the corresponding principal component. An information retention threshold (e.g., 90%) is set. The eigenvalues are sorted from large to small, and the first N principal components whose eigenvalue proportion reaches the threshold are selected as the core principal components by accumulation. N is the target dimension after dimension reduction (e.g., 64 dimensions). The modal unified feature is projected into the feature vector space corresponding to the core principal components to generate the dimension-reduced feature. This feature eliminates the redundant correlation dimensions, retains the key information for fault prediction, and reduces the subsequent computational complexity.

[0067] Preferably, in the specific technical implementation of step 332, a feature importance evaluation model is first constructed, which integrates three functional modules of fault correlation analysis, information gain calculation, and historical contribution statistics, and adapts to the evaluation needs of warehouse equipment fault features; the dimension-reduced features are input into the fault correlation analysis module, combined with the feature fault correlation matrix of step 131, to calculate the average correlation strength of each feature dimension with each type of potential fault, and the higher the correlation strength, the higher the feature importance; through the information gain calculation module, the information contribution of each feature dimension to fault type discrimination is quantified, and the information gain is the difference in fault discrimination accuracy when including and not including the feature, and the larger the difference, the stronger the feature's ability to distinguish faults; the historical contribution statistics module is called to statistically analyze the participation of each feature dimension in correct prediction cases based on fault prediction data in the past period (i.e., the proportion of cases in which the feature dimension provides key support for prediction); a weighted summation algorithm is used to integrate the evaluation results of the three modules to generate the importance score of each feature dimension, with the weight distribution being 40% for fault correlation, 35% for information gain, and 25% for historical contribution; a threshold value (e.g., 0.6) is set for the importance score, and a reinforcement weight (e.g., 1.2-1.5 times the original weight) is given to key feature dimensions with a score higher than the threshold value, while the original weight or a proper reduction (e.g., 0.8 times the original weight) is maintained for non-key feature dimensions with a score lower than the threshold value; the dimension-reduced features are multiplied by the corresponding reinforcement weights to generate weight-reinforced features, which highlight the key information of close fault correlation and improve the discriminant sensitivity of the features.

[0068] Preferably, the specific implementation process of step 333 is as follows: first, define the extraction dimension of the equipment operation state trend evolution data, based on the evolution law of warehouse equipment failure, the extraction dimension includes parameter change trend, feature mutation frequency, steady state duration, decay rate correlation data, covering vibration frequency trend, temperature rise trend, energy consumption growth trend and other core evolution information; set the trend analysis time window (for example 1 hour), retrieve the equipment operation state time series data in the time window from the warehouse equipment digital twin prediction model, including real-time parameters, feature calculation results and virtual simulation output values of each key component; classify and process the time series data according to the extraction dimension, the parameter change trend is calculated by linear fitting algorithm, the slope of the time series data, the slope is positive, which represents the upward trend of the parameter, the greater the absolute value of the slope, the more obvious the trend, for example, the fitting slope of temperature time series data is 0.02℃ / min, which represents a slow warming trend; the feature mutation frequency is calculated by setting a mutation threshold (for example, the parameter change is more than 3 times the standard deviation), and the number of mutations in the time window is counted, for example, the mutation threshold of vibration amplitude is 0.03mm, and the number of mutations in the time window is 2; the steady state duration is the continuous time when the parameter value is in a reasonable range and the fluctuation is less than a set threshold (for example 5%); the decay rate correlation data is calculated by calculating the difference between the current parameter and the historical health state parameter, and the decay amount per unit time is obtained by combining the running time, for example, the current value of bearing clearance is 0.15mm, the historical health value is 0.1mm, and the running time is 100 hours, then the decay rate is 0.0005mm / hour; standardize the trend data of each dimension, unify the data format and dimension, and generate structured equipment trend evolution data, which can reflect the dynamic change law of equipment operation state and provide trend support for fault prediction.

[0069] Preferably, in the specific technical implementation of step 334, a feature hierarchy embedding model is first constructed, which includes a dimension alignment module, a hierarchy correlation module, and a fusion output module, which adapts to the heterogeneous fusion needs of the weight-enhanced features and the trend evolution data; the weight-enhanced features (e.g., 64 dimensions) and the device trend evolution data (e.g., 16 dimensions) are input into the dimension alignment module, and the trend evolution data is adjusted to the same dimension (e.g., 64 dimensions) as the weight-enhanced features through zero padding or dimension compression to ensure the dimension consistency of the hierarchy embedding; the hierarchy correlation module is started to construct a feature-trend correlation matrix, where the rows represent the dimensions of the weight-enhanced features, the columns represent the original dimensions of the trend evolution data, and the matrix elements represent the fault correlation tightness between the feature dimensions and the trend dimensions, for example, the correlation strength between the temperature-related feature dimension and the temperature rise trend dimension is 0.85; based on the correlation matrix, an attention mechanism is used to assign a trend attention weight to each feature dimension, the higher the correlation strength, the greater the corresponding trend data weight, realizing the directional correlation between features and trends; deep embedding is performed through a hierarchical fusion algorithm, and the fusion process is divided into basic layer embedding and enhanced layer embedding. The basic layer embedding uses element-wise multiplication to multiply the aligned trend data and the weight-enhanced features, and the enhanced layer embedding retains the original feature information through residual connection while integrating the trend correlation gain; the fused features are regularized to avoid overfitting, and the final fault prediction core feature set is generated, which contains not only high-recognizability static feature information but also dynamic trend evolution rules, and has comprehensive and accurate fault discrimination ability.

[0070] Optionally, step 4 includes the following sub-steps: step 41, based on the warehouse equipment failure mechanism data and historical fault cases, a multi-dimensional failure mode feature graph library is constructed, which includes feature graphs corresponding to different failure types; step 42, input the fault prediction core feature set into the probability evolution modeling model to generate fault type and evolution rate prediction results through feature graph cosine matching and evolution path time sequence analysis; step 43, fuse the matching degree of the feature graph cosine matching and the historical prediction accuracy of the probability evolution modeling model to calculate the reliability of the prediction result, and form a warehouse equipment fault prediction report through fault prediction element heterogenization packaging.

[0071] Optionally, step 41 comprises the following sub-steps: step 411, collecting failure mechanism literature of the warehouse equipment, and analyzing the causes, evolution process and characteristic performance of different failure types through a mechanism analysis model; step 412, sorting historical fault case data of the warehouse equipment, and extracting fault characteristics, environmental conditions and equipment states in each case; step 413, performing clustering operation on the failure mechanism analysis results and historical fault characteristics by using a feature clustering analysis model, to form a characteristic map corresponding to different failure types; step 414, performing classification indexing and structured storage on the characteristic map according to the failure types, and constructing a multi-dimensional failure mode characteristic map library.

[0072] Preferably, the specific implementation process of step 411 is as follows: first, the core failure type division dimensions of the warehouse equipment are determined, and the warehouse equipment (such as stacker, conveyor, AGV, etc.) is divided into four categories of mechanical transmission failure, electrical control failure, hydraulic and pneumatic failure, and environmental adaptation failure according to the structural characteristics and running scenarios of the warehouse equipment, and each category is further divided into specific failure subtypes (such as mechanical transmission failure including bearing wear, gear fracture, and conveyor belt deviation, etc.); a literature collection specification is formulated, focusing on failure mechanism research papers in the field of warehouse equipment, industry technical manuals, equipment manufacturer fault analysis reports, and failure judgment specifications in national standards / industry standards, to ensure the authority and pertinence of the data sources; a mechanism analysis model is constructed, which includes a literature deconstruction module, a causal chain analysis module, and a feature extraction module, the literature deconstruction module extracts failure-related key information (such as failure components, triggering conditions, and influencing factors) in the literature through natural language processing technology, the causal chain analysis module sorts out the logical relationship of “cause-evolution path-failure result”, for example, the causes of bearing wear include insufficient lubrication and excessive load, the evolution path is surface wear → gap enlargement → vibration aggravation → failure shutdown, and the feature extraction module quantifies the key characteristic performance in the failure process (such as vibration frequency change range, temperature threshold, and wear amount critical value); the mechanism analysis results of each type of failure are structured and sorted, to form a mechanism characteristic data set containing failure causes, evolution stages, characteristic parameters, and judgment standards, for example, the mechanism characteristic data set of gear fracture includes: causes (material fatigue, impact load), evolution stages (crack initiation → crack propagation → fracture), characteristic parameters (vibration frequency change, stress concentration value), and judgment standards (crack length ≥ 0.5 mm is a dangerous state).

[0073] Preferably, in the specific technical implementation of step 412, the collection range of historical failure cases is first defined, covering various types of failure events that have occurred during the operation of the warehouse equipment, and the collection channels include the failure records of the equipment maintenance management system, the on-site inspection report, the maintenance work order, and the failure diagnosis log. The case time span is not less than a set period of time (for example, 3 years) to ensure the timeliness and sufficient sample size of the data. A case data standardization and sorting specification is developed to clearly define the core fields that each case must include: equipment model and number, failure occurrence time, running conditions (load rate, running time, environmental parameters), failure site and failure type, failure precursor phenomenon, failure characteristic data (abnormal records of parameters such as vibration, temperature, and energy consumption), maintenance measures, and failure root cause confirmation results. The collected raw case data is pre-processed by cleaning, eliminating invalid cases with incomplete information (such as missing key characteristic data, unclear failure root cause), and recording errors, merging and deduplicating repeated cases to generate a clean case data set. A case feature extraction model is constructed to extract information in three dimensions of "failure characteristics-environmental conditions-equipment state". The failure characteristic dimension extracts specific parameter abnormal values, change trends, and mutation characteristics (such as vibration amplitude mutation value, temperature rise rate). The environmental condition dimension extracts environmental parameters such as temperature, humidity, and dust concentration at the time of failure. The equipment state dimension extracts information such as running time, load condition, and maintenance cycle before failure. The extracted three-dimensional features are associated with the failure type of the case to form a structured case feature data set. For example, the feature annotation of a certain conveyor belt deviation case is: failure characteristics (lateral vibration amplitude ≥ 0.3 mm, deviation amount ≥ 50 mm), environmental conditions (humidity 65%, dust concentration 0.8 mg / m 3 ), equipment state (continuous running for 8 hours, load rate 90%, 30 days since last maintenance).

[0074] Preferably, the specific implementation process of step 413 is as follows: first, integrate the mechanism feature dataset of step 411 and the case feature dataset of step 412, preliminarily group by failure type, and form a mixed feature set corresponding to each failure type (containing theoretical mechanism features and actual case features); construct a feature clustering analysis model, which includes a feature standardization module, a similarity calculation module, an adaptive clustering module, and a graph generation module. For numerical features (such as vibration frequency, temperature value) and categorical features (such as failure inducement, environmental type) in the mixed feature set, the feature standardization module uses different processing methods. Numerical features are converted into dimensionless data through Z-score standardization, and categorical features are converted into numerical vectors through one-hot encoding. The similarity calculation module uses an improved cosine similarity algorithm, introduces a feature weight factor (based on the importance of the feature to failure determination, such as mechanism feature weight accounting for 45% and case high-frequency feature weight accounting for 55%), and calculates the correlation similarity between different feature vectors. The adaptive clustering module is based on the density peak clustering algorithm, which does not need to pre-set the number of clusters, automatically identifies the core feature cluster under each failure type, and the core feature cluster is a feature set with a similarity higher than a set threshold (for example, 0.75), representing the typical feature performance of this failure type. The graph generation module associates each core feature cluster with the corresponding failure type and evolution stage, constructs the hierarchical structure of the feature graph, and the graph includes the top layer (failure category), the middle layer (failure subtype), and the bottom layer (core feature cluster). Each feature cluster is labeled with feature parameter range, occurrence probability, and association strength with other features. For example, the feature graph bottom layer of bearing wear failure includes: vibration feature cluster (2 times frequency amplitude 0.1-0.3 mm, peak factor ≥ 3.5), temperature feature cluster (temperature ≥ 70℃, temperature rise rate ≥ 0.5℃ / min), and running state associated feature cluster (continuous running ≥ 10 hours, lubrication period overdue).

[0075] Preferably, in the specific technical implementation of step 414, first design a multi-dimensional classification index system, construct an index based on four dimensions of failure type, equipment component, fault severity, and feature data type, the failure type index corresponds to major categories such as mechanical transmission and electrical control and subtypes, the equipment component index corresponds to specific components such as bearings, gears, and controllers, the fault severity index is divided into four levels of slight fault, general fault, serious fault, and fatal fault, and the feature data type index corresponds to feature categories such as vibration, temperature, and energy consumption; develop a structured storage specification for feature maps, adopt a storage structure of "map metadata + core feature data + association relationship data", the map metadata includes map number, failure type name, version number, update time, and data source (mechanism literature / case), the core feature data includes feature parameter name, parameter range, data type, and determination threshold, and the association relationship data includes the association probability of the failure type with other failure types and the causal relationship weight between features; select an appropriate storage architecture, use a relational database (such as MySQL) to store structured indexes and metadata, and use a graph database (such as Neo4j) to store the hierarchical relationship and association relationship of feature maps, to ensure efficient query and association analysis capabilities of data; establish an update and maintenance mechanism for stored data, set a regular update cycle (e.g., every 6 months), synchronize newly added failure mechanism research results and historical fault case data, iteratively optimize the original feature map, and record the change log of each update, including newly added features, adjusted parameter ranges, deleted obsolete features, etc.; through index construction and structured storage, form a complete multi-dimensional failure mode feature map library, support multi-dimensional quick retrieval according to failure type, equipment component, feature parameter, etc., and provide data support for subsequent fault matching and prediction.

[0076] Optionally, step 42 includes the following sub-steps: step 421, constructing a probability evolution modeling model containing a feature matching module, a path analysis module, and a rate calculation module; step 422, inputting the fault prediction core feature set into the feature matching module, performing cosine similarity matching with the feature maps in the multi-dimensional failure mode feature map library, and determining the failure type with the highest matching degree; step 423, analyzing the typical evolution path of the failure type through the path analysis module, and determining the evolution stage by fusing the current equipment state data; step 424, calculating the evolution rate of the failure type based on the feature change rate and historical evolution data through the rate calculation module, using a rate estimation model to generate a fault type and evolution rate prediction result.

[0077] Preferably, the specific implementation process of step 421 is as follows: first, the core design goal of the model is determined, that is, based on the core feature set of fault prediction, combined with the multi-dimensional failure mode feature spectrum library, the whole process prediction of "fault type matching-evolution stage positioning-evolution rate calculation" is realized, the model needs to adapt to the complex scene of multiple failure types and multiple running conditions of the warehouse equipment; a "three-module integrated" probability evolution modeling model architecture is constructed, the feature matching module is responsible for the accurate matching of core features and failure spectrum, the path analysis module is responsible for tracing the failure evolution path and positioning the current stage, and the rate calculation module is responsible for quantifying the dynamic rate of failure evolution, the three modules realize the bidirectional flow of features and results through the data interface; the core functional units and adaptation rules of each module are designed, the feature matching module includes feature standardization unit, similarity calculation unit and matching threshold judgment unit, aiming at the dimension characteristics (such as high dimension, nonlinearity) of the fault features of the warehouse equipment, the improved cosine similarity algorithm is adopted, and the feature weight factor (based on the feature importance evaluation result) is introduced; the path analysis module includes evolution path library unit, state matching unit and stage judgment unit, the evolution path library unit stores the "stage-feature" mapping relationship of various failure types, and the state matching unit realizes positioning through the comparison of the current equipment state data and the stage features; the rate calculation module includes rate factor extraction unit, historical data fitting unit and dynamic rate calculation unit, the rate factor covers feature change rate, working condition influence coefficient and environment correction coefficient, to ensure the dynamic adaptability of rate calculation; through the parameter linkage and data verification mechanism among the modules, a complete probability evolution modeling model is formed, for example, the matching result of the feature matching module is taken as the input of the path analysis module, and the stage positioning result of the path analysis module provides the stage benchmark parameter for the rate calculation module.

[0078] Preferably, in the specific technical implementation of step 422, the fault prediction core feature set is first input into the feature standardization unit of the feature matching module. The core feature set contains the static features and the dynamic features of trend evolution after weight enhancement (such as a 64-dimensional vector). The standardization unit adopts the same standardization rules (Z-score standardization) as the failure mode feature map library to convert the core features into dimensionless standard feature vectors, ensuring the consistency of the matching reference. The similarity calculation unit is called to load the standard feature maps of various failure types in the multi-dimensional failure mode feature map library (each map corresponds to a standard feature vector), and the improved cosine similarity between the standard feature vector and the core feature vector is calculated. The improvement point is to give higher calculation weight to the key dimensions in the core features after weight enhancement, for example, the weight of the vibration-related key dimension is 1.3 times that of the ordinary dimension, which improves the accuracy of matching. The similarity matching threshold is set, which is determined based on the statistical analysis results of the historical fault matching data of the warehouse equipment (for example, 0.7). The failure types higher than the threshold enter the candidate matching list, and the failure types lower than the threshold are determined as "no matching to known failure types", and the abnormal feature recording mechanism is triggered. The failure types in the candidate matching list are sorted by similarity, and the failure type with the highest similarity is selected as the prediction result. If there are multiple failure types with similar similarity (difference less than the set threshold, for example, 0.05), the final failure type is determined through the probability voting mechanism combined with the current equipment operating conditions (such as load rate, environmental parameters) and historical fault records. The failure type judgment result and the corresponding matching similarity value are output, for example, "predicted failure type: bearing wear, matching similarity: 0.82", which provides a basis for subsequent evolution analysis.

[0079] Preferably, the specific implementation process of step 423 is as follows: first, based on the failure type determined in step 422, the corresponding typical evolution path is retrieved from the evolution path library unit of the path analysis module. The evolution path is logically divided into "initial stage-development stage-danger stage-failure stage", and each stage clearly defines the core feature performance, state parameter range, and evolution trigger condition. For example, the typical evolution path of bearing wear: initial stage (normal lubrication, wear <0.1mm, stable vibration frequency) → development stage (lubrication attenuation, wear 0.1-0.3mm, slight fluctuation of vibration frequency) → danger stage (insufficient lubrication, wear 0.3-0.5mm, vibration frequency rises significantly) → failure stage (wear ≥0.5mm, bearing jam); collect current equipment state data, including real-time operating parameters (load rate, speed, environmental temperature and humidity), characteristic monitoring data (vibration, temperature, stress value), maintenance records (recent maintenance time, lubrication condition), form a current state data set; start the state matching unit, compare the current state data set with the characteristic parameters of each stage of the evolution path one by one, calculate the state matching degree, and the matching degree calculation covers three dimensions of characteristic parameter consistency, working condition adaptation, and trend consistency. For example, the current bearing vibration frequency is 25Hz, the development stage vibration frequency range is 20-30Hz, the consistency is 1.0, and the current load rate is 85% and the development stage typical load rate (80-90%) adaptation degree is 0.9; through the stage determination unit, the matching degrees of each dimension are integrated to determine the failure evolution stage of the current equipment. If the comprehensive matching degree of a stage is higher than the set threshold (for example, 0.8), the stage is directly located; if it is in the stage transition interval (the comprehensive matching degree is between the threshold values of the adjacent two stages), the evolution direction is determined combined with the characteristic change trend, for example, when transitioning from the development stage to the danger stage, the vibration frequency shows a continuous upward trend, and then it is located as "the end of the development stage, about to enter the danger stage"; the evolution stage positioning result and stage feature matching details are output, providing a stage benchmark for evolution rate calculation.

[0080] Preferably, in the specific technical implementation of step 424, first extract the core calculation factor from the rate factor extraction unit of the rate calculation module, based on the current failure type and evolution stage, the extracted factors include: characteristic change rate (such as daily growth rate of vibration frequency, temperature rise rate, monthly increment of wear amount), working condition influence coefficient (the higher the load rate, the larger the coefficient, for example, the coefficient is 1.2 when the load rate is 100%, and the coefficient is 0.8 when the load rate is 60%), environmental correction coefficient (influence of humidity, dust concentration on evolution rate, for example, the coefficient is 1.1 when the humidity is greater than 70%); call the historical data fitting unit to load the historical failure evolution data of the same evolution stage and similar working conditions of the failure type, establish the mapping model of "rate factor-evolution rate" by combining linear regression and nonlinear fitting, for example, the evolution rate model of bearing wear in the development stage is: wear rate = 0.02 mm / month x load coefficient x environment coefficient + vibration frequency growth rate x 0.01; input the current extracted rate factor value, solve the evolution rate through the dynamic rate calculation unit, and at the same time introduce a real-time correction mechanism, dynamically adjust the rate value based on the characteristic change data of the recent setting period (for example, 7 days), avoid the limitation of historical data, for example, the bearing wear increment in the last 7 days is 0.03 mm, which is higher than the historical same period average, the evolution rate is adjusted by a certain proportion (for example, 10%); convert the evolution rate into intuitive prediction indicators, such as "predicted time length from the end of the current development stage to the dangerous stage: 15 days" "predicted remaining life from the current stage to the failure stage: 30 days"; integrate the failure type determination result, evolution stage positioning result and evolution rate calculation result to generate a fault type and evolution rate prediction result set, including prediction conclusion, key supporting data, risk level prompt (such as high risk in dangerous stage, which needs to be maintained immediately).

[0081] Optionally, step 43 includes the following sub-steps: step 431, calculating the feature matching degree of the fault prediction core feature set and the matching feature map, taking the feature matching degree as the first reference index of the prediction reliability; step 432, statistics the historical prediction accuracy data of the probability evolution modeling model, taking the historical prediction accuracy data as the second reference index of the prediction reliability; step 433, using a weighted fusion operation model to perform weighted calculation on the first reference index and the second reference index, outputting the prediction confidence; step 434, performing normalized embedding encapsulation of the fault type, evolution rate and prediction confidence to form a standardized warehouse equipment fault prediction report.

[0082] Preferably, the specific implementation process of step 431 is as follows: first, the calculation object of the feature matching degree is determined, that is, the standard feature spectrum corresponding to the matching failure type of the fault pre-judgment core feature set determined in step 422, the core feature set contains the static features and trend evolution dynamic features (such as 64-dimensional vector) after weight enhancement, and the standard feature spectrum is the reference feature vector of the failure type in the multi-dimensional failure mode feature spectrum library; the improved cosine similarity algorithm consistent with step 422 is adopted to ensure the consistency of the calculation logic, and the feature weight factor is introduced in the algorithm, and higher weight is given to the key fault features (such as vibration frequency, temperature change rate, etc.) in the core feature set (for example, the weight of the key features is 1.2-1.5 times of the ordinary features), so as to improve the pertinence of the matching degree calculation; the feature dimension level matching details are output synchronously in the calculation process, that is, the matching scores of each dimension in the core feature set and the corresponding dimension in the standard feature spectrum, for example, the vibration frequency dimension matching score is 0.85, and the temperature dimension matching score is 0.78, which is convenient for subsequent result tracing; the dimension level matching scores are weighted and summed according to the weight to obtain the overall feature matching degree, and the matching degree value range is [0, 1], and the closer the value is to 1, the higher the matching degree of the core feature and the standard spectrum; the overall feature matching degree is directly used as the first reference index of the pre-judgment reliability, for example, the feature matching degree is 0.82, then the first reference index value is 0.82, and the dimension level matching details are recorded as auxiliary description, which provides basis for subsequent confidence analysis.

[0083] Preferably, in the specific implementation of step 432, first, the statistical range of historical prediction accuracy is defined, the probability evolution modeling model is used to statistically analyze the historical prediction cases of the current failure type, similar operating conditions (such as load rate, environmental temperature and humidity), and the same evolution stage, with a time span of not less than a set period (for example, 2 years) to ensure the representativeness of the samples; a calculation standard of historical prediction accuracy is established, the accuracy indicators include type prediction accuracy, evolution rate error rate, and stage positioning accuracy, type prediction accuracy = number of correctly predicted cases / total number of predicted cases x 100%, evolution rate error rate = | predicted evolution rate - actual evolution rate | / actual evolution rate x 100% (taking the average value), and stage positioning accuracy = number of cases with correct stage positioning / total number of predicted cases x 100%; an accuracy statistical model is constructed, which includes a case screening module, an indicator calculation module, and a weight distribution module, the case screening module screens target cases from the historical database according to the three-dimensional "failure type-condition-stage", the indicator calculation module calculates the three types of accuracy indicators respectively, and the weight distribution module distributes weights based on the influence of each indicator on the prediction reliability (for example, type prediction accuracy weight 0.4, evolution rate error rate weight 0.3, and stage positioning accuracy weight 0.3); the three types of accuracy indicators are normalized (converted to [0, 1] interval values), for example, type prediction accuracy 85% is normalized to 0.85, and evolution rate error rate 10% is normalized to 0.9 (the lower the error rate, the higher the normalized value); the historical prediction accuracy comprehensive value is calculated by weighted summation, which is the second reference indicator of prediction reliability, for example, type prediction accuracy 0.85, evolution rate error rate normalized value 0.9, and stage positioning accuracy 0.8, the second reference indicator value after weighted summation is 0.85 x 0.4 + 0.9 x 0.3 + 0.8 x 0.3 = 0.85.

[0084] Preferably, the specific implementation process of step 433 is as follows: first, a weighted fusion operation model is constructed, which includes a weight configuration module, a fusion calculation module, and a result calibration module, the weight configuration module sets the fusion weights of the first reference indicator (feature matching degree) and the second reference indicator (historical prediction accuracy) based on the practical experience of warehouse equipment failure prediction and the historical data verification results, and the default weight configuration is that the first reference indicator accounts for 0.55 and the second reference indicator accounts for 0.45, while supporting dynamic adjustment according to the actual scene (for example, when the sample size of historical cases is small, the weight of the first reference indicator is increased to 0.65); the first reference indicator value and the second reference indicator value are input into the fusion calculation module, and the initial confidence is calculated by the weighted summation formula, initial confidence = first reference indicator value x weight 1 + second reference indicator value x weight 2, for example, first reference indicator 0.82, second reference indicator 0.85, and initial confidence = 0.82 x 0.55 +0.85x0.45=0.8335; start the result calibration module, introduce the scene correction factor to calibrate the initial confidence, the scene correction factor is determined based on the special running state of the current device (such as overload running, abnormal environmental parameters, overdue maintenance period), for example, the current load rate of the device is 110% (overload), the correction factor is 0.98, and the correction factor is 1.0 when the environmental parameters are normal; the calibrated confidence = initial confidence x scene correction factor, the final confidence result is retained to two decimal places, the value range is [0, 1], for example, the calibrated confidence is 0.82, and the confidence level (high confidence: ≥0.8, medium confidence: 0.6-0.79, low confidence: <0.6) is output at the same time, which provides intuitive reference for report interpretation.

[0085] Preferably, in the specific technical implementation of step 434, first, the normalization packaging specification of the pre-judgment result elements is formulated, and the presentation format and standard terms of each element are specified. The fault type needs to use the unified naming in the multi-dimensional failure mode feature graph library (such as “mechanical transmission failure-bearing wear”). The evolution rate needs to present both quantitative values and intuitive expressions (such as “evolution rate: 0.03 mm / day, and it is expected to enter the dangerous stage in 15 days”). The pre-judgment confidence needs to be labeled with values and levels (such as “confidence: 0.82 (high confidence)”). A structured framework of the report is constructed, which includes four modules: device basic information, pre-judgment core results, supporting data explanation, and maintenance suggestions. The device basic information module includes device model, number, pre-judgment time, and running condition summary. The pre-judgment core result module presents fault type, evolution stage, evolution rate, and pre-judgment confidence. The supporting data explanation module includes feature matching details, historical precision statistical data, and scene correction factor explanation. The maintenance suggestion module gives targeted measures based on the pre-judgment results (such as “immediately supplement the lubricating grease, and recheck the bearing wear within 10 days”). The normalized pre-judgment elements are filled into the report template according to the structured framework by using the templating generation method, while supporting the embedding of visual charts (such as evolution trend curve graph and feature matching degree radar graph), which improves the readability of the report. The report is checked for integrity and standardization to check whether there are missing elements, format errors, and non-uniform terms. After passing the check, a standardized warehouse equipment fault pre-judgment report is generated. The report format supports common formats such as PDF and Word for export, and is simultaneously stored in the device fault pre-judgment database for subsequent traceability and data analysis.

[0086] As Figure 2As shown, this is a fault prediction device for warehousing equipment based on digital twins and multimodal feature transfer. It includes: a data acquisition and purification module, used to collect multi-source heterogeneous data on the entire lifecycle of the warehousing equipment, environmental perception data, and physical structure data; processing the raw data through an adaptive noise reduction and purification algorithm to generate a basic dataset for equipment twin modeling; a digital twin modeling module, used to construct a virtual model of the equipment based on the basic dataset using modal feature embedded mapping technology, and combining a virtual-real dynamic calibration mechanism to correct model parameter deviations, generating a digital twin prediction model for the warehousing equipment; a feature extraction and fusion module, used to extract multimodal features and trend evolution data of equipment operation from the digital twin prediction model of the warehousing equipment; processing these features through a feature transfer fusion algorithm and kernel space dimension reduction to focus on core fault-related information, generating a core feature set for fault prediction; and a fault prediction and report generation module, used to activate a failure mode Bayesian inference engine and a probabilistic evolution modeling model based on the core feature set for fault prediction, analyze potential fault types, infer fault evolution rates, calculate prediction confidence, and finally generate a standardized fault prediction report for the warehousing equipment.

[0087] like Figure 3 The diagram shows an electronic device comprising: a data storage unit for storing full lifecycle data of warehousing equipment, environmental perception data, physical structure data, basic dataset for equipment twin modeling, core feature set for fault prediction, and historical fault case data; a model running unit for running the digital twin prediction model, feature transfer fusion algorithm, kernel space dimension reduction model, failure mode Bayesian inference engine, and probabilistic evolution modeling model in the aforementioned warehousing equipment fault prediction method based on digital twins and multimodal feature transfer; a data interaction unit for collecting real-time operating data, environmental sensing data, and structural stress data of the warehousing equipment entity, synchronizing them to the model running unit, and simultaneously sending the generated warehousing equipment fault prediction report to the warehousing operation and maintenance management system; and a control unit for coordinating the working sequence of the data storage unit, model running unit, and data interaction unit, triggering the entire process of data acquisition, model calibration, feature processing, and prediction report generation to achieve accurate prediction of warehousing equipment faults.

Claims

1. A method for predicting faults in warehousing equipment based on digital twins and multimodal feature transfer, characterized in that, Includes the following steps: Step 1: Collect multi-source heterogeneous data, environmental perception data, and physical structure data of warehousing equipment throughout its entire life cycle, and perform adaptive noise reduction and purification to generate a basic dataset for equipment twin modeling. Step 2: Based on the basic dataset for equipment twin modeling, construct a digital twin prediction model for warehousing equipment through modal feature embedded mapping and virtual-real dynamic calibration; Step 3: Extract multimodal features and trend evolution data of equipment operation from the digital twin prediction model of warehousing equipment, and generate a core feature set for fault prediction through feature transfer fusion and kernel space dimension reduction. Step 4: Based on the core feature set of fault prediction, generate a fault prediction report for warehousing equipment through failure mode Bayesian inference and probabilistic evolution modeling. The fault prediction report for warehousing equipment includes potential fault types, evolution rates and prediction confidence.

2. The method for predicting warehouse equipment faults based on digital twins and multimodal feature transfer according to claim 1, characterized in that, Step 1 includes the following sub-steps: Step 11: Collect design parameter data, operation history data, maintenance record data, environmental sensor data, and structural stress data of the storage equipment to form a multi-source raw dataset of the equipment; Step 12: Perform adaptive noise filtering on the multi-source raw dataset of the device to remove abnormal fluctuation data and acquisition error data, so as to generate a denoised dataset of the device. Step 13: Perform feature fault correlation mining analysis on the equipment denoising dataset, retain data fields related to equipment faults, and form the basic dataset for equipment twin modeling by filtering and aggregating the data fields.

3. The method for predicting warehouse equipment faults based on digital twins and multimodal feature transfer according to claim 2, characterized in that, Step 11 includes the following sub-steps: Step 111: Perform structured parsing of the equipment design documents to obtain the equipment's structural dimensions, material parameters, rated load, rated speed, and rated power to form design parameter data; Step 112: Collect the equipment's operating parameters, start-stop count, operation duration, mean time between failures, and mean time to repair from the sampling equipment control system bus interface to form historical operating data; Step 113: Capture the equipment's maintenance records, component replacement information, fault handling process, and inspection item test values ​​retrieved from the maintenance management system interface to form maintenance record data; Step 114: Collect temperature, humidity, dust concentration, PM2.5 content, CO concentration, and SO2 concentration in the operating area of ​​the equipment from the distributed environmental sensor array to form environmental sensing data; Step 115: Analyze the stress changes, deformation data, vibration frequency and amplitude of key components of the equipment collected by the microelectromechanical system stress sensing module to form structural stress data; Step 116: Perform heterogeneous data fusion on design parameter data, operation history data, maintenance record data, environmental sensor data, and structural stress data to form a multi-source original dataset for the equipment.

4. The method for predicting warehouse equipment faults based on digital twins and multimodal feature transfer according to claim 2, characterized in that, Step 12 includes the following sub-steps: Step 121: Use the kernel density estimation statistical filtering model to detect outliers in the continuous data of the multi-source raw dataset of the equipment, and mark the data that exceeds the reasonable distribution range; Step 122: Adaptively replace the marked outliers using an interpolation completion model to output preprocessed data with continuous temporal sequence. Step 123: Use a hash-based deduplication model to remove redundant records from the preprocessed data to generate a device-denoised dataset.

5. The method for predicting warehouse equipment faults based on digital twins and multimodal feature transfer according to claim 2, characterized in that, Step 13 includes the following sub-steps: Step 131: Construct a feature fault correlation matrix. The rows of the matrix represent the potential fault types of the equipment, the columns represent the data fields in the equipment denoised dataset, and the matrix elements represent the correlation strength between the fields and the fault types. Step 132: Based on the feature fault correlation matrix, filter out data fields whose correlation strength meets the set conditions through the correlation strength threshold screening model; Step 133: Link and integrate the datasets corresponding to the filtered data fields, and remove unrelated data items to generate the basic dataset for device twin modeling.

6. The method for predicting warehouse equipment faults based on digital twins and multimodal feature transfer according to claim 1, characterized in that, Step 2 includes the following sub-steps: Step 21: Based on the physical structure data of the equipment in the equipment twin modeling basic dataset, perform equipment structure modeling through three-dimensional modal decomposition and reconstruction model to generate equipment structure twin model; Step 22: Embed the multi-source features in the basic dataset for equipment twin modeling into the equipment structure twin model through feature embedding mapping rules, establish a one-to-one mapping relationship between features and structural components, and generate a feature-embedded twin model. Step 23: Dynamically compare and calibrate the virtual state data output by the feature-embedded twin model with the real-time sensor data to correct the parameter deviation of the feature-embedded twin model, so as to generate a digital twin prediction model for warehousing equipment.

7. The method for predicting warehouse equipment faults based on digital twins and multimodal feature transfer according to claim 6, characterized in that, Step 21 includes the following sub-steps: Step 211: Perform structured semantic parsing on the physical structure data of the equipment in the equipment twin modeling basic dataset to extract component composition, assembly relationship, connection method, structural dimensions and material parameters to form equipment structural information; Step 212: Using three-dimensional modal decomposition to reconstruct the model, the equipment structure is decomposed into substructures according to functional modules based on the equipment structure information, and three-dimensional geometric modeling is performed on each substructure to generate a three-dimensional model of the substructure. Step 213: Perform topological combination and integration of the three-dimensional models of each substructure according to the assembly relationship to form a twin model of the equipment structure.

8. The method for predicting warehouse equipment faults based on digital twins and multimodal feature transfer according to claim 6, characterized in that, Step 22 includes the following sub-steps: Step 221: Extract multi-source feature metadata from the equipment twin modeling basic dataset, and divide it into structural features, operational features, and environmental features according to data type to obtain structural feature objects; Step 222: Construct a feature embedding mapping rule table to define a one-to-one mapping relationship between structural feature objects and corresponding components in the device structure twin model; Step 223: According to the feature embedding mapping rule table, embed the structural feature object into the corresponding component of the device structural twin model to generate a feature-embedded twin model.

9. The method for predicting warehouse equipment faults based on digital twins and multimodal feature transfer according to claim 6, characterized in that, Step 23 includes the following sub-steps: Step 231: Obtain the operating status data of the device entity collected through the real-time sensor network and dynamically compare it with the virtual status data output by the feature-embedded twin model; Step 232: Calculate the deviation value between the two types of data using the deviation value calculation model. If the deviation value exceeds the set threshold range, adjust the feature weights and structural parameters of the feature embedding twin model through the parameter adaptive adjustment model. Step 233: Repeat the dynamic comparison and parameter adjustment process until the deviation value is within a reasonable range to generate a digital twin prediction model for the warehousing equipment.

10. The method for predicting warehouse equipment faults based on digital twins and multimodal feature transfer according to claim 1, characterized in that, Step 3 includes the following sub-steps: Step 31: Extract vibration characteristics, temperature characteristics, energy consumption characteristics, action response characteristics and environmental correlation characteristics of the equipment from the digital twin prediction model of the warehousing equipment to form the original multimodal characteristics of the equipment; Step 32: Use a transfer learning mapping model to perform spatial mapping transformation on the original multimodal features of the device, mapping different modal features to a preset unified feature space to generate unified modal features; Step 33: Perform redundant dimension orthogonal elimination and key feature weight enhancement on the modal unified features, and integrate the equipment trend evolution data extracted from the digital twin prediction model of warehousing equipment to generate a core feature set for fault prediction.

Images