Equipment state monitoring and abnormity early warning method based on multi-source data fusion

Abstract

The invention discloses an equipment state monitoring and abnormity early warning method based on multi-source data fusion, and particularly relates to the technical field of monitoring and early warning. The method comprises the following steps: segmenting original monitoring data according to a sampling frequency difference to generate a time sequence subset, extracting transient characteristics of the time sequence subset, constructing an asynchronous interference identification index, and performing time domain dislocation evaluation and correction on the transient characteristics to obtain a synchronous characteristic set; performing holographic spectrum conversion on the synchronous feature set, constructing a holographic spectrum feature initial matrix, and screening through a feature sparse discrimination mechanism to obtain a sparse feature matrix; and training the multi-modal mapping model by using the target equipment fault diagnosis historical data, and outputting an abnormal risk score. And generating a state label of the target equipment and a corresponding early warning result according to a preset abnormal threshold value, thereby realizing high-precision and low-delay prediction and alarm of the equipment fault under the complex unsteady state working condition.

Description

Technical Field

[0001] The present invention relates to the technical field of monitoring and early warning, and more specifically, to a method for equipment status monitoring and abnormal early warning based on multi-source data fusion. Background Art

[0002] Existing equipment status monitoring methods mostly rely on single sensor signals or feature extraction strategies under steady-state working conditions, and it is difficult to adapt to the time-series inconsistency and feature expression complexity problems of multi-source heterogeneous data during non-steady-state operation. In the scenario of multi-source data fusion, due to differences in sampling frequencies and response delays of different sensors, fault information often appears misaligned in the time domain, making it difficult to align transient features and resulting in serious fusion distortion. At the same time, although current spectrum feature construction methods can integrate multi-source information, the effective feature components are sparsely distributed, and key fault information is easily submerged by redundant features, resulting in a decline in diagnostic performance and waste of computing resources.

[0003] In order to solve the above problems, a technical solution is provided now. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a method for equipment status monitoring and abnormal early warning based on multi-source data fusion to solve the problems raised in the above background art.

[0005] To achieve the above object, the present invention provides the following technical solutions: A method for equipment status monitoring and abnormal early warning based on multi-source data fusion, comprising the following steps: Collect the original monitoring data of the target equipment under non-steady-state working conditions, and perform segmentation processing according to the sampling frequency difference to obtain multiple time series subsets; Extract the transient features of each time series subset according to the operating state characteristics of the target equipment under non-steady-state working conditions, and construct a multi-scale transient feature set; Establish an asynchronous interference recognition index, perform time-domain misalignment evaluation and correction on each transient feature, and generate a corrected synchronous transient feature set; Perform holographic spectrum transformation on the synchronous transient feature set to obtain an initial holographic spectrum feature matrix, and use a feature sparse discrimination mechanism to evaluate the sparsity of the initial holographic spectrum feature matrix and remove redundant features to form a sparse feature matrix; Establish a multi-modal mapping model between the sparse feature matrix and the status of the target equipment based on the historical data of the target equipment fault diagnosis, input the sparse feature matrix, and obtain an abnormal risk score value; Determine the status label of the target equipment according to the comparison between the abnormal risk score value and a preset abnormal threshold, and perform an abnormal early warning result of the status of the target equipment based on the status label of the target equipment.

[0006] In a preferred embodiment, the original monitoring data of the target device under unsteady operating conditions is collected and segmented according to the sampling frequency differences to obtain multiple time series subsets, specifically: Obtain the original monitoring data of multiple different types of sensors of the target device under unsteady operating conditions, and determine the sampling interval of the time series according to the sampling frequency of each sensor; Perform time series segmentation on the original monitoring data respectively according to the sampling interval to form independent time series subsets corresponding to the sampling frequencies of each type of sensor; Mark each independent time series subset according to a unified time reference to obtain multiple time series subsets.

[0007] In a preferred embodiment, according to the operating state characteristics of the target device under unsteady operating conditions, the transient characteristics of each time series subset are extracted to construct a multi-scale transient feature set, specifically: Perform sliding slicing on each time series subset according to a fixed window length to generate multiple time window segments; In each time window segment, extract the time-domain variation characteristics including signal fluctuation mutation, edge jump, short-term wave peak and periodic perturbation to construct a single-scale transient feature vector; Unify the coding of multiple single-scale transient feature vectors generated under different window lengths, and combine them according to the time window scale hierarchy to generate a multi-scale transient feature set.

[0008] In a preferred embodiment, an asynchronous interference recognition index is established to perform time-domain misalignment evaluation and correction on each transient feature to generate a corrected synchronous transient feature set, specifically: Calculate the time-domain offset between each single-scale transient feature vector and a preset synchronization reference, and determine the degree of asynchronous interference according to the time-domain offset; Establish an asynchronous interference recognition index according to the degree of asynchronous interference, and use the asynchronous interference recognition index to perform time-domain misalignment correction on each single-scale transient feature vector; Perform time-domain reconstruction on all single-scale transient feature vectors that have completed time-domain misalignment correction to generate a corrected synchronous transient feature set.

[0009] In a preferred embodiment, perform holographic spectrum transformation on the synchronous transient feature set to obtain an initial holographic spectrum feature matrix, and use a feature sparse discrimination mechanism to evaluate the sparsity of the initial holographic spectrum feature matrix and remove redundant features to form a sparse feature matrix, specifically: Perform holographic spectrum transformation on the corrected single-scale transient feature vectors in the synchronous transient feature set to obtain the frequency spectrum diagrams corresponding to each corrected single-scale transient feature vector respectively; Unfold each frequency spectrogram according to the unified frequency axis coordinates and form an initial high-dimensional frequency feature space to generate an initial holographic spectrum feature matrix; Calculate the sparsity of each spectral feature component in the initial holographic spectrum feature matrix, and determine the contribution value of each spectral feature component in the initial holographic spectrum feature matrix; Set a sparsity threshold, use the sparsity threshold to screen and remove redundant features in the initial holographic spectrum feature matrix, extract spectral feature components with contribution values higher than the sparsity threshold, and form a sparse feature matrix.

[0010] In a preferred embodiment, establish a multi-modal mapping model between the sparse feature matrix and the target device state based on the target device fault diagnosis historical data, input the sparse feature matrix, and obtain an abnormal risk score value. Specifically: Construct a data set containing the target device historical operating state data and the corresponding state labels, standardize the historical operating state data, and form a historical sparse feature sample set; Perform sample pairing processing on the historical sparse feature sample set and the corresponding state labels to establish a set of correspondence data pairs between the sparse features and the state labels; Construct a multi-modal mapping model based on the historical sparse feature sample set and the corresponding relationship data pair set; the input of the multi-modal mapping model is the sparse feature matrix, and the output of the multi-modal mapping model is the abnormal risk score value; Input the sparse feature matrix into the multi-modal mapping model, perform mapping inference operations, and obtain an abnormal risk score value corresponding to the real-time operating state of the target device.

[0011] In a preferred embodiment, determine the state label of the target device according to the comparison between the abnormal risk score value and the preset abnormal threshold, and output the state abnormal warning result of the target device based on the state label of the target device. Specifically: Preset a first abnormal threshold and a second abnormal threshold; the first abnormal threshold is less than the second abnormal threshold; When the abnormal risk score value is less than or equal to the first abnormal threshold, determine that the state label of the target device is in a normal state; When the score value is greater than the first abnormal threshold and less than the second abnormal threshold, determine that the state label of the target device is in a warning state; When the score value is greater than or equal to the second abnormal threshold, determine that the state label of the target device is in an abnormal state; Trigger corresponding warning signals according to the state label of the target device. No warning is triggered for the normal state, a primary alarm prompt is triggered for the warning state, and an emergency stop instruction is triggered for the abnormal state and a fault report is generated.

[0012] Technical effects and advantages of a device status monitoring and anomaly warning method based on multi-source data fusion according to the present invention: Through time series segmentation based on sampling frequency differences, multi-scale transient feature extraction, and correction based on asynchronous interference recognition metrics, precise time-domain alignment of multi-sensor non-steady-state fault signals is achieved, reducing diagnostic delays and misjudgments caused by signal misalignment; at the same time, combined with holographic spectrum conversion and sparse discrimination mechanisms, the high-dimensional spectrum space is effectively compressed, high-sensitivity frequency components are retained, redundant information is reduced, and efficient capture of key faults is realized; based on a multi-modal mapping model, a mapping relationship between sparse features and device status is established, an anomaly risk score value is output, and status label determination and hierarchical warning are carried out in combination with a preset anomaly threshold, realizing automation from raw data to anomaly warning. It improves the accuracy and robustness of device fault identification, reduces the computational burden, has good real-time performance and scalability, and is applicable to online monitoring and warning under various complex working conditions. At the same time, it also has the dual response ability to sudden faults and progressive faults. Description of the Drawings

[0013] Figure 1 It is a schematic diagram of a device status monitoring and anomaly warning method based on multi-source data fusion according to the present invention. Detailed Embodiments

[0014] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0015] Embodiment:

[0016] Figure 1 A device status monitoring and anomaly warning method based on multi-source data fusion according to the present invention is given, which includes the following steps: Collect the original monitoring data of the target device under non-steady-state working conditions, and perform segmentation processing according to the sampling frequency differences to obtain multiple time series subsets; According to the operating state characteristics of the target device under non-steady-state working conditions, extract the transient features of each time series subset, and construct a multi-scale transient feature set; Establish an asynchronous interference recognition metric, evaluate and correct the time-domain misalignment of each transient feature, and generate a corrected synchronous transient feature set; Perform holographic spectrum conversion on the synchronous transient feature set to obtain an initial holographic spectrum feature matrix, and use the feature sparse discrimination mechanism to evaluate the sparsity of the initial holographic spectrum feature matrix and remove redundant features to form a sparse feature matrix; Based on the historical data of the target device's fault diagnosis, a multi-modal mapping model between the sparse feature matrix and the target device's state is established. Input the sparse feature matrix to obtain the abnormal risk score value; According to the comparison between the abnormal risk score value and the preset abnormal threshold, determine the state label of the target device, and based on the state label of the target device, issue an early warning result for the state abnormality of the target device.

[0017] Specifically, collect the original monitoring data of the target device under unsteady working conditions, and perform segmentation processing according to the sampling frequency difference to obtain multiple time series subsets, including: Obtain the original monitoring data of multiple different types of sensors of the target device under unsteady working conditions, and determine the sampling interval of the time series according to the sampling frequency of each sensor; Specifically, deploy multiple types of sensors to collect the working state of the target device in real time under unsteady working conditions. The types of sensors include but are not limited to: A structure vibration sensor for monitoring the acceleration change information on the surface of the target device; A temperature sensor for monitoring the heat accumulation during operation; A current sensor for recording the current dynamics of the drive circuit; A stress-strain sensor for obtaining the mechanical response characteristics of the materials at key parts; A displacement sensor for monitoring the relative movement of structural components.

[0018] Each type of sensor has an independent sampling mechanism. The sampling frequency refers to the number of data points generated by the sensor per unit time, and different types of sensors have different sampling frequency settings.

[0019] For example: The structure vibration sensor uses high-frequency sampling, and the single sampling interval is a short time period; since the temperature sensor has a slow data change speed, it uses low-frequency sampling and the sampling interval is relatively long; the current sensor uses medium-frequency sampling, and the sampling interval is between low-frequency sampling and high-frequency sampling.

[0020] Based on the reciprocal relationship between the sampling frequency and the sampling interval, calculate the time series sampling interval for each type of sensor. The sampling interval can be calculated by the following method: If the sampling frequency is set as the number of data points per unit time, then the sampling interval is the unit time divided by the number of data points, which assigns accurate time interval parameters to the data of each type of sensor.

[0021] Perform time series segmentation on the original monitoring data according to the sampling interval respectively to form independent time series subsets corresponding to the sampling frequencies of each sensor; Specifically, under the premise of the sampling interval of each type of sensor, the original monitoring data is segmented and processed on the time axis according to the fixed sampling interval. The data collected by each type of sensor is divided into several time segments, and each segment represents a set of observations in a continuous time window. The number of data in the time window is automatically determined according to the sampling interval of the sensor, forming a uniformly distributed sequence of data points.

[0022] Exemplarily, the sampling interval of the structural vibration sensor is a short time period, and the monitoring data is segmented into high-density time window segments; The sampling interval of the temperature sensor is relatively long, and the segmented time segments contain fewer data points.

[0023] The segmentation operation ensures that each type of sensor forms a time series subset with clear structure and stable sampling interval, which helps to construct multi-scale windows in the transient feature extraction stage.

[0024] Mark each independent time series subset according to a unified time reference to obtain multiple time series subsets; Specifically, in order to ensure that each independent time series subset has a unified time index structure, each independent time series subset needs to be marked with a unified time reference, specifically: For each independent time series subset, the corresponding theoretical timestamp sequence is generated by recursive calculation according to the sampling start time and sampling interval of the sensor to which it belongs. Each theoretical timestamp is generated by taking the sampling start time as the initial time, continuously accumulating the fixed sampling interval, and obtaining the theoretical appearance time corresponding to each data point in turn, and obtaining a single-channel time index list in the order of the sampling point number.

[0025] A unified time base is constructed, and the monitoring start time of the non-steady-state operation phase of the target device is used as the unified starting point to map and transform the theoretical timestamps of each sensor. The mapping and transformation process includes: adding the start offset time of the corresponding channel to each theoretical timestamp to convert it into a global timestamp under a unified reference coordinate; through the global timestamp, the time series subsets of different sensors are processed with consistent time indexes.

[0026] After the mapping is completed, each time series subset retains the original sensor properties, but has comparability and time alignment capabilities with other sensor data in the time axis dimension. Ultimately, multiple time series subsets with a unified time tag structure are formed, providing a structurally consistent data foundation for multimodal fusion, feature alignment, and interference correction processing.

[0027] Specifically, according to the operating state characteristics of the target equipment under non-steady-state conditions, the transient features of each time series subset are extracted to construct a multi-scale transient feature set, including: For each subset of time series, perform sliding slicing processing according to a fixed window length to generate multiple time window segments; Specifically, for each subset of time series that has completed unified time marking, perform data slicing operations according to a preset sliding window strategy. The sliding window strategy includes two key parameters: window length, which defines the time span covered by each segment; and sliding step, which defines the start time interval between two adjacent segments.

[0028] The execution method of the sliding window operation is as follows: starting from the first data point of each subset of time series, extract a continuous number of data points as the first window segment, then slide the window forward by one step, and extract the same length of data to form the second window segment until the end of the time series.

[0029] The selection of the window length is usually determined according to the typical change period of the sensor data. If the structural response reflected by the vibration data has multiple periodic components, multiple window lengths of different scales need to be set to adapt to different dynamic response speeds. The sliding-generated time window segments have local time continuity and can capture the local response change characteristics in the non-steady process.

[0030] In each time window segment, extract the time-domain change characteristics including signal fluctuation mutation, edge jump, short-term wave peak, and periodic perturbation, and construct a single-scale transient feature vector; Specifically, in each time window segment, perform local transient feature extraction processing based on the time-domain structure: Signal fluctuation mutation feature: Calculate the difference sequence between adjacent data points within the current window segment, and record the maximum difference amplitude, average fluctuation amplitude, and the number of fluctuations in the direction change to reflect the degree of local sudden perturbation.

[0031] Edge jump feature: Detect the change in the edge slope in the data sequence, and evaluate the direction and intensity of the jump trend through the rising gradient and falling gradient between two consecutive points or multiple points.

[0032] Short-term wave peak feature: Count the number of extreme points exceeding the preset threshold within the time window segment, including the number of positive wave peaks and negative valleys, and record the amplitude and position index respectively to identify abnormal fluctuation points.

[0033] Periodic perturbation feature: Adopt the envelope analysis method of the signal within the window to measure the main frequency band position and energy distribution trend of the periodic perturbation, and reflect the existence of short-period characteristics.

[0034] The above various transient change characteristics will be uniformly encoded to form a vector structure containing fixed-length data items, which is defined as a single-scale transient feature vector. Each single-scale transient feature vector completely corresponds to a time window segment, retaining the original time index and channel identification information.

[0035] Unify the encoding of multiple single-scale transient feature vectors generated under different window lengths, and combine them according to the time window scale hierarchy to generate a multi-scale transient feature set; Specifically, in order to fully capture the non-steady operation characteristics at different time granularities, repeated sliding slicing and transient feature extraction are performed on the same time series subset with multiple different window lengths. The feature vectors extracted under different window lengths have complementary scales: A smaller window length can capture short-period shock features and is suitable for identifying minor perturbations; A larger window length can extract stable trend changes and low-frequency variations, and is suitable for identifying slow aging or wear patterns.

[0036] For the single-scale transient feature vectors obtained under all different window lengths, perform unified dimensional encoding to ensure that the single-scale transient feature vectors under all window lengths have a consistent structural format, and construct a multi-level feature structure with scale labels, that is, a multi-scale transient feature set. The unified encoding operation includes but is not limited to: keeping the dimension number of all single-scale transient feature vectors, the arrangement order of various transient features, and the mapping method of the value range of feature values consistent.

[0037] Specifically, establish an asynchronous interference recognition index, perform time-domain misalignment evaluation and correction on each transient feature, and generate a corrected synchronous transient feature set, including: Calculate the time-domain offset between each single-scale transient feature vector and a preset synchronization reference, and determine the degree of asynchronous interference according to the time-domain offset; Specifically, based on multiple time series subsets with unified time markings and the single-scale transient feature vectors generated therefrom, select a sensor channel with high time resolution as the synchronization reference channel. Exemplarily, select a high-sampling-frequency channel composed of vibration sensors as the reference channel.

[0038] For the single-scale transient feature vector of each non-reference channel, perform a sliding comparison with the single-scale transient feature vector in the reference channel in chronological order. By calculating the similarity change trend between the single-scale transient feature vectors of the non-reference channel and the reference channel in adjacent time periods, find the position points where local feature alignment occurs. The time offset value corresponding to the position point is the time-domain offset of the non-reference channel relative to the reference channel.

[0039] The time-domain offset is represented by the sliding window displacement value in integer form, which reflects the response delay degree between channels when observing the same fault feature. After measuring the time-domain offset for all channels respectively, summarize the time-domain offsets of each channel, construct the asynchronous interference degree matrix, and establish the asynchronous interference degree matrix. The asynchronous interference degree is the quantification of the relative offset degree between multi-channel signals. The larger the time offset value, the more serious the asynchronous interference between channels. The fault feature refers to the cross-channel transient response feature caused by local faults or disturbances of the target device in the same physical time period when multiple different channels record the operation data of the target device.

[0040] Establish an asynchronous interference recognition index according to the asynchronous interference degree, and use the asynchronous interference recognition index to perform time-domain misalignment correction on each single-scale transient feature vector; Specifically, according to the asynchronous interference degree matrix, generate the corresponding asynchronous interference recognition index for each channel. The asynchronous interference recognition index is defined as the product of the time-domain offset of each channel and the channel sampling period, which represents the time value that the current channel data needs to be shifted forward or backward on the unified time basis.

[0041] Taking each single-scale transient feature vector as the processing unit, correct its time index information, that is, adjust its original timestamp according to the time value that needs to be shifted forward or backward calculated by the asynchronous interference recognition index. If the time value that needs to be shifted is a positive number, then shift backward by the corresponding number of sampling intervals; if the time value that needs to be shifted is a negative number, then adjust forward by the corresponding time step. Ensure that when the same physical event occurs for all channels, the positioning of the single-scale transient feature vectors on the time axis is maximally aligned.

[0042] The time-domain misalignment correction operation does not change the original eigenvalue structure, only performs translation processing on the time label, and retains the original feature vector content and sampling source information.

[0043] Perform time-domain reconstruction on all single-scale transient feature vectors that have completed time-domain misalignment correction to generate a set of synchronized transient features after correction; Specifically, reorganize all single-scale transient feature vectors that have completed time-domain misalignment correction processing under the unified time axis, and perform time series reconstruction according to their newly corrected time index. The specific reconstruction process is as follows: Establish a global time index structure with the unified time basis as the horizontal axis; Attach each single-scale transient feature vector to the corresponding time node according to the corrected timestamp; Number and organize the single-scale transient feature vectors from different channels and different window lengths at the same time node to construct a complete set of synchronized feature set units.

[0044] After the reconstruction is completed, the positions of all single-scale transient feature vectors on the time axis have been unified, forming a set of synchronous transient feature sets with consistent structures, comparable times, and corresponding channels, providing strictly aligned data input conditions for spectral transformation and fusion modeling.

[0045] Specifically, perform a holographic spectrum transformation on the synchronous transient feature set to obtain an initial holographic spectrum feature matrix, and use a feature sparse discrimination mechanism to evaluate the sparsity and remove redundant features of the initial holographic spectrum feature matrix, forming a sparse feature matrix, including: Perform a holographic spectrum transformation on the single-scale transient feature vectors after calibration in the synchronous transient feature set, and respectively obtain the frequency spectrograms corresponding to each single-scale transient feature vector after calibration. Specifically, for each single-scale transient feature vector that has completed time-domain alignment processing, perform a frequency conversion operation to map its expression form from the time domain to the frequency domain. In the conversion process, a preset window function is used to process the boundary of the feature vector, and then through frequency response mapping, the energy distribution characteristics in the frequency dimension are extracted. Each frequency response output corresponds to a frequency spectrogram, and each item in the frequency spectrogram represents the amplitude response degree at that frequency position.

[0046] The frequency spectrogram reflects the response intensity of the single-scale transient feature to different frequency components in the frequency domain, and can reveal weak oscillations or periodic perturbations in the time domain.

[0047] Unfold each frequency spectrogram according to the unified frequency axis coordinates and form an initial high-dimensional frequency feature space to generate an initial holographic spectrum feature matrix; Specifically, linearly unfold each frequency spectrogram according to the predefined frequency axis coordinates, and arrange the response amplitudes corresponding to each frequency point in a vector structure. The vector lengths of all frequency spectrograms are the same, and the frequency coordinates are aligned to form a frequency response matrix with a unified coordinate system.

[0048] After all the frequency spectrograms are unfolded, arrange them row by row according to the time nodes and channel order to form a frequency feature matrix. The number of rows of the frequency feature matrix is equal to the number of single-scale transient feature vectors, and the number of columns is equal to the number of frequency components covered in the frequency spectrogram. The frequency feature matrix is the initial holographic spectrum feature matrix and is the input basis for sparse feature discrimination and redundant information removal.

[0049] Calculate the sparsity degree of each spectral feature component in the initial holographic spectrum feature matrix, and determine the contribution value of each spectral feature component in the initial holographic spectrum feature matrix; Specifically, each column of frequency components in the initial holographic spectrum feature matrix corresponds to a spectrum feature component. For each column of frequency components in the initial holographic spectrum feature matrix, the distribution density of non-zero response amplitudes is statistically analyzed. The specific processing method is as follows: count the number of data points with non-zero response amplitudes in each column of frequency components, and calculate the ratio with the total number of samples to obtain the response density value of the current spectrum feature component.

[0050] Jointly measure the amplitude variance and concentration degree of each column of spectrum feature components, and calculate the contribution of each column of spectrum feature components to the difference degree of feature distribution in the overall spectrum space. Specifically: for each column of frequency components, first calculate the response amplitude variance in all samples; then calculate the response concentration degree of the frequency components under different target device state labels, such as by calculating the entropy value; after normalizing the response amplitude variance and entropy value, perform weighted summation to obtain the feature discrimination contribution value of the frequency component in the spectrum space. The higher the feature discrimination contribution value, the stronger the discrimination ability of the response change of the frequency point between different target device states.

[0051] Combine the response density of the spectrum feature components with the feature discrimination contribution value to form a sparse discrimination index sequence for feature screening operations.

[0052] Set a sparsity threshold, use the sparsity threshold to screen and remove redundant features in the initial holographic spectrum feature matrix, extract the spectrum feature components with contribution values higher than the sparsity threshold, and form a sparse feature matrix; Specifically, set a sparsity threshold to determine which spectrum feature components have sufficient diagnostic value. The sparsity threshold is determined according to the distribution of historical data samples of the target device, usually the median of the distribution of feature discrimination contribution values. The setting principle of the sparsity threshold is to retain the spectrum feature components that are representative features for identifying the state of the target device. Retain the spectrum feature components with feature discrimination contribution values higher than the sparsity threshold in the sparse discrimination index sequence, and regard the remaining spectrum feature components as redundant features for elimination.

[0053] The retained spectrum feature components form a new frequency feature sub-matrix, which is the sparse feature matrix and can effectively improve the modeling efficiency and discrimination ability in the modeling of the multi-modal mapping model.

[0054] Specifically, establish a multi-modal mapping model of the sparse feature matrix and the target device state based on the historical data of the target device fault diagnosis. Input the sparse feature matrix to obtain the abnormal risk score value, including: Construct a data set containing the historical operation state data of the target device and the corresponding state labels, standardize the historical operation state data, and form a historical sparse feature sample set; Specifically, collect the fault diagnosis historical data of the target device under multiple typical operating cycles, including various samples in the normal state, warning state, and abnormal state. Perform time series subset partitioning, multi-scale transient feature extraction, asynchronous interference identification and correction, spectrum conversion, and sparse feature extraction operations on the fault diagnosis historical data in sequence to generate a historical sparse feature matrix.

[0055] Perform standardization processing on each column of frequency components in the historical sparse feature matrix. The standardization method is: subtract the mean of the response amplitudes of all samples in each column from the response amplitude of each column, and then divide by the standard deviation of the response amplitude of the corresponding column. The data after standardization is the historical sparse feature sample set.

[0056] Perform sample pairing processing on the historical sparse feature sample set and the corresponding state labels to establish a set of correspondence data pairs between the sparse features and the state labels; Specifically, for each sample entry in the historical sparse feature sample set, find the corresponding real operating state record in the historical device monitoring records. The real operating state is obtained by expert annotation or automatic recognition in the operation logs, and specifically includes but is not limited to state labels such as normal, warning, and abnormal.

[0057] Pair each sparse feature vector with the corresponding state label to construct a set of correspondence data pairs between the samples and the labels. Each data pair contains two fields: the standardized sparse feature vector and the state classification label of the target device at the corresponding moment. The set of correspondence data pairs is used as the training data set of the multi-modal mapping model, which contains state distribution and sparse response feature information.

[0058] Construct a multi-modal mapping model based on the historical sparse feature sample set and the set of correspondence data pairs; the input of the multi-modal mapping model is the sparse feature matrix, and the output of the multi-modal mapping model is the abnormal risk score value; Specifically, the larger the abnormal risk score value, the greater the possibility that the target device has an abnormal risk.

[0059] Input the sparse feature matrix into the multi-modal mapping model and perform mapping inference operations to obtain the abnormal risk score value corresponding to the real-time operating state of the target device; Specifically, perform mapping inference operations, that is, input the sparse feature matrix into the set of mapping functions during the inference stage of the multi-modal mapping model, and output the abnormal risk score value corresponding to the real-time operating state of the target device according to the structural parameters and response rules in the multi-modal mapping model.

[0060] Specifically, based on the comparison between the abnormal risk score value and the preset abnormal threshold, determine the state label of the target device, and output the state abnormal warning result of the target device based on the state label of the target device, including: Set a first abnormal threshold and a second abnormal threshold; the first abnormal threshold is less than the second abnormal threshold; Specifically, the setting of the first abnormal threshold and the second abnormal threshold is based on the statistical distribution characteristics of the historical abnormal risk score values corresponding to the target device in the historical operation data. Specifically: count all the historical abnormal risk score values, and classify and group the historical abnormal risk score values according to the status labels; select the upper quartile of the historical abnormal risk score values corresponding to the historical normal state as the first abnormal risk score threshold, and select the lower quartile of the historical abnormal risk score values corresponding to the historical abnormal state as the second abnormal risk score threshold.

[0061] When the abnormal risk score value is less than or equal to the first abnormal threshold, it is determined that the status label of the target device is in the normal state; When the score value is greater than the first abnormal threshold and less than the second abnormal threshold, it is determined that the status label of the target device is in the warning state; When the score value is greater than or equal to the second abnormal threshold, it is determined that the status label of the target device is in the abnormal state; Trigger corresponding warning signals according to the status label of the target device. No warning is triggered in the normal state, a primary alarm prompt is triggered in the warning state, and an emergency shutdown instruction is triggered and a fault report is generated in the abnormal state.

[0062] The above formulas are all dimensionless and take their numerical values for calculation. The formulas are obtained by collecting a large amount of data for software simulation to get a formula closest to the actual situation. The preset parameters and threshold selection in the formulas are set by those skilled in the art according to the actual situation.

[0063] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (such as infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or a data center that contains one or more collections of available media. The available medium can be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0064] Those of ordinary skill in the art will realize that the modules and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or in a combination of computer software and electronic hardware. Whether these functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. Skilled artisans can use different methods for each specific application to implement the described functions, but such implementation should not be considered to exceed the scope of the present application.

[0065] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the systems, devices, and modules described above can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein.

[0066] In several embodiments provided in the present application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the modules is only a logical function division, and there can be other division methods in actual implementation. For example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the couplings, direct couplings, or communication connections shown or discussed with each other can be through some interfaces, and the indirect couplings or communication connections of the devices or modules can be in electrical, mechanical, or other forms.

[0067] The module described as a separation component may or may not be physically separated. The component shown as a module may or may not be a physical module. It may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0068] In addition, in each embodiment of this application, the functional modules can be integrated into one processing module, or each module can exist physically alone, or two or more modules can be integrated into one module.

[0069] If the described function is implemented in the form of a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in each embodiment of this application. The aforementioned storage medium includes: USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs and other various media that can store program codes.

[0070] The above is only the specific implementation manner of this application, but the protection scope of this application is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed in this application, and all should be covered by the protection scope of this application. Therefore, the protection scope of this application should be subject to the protection scope of the claimed rights.

[0071] Finally: The above is only the preferred embodiment of the present invention and is not used to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A device status monitoring and anomaly warning method based on multi-source data fusion, characterized in that It includes the following steps: Collect the original monitoring data of the target device under unsteady operating conditions, and perform segmentation processing according to the sampling frequency differences to obtain multiple time series subsets; Extract the transient features of each time series subset according to the operating state characteristics of the target device under unsteady operating conditions, and construct a multi-scale transient feature set; Establish an asynchronous interference recognition index, perform time-domain misalignment evaluation and correction on each transient feature, and generate a corrected synchronous transient feature set; Perform holographic spectrum transformation on the synchronous transient feature set to obtain an initial holographic spectrum feature matrix, and use a feature sparse discrimination mechanism to evaluate the sparsity of the initial holographic spectrum feature matrix and remove redundant features to form a sparse feature matrix; Establish a multi-modal mapping model between the sparse feature matrix and the state of the target device based on the historical data of the target device's fault diagnosis. Input the sparse feature matrix to obtain an abnormal risk score value; Determine the state label of the target device according to the comparison between the abnormal risk score value and the preset abnormal threshold, and issue a state anomaly warning result for the target device based on the state label of the target device.

2. The device status monitoring and anomaly warning method based on multi-source data fusion according to claim 1, wherein, Collect the original monitoring data of the target device under unsteady operating conditions, and perform segmentation processing according to the sampling frequency differences to obtain multiple time series subsets. Specifically: Obtain the original monitoring data of multiple different types of sensors of the target device under unsteady operating conditions, and determine the sampling interval of the time series according to the sampling frequency of each sensor; Perform time-series segmentation on the original monitoring data according to the sampling intervals respectively to form independent time series subsets corresponding to the sampling frequencies of each type of sensor; Mark each independent time series subset according to a unified time reference to obtain multiple time series subsets.

3. The device status monitoring and anomaly warning method based on multi-source data fusion according to claim 2, characterized in that, Extract the transient features of each time series subset according to the operating state characteristics of the target device under unsteady operating conditions, and construct a multi-scale transient feature set. Specifically: Perform sliding slicing processing on each time series subset according to a fixed window length to generate multiple time window segments; In each time window segment, extract time-domain change features including signal fluctuation mutation, edge jump, short-term wave peak, and periodic perturbation to construct a single-scale transient feature vector; Perform unified coding on multiple single-scale transient feature vectors generated under different window lengths, and combine them according to the time window scale hierarchy to generate a multi-scale transient feature set.

4. A device status monitoring and anomaly warning method based on multi-source data fusion according to claim 3, characterized in that, Establish an asynchronous interference recognition index, perform time-domain misalignment evaluation and correction on each transient feature, and generate a corrected synchronous transient feature set. Specifically: Calculate the time-domain offset between each single-scale transient feature vector and a preset synchronous reference, and determine the degree of asynchronous interference according to the time-domain offset; Establish an asynchronous interference recognition index according to the degree of asynchronous interference, and use the asynchronous interference recognition index to perform time-domain misalignment correction on each single-scale transient feature vector; Perform time-domain reconstruction on all single-scale transient feature vectors that have completed time-domain misalignment correction to generate a corrected synchronous transient feature set.

5. The device status monitoring and anomaly warning method based on multi-source data fusion according to claim 4, characterized in that Perform holographic spectrum transformation on the synchronous transient feature set to obtain an initial holographic spectrum feature matrix, and use a feature sparse discrimination mechanism to evaluate the sparsity of the initial holographic spectrum feature matrix and remove redundant features to form a sparse feature matrix. Specifically: Perform holographic spectrum transformation on the corrected single-scale transient feature vectors in the synchronous transient feature set to obtain the frequency spectrograms corresponding to each corrected single-scale transient feature vector respectively; Unfold each frequency spectrogram along the unified frequency axis coordinates and form an initial high-dimensional frequency feature space to generate an initial holographic spectrum feature matrix; Calculate the sparsity of each spectral feature component in the initial holographic spectrum feature matrix and determine the contribution value of each spectral feature component in the initial holographic spectrum feature matrix; Set a sparsity threshold, use the sparsity threshold to screen and remove redundant features in the initial holographic spectrum feature matrix, and extract the spectral feature components with contribution values higher than the sparsity threshold to form a sparse feature matrix.

6. The device status monitoring and anomaly warning method based on multi-source data fusion according to claim 5, characterized in that, Establish a multi-modal mapping model between the sparse feature matrix and the target device state based on the target device fault diagnosis historical data. Input the sparse feature matrix to obtain an abnormal risk score value, specifically: Construct a data set containing the target device historical operating state data and the corresponding state labels, standardize the historical operating state data to form a historical sparse feature sample set; Perform sample pairing processing on the historical sparse feature sample set and the corresponding state labels to establish a set of correspondence data pairs between the sparse features and the state labels; Construct a multi-modal mapping model based on the historical sparse feature sample set and the set of correspondence data pairs; the input of the multi-modal mapping model is the sparse feature matrix, and the output of the multi-modal mapping model is the abnormal risk score value; Input the sparse feature matrix into the multi-modal mapping model, perform mapping inference operations, and obtain the abnormal risk score value corresponding to the real-time operating state of the target device.

7. A device status monitoring and anomaly warning method based on multi-source data fusion according to claim 6, characterized in that, According to the comparison between the abnormal risk score value and the preset abnormal threshold, determine the state label of the target device, and output the state abnormal warning result of the target device based on the state label of the target device, specifically: Preset a first abnormal threshold and a second abnormal threshold; the first abnormal threshold is less than the second abnormal threshold; When the abnormal risk score value is less than or equal to the first abnormal threshold, determine that the state label of the target device is in a normal state; When the score value is greater than the first abnormal threshold and less than the second abnormal threshold, determine that the state label of the target device is in a warning state; When the score value is greater than or equal to the second abnormal threshold, determine that the state label of the target device is in an abnormal state; Trigger corresponding warning signals according to the state label of the target device. No warning is triggered for the normal state, a primary alarm prompt is triggered for the warning state, and an emergency stop instruction is triggered for the abnormal state and a fault report is generated.

Images