Method for predictive maintenance of equipment based on industrial internet and related device

Abstract

The application relates to the technical field of industrial internet, and provides a device fault prediction maintenance method based on industrial internet and related equipment. By collecting multi-source data and aligning to generate a time sequence matrix according to a running cycle, the matrix is upgraded to a tensor and nonlinear principal components are extracted, a degradation tensor sequence is generated by similarity modulation, a deviation field is constructed and a system potential energy function is calculated, a fault symptom index is generated by using the change rate of the potential energy with time, the device can be accurately identified before entering the rapid deterioration stage, the abnormal tensor position is traced back to the original feature space and mapped to a specific fault type after triggering the early warning, a mathematical anomaly is converted into an operable maintenance suggestion, early, interpretable and adaptive predictive maintenance is realized without a fault label, and the efficiency and reliability of industrial production are improved.

Description

Technical Field

[0001] This application relates to the field of industrial internet technology, and more specifically, to a method and related equipment for predictive maintenance of equipment faults based on industrial internet. Background Technology

[0002] In the field of predictive maintenance for industrial equipment, existing methods typically rely on statistical analysis of signals from single types of sensors or on training classification models based on historical fault data to identify equipment anomalies. These methods are limited to analyzing local variations in individual signals, and model training heavily depends on a large amount of labeled fault sample data. Furthermore, after issuing warnings, it is difficult to explain the specific source of the signal causing the anomaly, often requiring secondary confirmation from human experience for maintenance decisions.

[0003] However, early signs of equipment degradation often exist in weak and scattered forms across multiple sensor channels simultaneously. Each channel appears within normal range when observed individually, making it difficult for existing methods to effectively capture this cross-channel coordinated change information. Furthermore, health benchmarks drift slowly after long-term equipment operation, leading to false alarms or missed alarms, which in turn results in lower efficiency and reliability in industrial production. Summary of the Invention

[0004] This application provides a method and related equipment for predictive maintenance of equipment failure based on the Industrial Internet, thereby solving the problem of low efficiency and reliability in industrial production.

[0005] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part from practice of this application.

[0006] According to one aspect of this application, a method for predictive maintenance of equipment faults based on the Industrial Internet is provided, comprising: collecting multi-source monitoring data during equipment operation; slicing and aligning the multi-source monitoring data according to the equipment's operating cycle to generate a time-series data matrix; converting the time-series data matrix into a three-dimensional data array; extracting the nonlinear principal components of the three-dimensional data array and calculating the similarity between the current three-dimensional data array and a reference tensor; modulating the three-dimensional data array according to the similarity to generate a degenerate tensor sequence; constructing an deviation field based on the degenerate tensor sequence and the reference tensor; calculating a system potential function based on the deviation field; generating a fault symptom index based on the time rate of change of the system potential function; issuing an early warning when the fault symptom index exceeds a threshold; responding to the early warning; tracing back to the time-series data matrix based on the position of the tensor element contributing the most in the deviation field to determine abnormal features; mapping the abnormal features to a fault prediction result; and outputting maintenance suggestions based on the time rate of change of the system potential function; obtaining feedback information after actual maintenance; and updating adjustable parameters in the system potential function using the feedback information.

[0007] In this application, based on the aforementioned scheme, the multi-source monitoring data includes at least sensor signals, operating parameters, and operation records; slicing and aligning the multi-source monitoring data according to the equipment's operating cycle to generate a time-series data matrix includes: determining the start and end trigger signals of the slicing window according to the equipment's operating cycle; slicing the sensor signals of each type in the multi-source monitoring data according to the slicing window to generate slice signals; extracting time-domain statistical features from the slice signals; and arranging the extracted time-domain statistical features, the operating parameters, and the operation record events according to a unified time axis to form a time-series data matrix.

[0008] In this application, based on the aforementioned scheme, the step of converting the time-series data matrix into a three-dimensional data array, extracting the nonlinear principal components of the three-dimensional data array, calculating the similarity between the current three-dimensional data array and the reference tensor, and modulating the three-dimensional data array according to the similarity to generate a degenerate tensor sequence includes: taking a preset number of historical lag time points forward from the current time as the center, using the feature vector at each historical lag time point as a time slice of the three-dimensional data array, stacking all time slices along the time dimension to construct an initial three-dimensional data array, wherein the first dimension of the initial three-dimensional data array is the time lag order, the second dimension is the original feature variables, and the third dimension is the nonlinear principal components to be extracted; mapping the original feature variables in the second dimension of the initial three-dimensional data array to a high-dimensional space using a kernel function, and in the high-dimensional... The first few nonlinear principal component directions are extracted from the space, and the original feature variable dimension is compressed into the nonlinear principal component dimension to obtain a compressed three-dimensional data array. The original feature variable part of the initial three-dimensional data array that has not been extracted by nonlinear principal components is merged with the compressed three-dimensional data array along the third dimension to form a three-dimensional data array to be matched. The average three-dimensional data array after the same processing of the historical data of the device in a healthy state is used as the three-dimensional data array. The similarity between the three-dimensional data array to be matched at the current moment and the reference tensor is measured by dynamic time warping distance. The similarity is used as a weighting factor to weight each element in the three-dimensional data array to be matched to generate the degenerate three-dimensional data array at the current moment. The degenerate three-dimensional data arrays at each moment are arranged in chronological order to obtain the degenerate tensor sequence.

[0009] In this application, based on the aforementioned scheme, the step of constructing the deviation field according to the degenerate tensor sequence and the reference tensor includes: subtracting the reference tensor from the degenerate tensor at the current moment in the degenerate tensor sequence to obtain the deviation field.

[0010] In this application, based on the aforementioned scheme, the calculation of the system potential function based on the off-field includes: determining an energy potential well term based on the square value of each element in the off-field, determining a nonlinear instability term based on the fourth power value of each element in the off-field, determining a curvature driving term based on the second-order mixed partial derivative of the off-field in the spatiotemporal dimension, and determining a long-range interaction term based on the interaction between elements at different positions in the off-field; summing the energy potential well term, the nonlinear instability term, the curvature driving term, and the long-range interaction term to obtain the system potential function value at the current moment.

[0011] In this application, based on the aforementioned scheme, the step of generating a fault symptom index according to the time rate of change of the system potential energy function, and issuing an early warning when the fault symptom index exceeds a threshold, includes: performing numerical differentiation on the system potential energy function to obtain the first and second derivatives of the system potential energy function with respect to time; combining the first and second derivatives and mapping them to a preset interval to obtain the fault symptom index; using the statistical quantile of the fault symptom index in the historical normal window as the threshold; issuing an early warning when the current fault symptom index exceeds the threshold, and recording the position of the tensor element that contributes the most in the deviation field.

[0012] In this application, based on the aforementioned scheme, the response to the warning, tracing back to the time-series data matrix based on the position of the tensor element contributing the most in the deviation field, determining the abnormal features, mapping the abnormal features to the fault prediction result, and outputting maintenance suggestions based on the time rate of change of the system potential function, includes: calculating the partial derivative values ​​of the tensor element positions with respect to each original feature, selecting abnormal features according to the partial derivative values; matching the abnormal features with preset fault mapping rules, outputting the fault prediction type based on the matching result; selecting the maintenance level based on the fault prediction type and the value of the first derivative of the system potential function, and outputting a maintenance work order containing the fault prediction type, maintenance level, and maintenance suggestions.

[0013] According to one aspect of this application, an apparatus for predictive maintenance of equipment faults based on the Industrial Internet is provided, comprising:

[0014] The acquisition module is used to collect multi-source monitoring data during the operation of the equipment, and to slice and align the multi-source monitoring data according to the operating cycle of the equipment to generate a time-series data matrix.

[0015] The conversion module is used to convert the time-series data matrix into a three-dimensional data array, extract the nonlinear principal components of the three-dimensional data array, calculate the similarity between the current three-dimensional data array and the reference tensor, and modulate the three-dimensional data array according to the similarity to generate a degenerate tensor sequence.

[0016] The early warning module is used to construct an deviation field based on the degenerate tensor sequence and the reference tensor, calculate the system potential energy function based on the deviation field, generate a fault symptom index based on the time rate of change of the system potential energy function, and issue an early warning when the fault symptom index exceeds a threshold.

[0017] The response module is used to respond to the warning, trace back to the time series data matrix based on the position of the tensor element that contributes the most in the deviation field, determine the abnormal features, map the abnormal features to the fault prediction result, and output maintenance suggestions based on the time rate of change of the system potential energy function.

[0018] The update module is used to obtain feedback information after actual maintenance and to update the adjustable parameters in the system potential energy function using the feedback information.

[0019] According to one aspect of this application, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the method for predictive maintenance of equipment faults based on the Industrial Internet as described in the above embodiments.

[0020] According to one aspect of this application, a computer device is provided, comprising: one or more processors; and a storage device for storing one or more programs, which, when executed by the one or more processors, cause the one or more processors to implement the industrial internet-based equipment fault prediction and maintenance method as described in the above embodiments.

[0021] According to one aspect of this application, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the industrial internet-based equipment fault prediction and maintenance method provided in the various optional implementations described above.

[0022] The main differences and technical effects of the technical solution in this application compared with the prior art are as follows:

[0023] This technical solution eliminates the time misalignment problem of heterogeneous data from different frequencies in industrial sites by collecting multi-source data and aligning it according to the operating cycle to generate a time-series matrix. The matrix is ​​then upgraded to a tensor, and nonlinear principal components are extracted. A degradation tensor sequence is generated using similarity modulation, enabling continuous quantitative characterization of the multi-dimensional nonlinear degradation process of equipment. Based on this, an off-field is constructed, and the system potential energy function is calculated. The rate of change of potential energy over time is used to generate fault symptom indicators, allowing accurate identification of equipment before it enters the rapid degradation stage. Dynamic thresholds effectively reduce false alarms. After triggering an early warning, the abnormal tensor position is traced back to the original feature space and mapped to a specific fault type, transforming mathematical anomalies into actionable maintenance suggestions, thus solving the problem of the uninterpretable nature of black-box models. Finally, maintenance feedback is used to update the potential energy function parameters, forming a self-evolving closed loop. This ensures that the predictive performance remains accurate as the equipment is used, thereby achieving early, interpretable, and adaptive predictive maintenance without the need for fault labels, improving the efficiency and reliability of industrial production.

[0024] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0025] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0026] Figure 1 A flowchart of a method for predictive maintenance of equipment failure based on the Industrial Internet is shown in one embodiment of this application.

[0027] Figure 2 A flowchart illustrating the generation of a time-series data matrix is ​​shown in one embodiment of this application.

[0028] Figure 3 A schematic diagram of an apparatus for predictive maintenance of equipment failure based on the Industrial Internet is shown in one embodiment of this application.

[0029] Figure 4 A schematic diagram of the structure of a computer system suitable for implementing the computer device of the present application is shown. Detailed Implementation

[0030] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided to make this application more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art.

[0031] The block diagrams shown in the attached figures are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, or in one or more hardware modules composed of smart chips, smart integrated circuits, or application-specific integrated circuits (ASICs), or in different network and / or processor devices and / or microcontroller devices.

[0032] The implementation details of the technical solution of this application are described below:

[0033] Figure 1 A flowchart illustrating a method for predictive maintenance of equipment failure based on the Industrial Internet of Things, according to an embodiment of this application, is shown. (Refer to...) Figure 1 As shown, the method for predictive maintenance of equipment faults based on the Industrial Internet includes at least steps S110 to S150, which are described in detail below:

[0034] S110: Collect multi-source monitoring data during equipment operation, slice and align the multi-source monitoring data according to the equipment's operating cycle, and generate a time-series data matrix.

[0035] In this embodiment, various sensors deployed on industrial equipment collect raw signals during equipment operation in real time, while simultaneously acquiring corresponding operating parameters and operation records from the industrial control system and manufacturing execution system. The multi-source monitoring data includes at least sensor signals, operating parameters, and operation records. Sensor signals include, but are not limited to, various types of time-domain signals such as vibration, temperature, pressure, current, and voltage; operating parameters include, but are not limited to, load values, speed values, and valve opening values; and operation records include event types and their corresponding timestamps from the equipment operation log.

[0036] like Figure 2 As shown, in one embodiment of this application, the multi-source monitoring data is sliced ​​and aligned according to the device's operating cycle to generate a time-series data matrix, including:

[0037] S210, determine the start and end trigger signals of the slicing window according to the equipment's operating cycle, slice the sensor signals of each type in the multi-source monitoring data according to the slicing window, and generate slice signals;

[0038] S220, Extract time-domain statistical features from the sliced ​​signal;

[0039] S230, the extracted time-domain statistical features, the operating condition parameters, and the running record events are arranged according to a unified time axis to form a time-series data matrix.

[0040] Specifically, due to the inconsistent sampling frequencies of various sensors (e.g., vibration signals can have a sampling rate of 10kHz while temperature signals are only 1Hz), the raw data needs to be sliced ​​and aligned. Specifically, the start and end trigger signals of the slicing window are determined based on the equipment's operating cycle. The operating cycle can be a single processing flow, a single start-stop cycle, or a single production batch. Using equipment operating status events as trigger signals, such as a spindle start signal as the window start trigger and a spindle stop signal as the window end trigger, the signals from various types of sensors are sliced ​​according to the slicing window to generate sliced ​​signals, so that the signal segment within each window corresponds to a complete physical process.

[0041] Temporal statistical features are extracted from the same type of sensor signals within each slice window. These features include mean, variance, peak value, and kurtosis. The statistical features of all sensor types within each window are combined to form a sensor feature vector corresponding to that window. Load values, speed values, and valve opening values ​​from the operating parameters are encoded into multi-dimensional operating state vectors. Event types and corresponding timestamps from the operation records are converted into discrete event labels. The extracted temporal statistical features, the operating parameters, and the operation record events are arranged along a unified time axis to form a time-series data matrix. The rows of this time-series data matrix correspond to each time window, and the columns correspond to the sum of the sensor feature dimension, operating parameter dimension, and event label dimension.

[0042] In practical applications, industrial sensors often have varying sampling frequencies, with high-frequency vibrations and low-frequency temperatures coexisting. Directly splicing these data can lead to timeline misalignment and feature distortion. The aforementioned process slices the data using the complete operating cycle of the equipment as a window, ensuring that each data segment corresponds to a complete physical process. Statistical features are extracted to compress high-frequency signals into equally spaced time points, unifying all data streams to the same time reference and eliminating asynchronous errors caused by sampling frequency differences. This step ensures the spatiotemporal consistency of subsequent modeling input data, avoiding feature misalignment and spurious correlations caused by data misalignment.

[0043] S120, the time-series data matrix is ​​converted into a three-dimensional data array, the nonlinear principal components of the three-dimensional data array are extracted, and the similarity between the current three-dimensional data array and the reference tensor is calculated. The three-dimensional data array is modulated according to the similarity to generate a degenerate tensor sequence.

[0044] In this embodiment, to characterize the device state in both time and space dimensions, the generated time-series data matrix is ​​converted into a three-dimensional tensor structure. Specifically, historical time points are extracted backward from the current moment, and the feature vectors at each time point are used as tensor slices to construct an initial tensor. Then, the similarity between the current initial tensor and the reference tensor is calculated, and the tensor is modulated based on the similarity to generate a degenerate tensor sequence.

[0045] In one embodiment of this application, the time-series data matrix is ​​converted into a three-dimensional data array, the nonlinear principal components of the three-dimensional data array are extracted, and the similarity between the current three-dimensional data array and the reference tensor is calculated. The three-dimensional data array is then modulated based on the similarity to generate a degenerate tensor sequence, including:

[0046] Taking the current moment as the center, a preset number of historical lag time points are extracted. The feature vector at each historical lag time point is used as a time slice of the three-dimensional data array. All time slices are stacked along the time dimension to construct an initial three-dimensional data array. The first dimension of the initial three-dimensional data array is the time lag order, the second dimension is the original feature variable, and the third dimension is the nonlinear principal component to be extracted.

[0047] The original feature variables in the second dimension of the initial three-dimensional data array are mapped to a high-dimensional space by a kernel function. The first few nonlinear principal component directions are extracted in the high-dimensional space, and the original feature variable dimension is compressed into the nonlinear principal component dimension to obtain the compressed three-dimensional data array.

[0048] The original feature variables that have not been extracted by nonlinear principal components in the initial three-dimensional data array are merged with the compressed three-dimensional data array along the third dimension to form a three-dimensional data array to be matched.

[0049] The average three-dimensional data array of the historical data of the device in a healthy state after the same processing is used as the three-dimensional data array. The dynamic time warping distance is used to measure the similarity between the three-dimensional data array to be matched at the current moment and the reference tensor.

[0050] Using the similarity as a weighting factor, each element in the three-dimensional data array to be matched is weighted to generate the degenerate three-dimensional data array at the current time. The degenerate three-dimensional data arrays at each time are arranged in chronological order to obtain the degenerate tensor sequence.

[0051] In this embodiment, the generated two-dimensional time-series data matrix is ​​converted into a three-dimensional data array with spatiotemporal correlation characteristics. Since a two-dimensional matrix can only express the planar relationship of a feature at a given moment, it cannot simultaneously characterize the mutual coupling between features and the temporal continuity of the degradation process. Therefore, this step extracts a preset number of P historical lag time points centered on the current moment. The complete feature vector at each time point (including sensor statistical features and operating condition codes) is used as a time slice of the three-dimensional data array. These slices are stacked sequentially along the time dimension to construct the initial three-dimensional data array. The three dimensions of this initial three-dimensional data array correspond to the time lag order (dimension P), the original feature variables (dimension Q, i.e., the number of columns in the time-series data matrix), and the nonlinear principal components to be extracted (dimension R is initially empty), providing a data carrier for subsequent mining of degradation patterns from the spatiotemporal joint domain.

[0052] After constructing the initial 3D data array, nonlinear dimensionality reduction is performed to extract hidden high-order coupling information in the original feature space. Since there are often complex nonlinear correlations between different sensor signals—for example, the rise of high-frequency vibration components may be accompanied by a slow temperature increase—traditional linear dimensionality reduction methods struggle to effectively capture such cooperative change patterns.

[0053] Therefore, this embodiment utilizes kernel principal component analysis (KPCA) to implicitly map the original feature variables in the second dimension of the initial 3D data array to a high-dimensional regenerative kernel Hilbert space using radial basis function (RBF) kernels. In this space, the covariance matrix is ​​calculated, and the first few nonlinear principal component directions with the largest contribution rates are extracted. This compresses the dimension of the original feature variables from Q to R (R is much smaller than Q), resulting in a compressed 3D data array. To balance the physical interpretability of the original features with the high-dimensional sensitivity of the nonlinear principal components in subsequent degradation metrics, the original feature variables from the initial 3D data array that were not extracted by the nonlinear principal components are merged with the compressed 3D data array along the third dimension to form a matching 3D data array. The third dimension of this array contains both the original eigenvalues ​​and the projection coefficients of the nonlinear principal components.

[0054] After obtaining the three-dimensional data array to be matched, it is further transformed into a degraded three-dimensional data array sequence that can quantitatively characterize the degree of device degradation. Specifically, the three-dimensional data array obtained by averaging the historical data of multiple complete operating cycles of the device in a brand-new healthy state after processing through the same process is used as the benchmark three-dimensional data array. This benchmark array represents the standard pattern of the device in a non-degraded state.

[0055] Dynamic Time Warping (DTW) distance is used to calculate the similarity between the current 3D data array to be matched and the reference 3D data array. DTW eliminates the normal temporal phase shift caused by fluctuations in equipment operating conditions (such as changes in speed or load) by allowing nonlinear bending of the time axis, ensuring that the calculated similarity only reflects morphological differences caused by inherent equipment degradation. The DTW distance is used as a weighting factor to weight each element in the 3D data array to be matched; a larger distance indicates a greater deviation from the healthy baseline, and the corresponding element has a higher weight. After weighting, a degraded 3D data array is generated for the current moment. Arranging the degraded 3D data arrays at each moment in chronological order yields a degraded 3D data array sequence, which serves as input for subsequent steps, used for calculating the system potential function and identifying fault symptoms.

[0056] The above process elevates the two-dimensional matrix into a three-dimensional tensor, preserving the data structure simultaneously across the dimensions of time lag, feature variables, and nonlinear principal components. Traditional linear dimensionality reduction cannot capture the nonlinear collaborative drift between features, while kernel principal component analysis (KPI) maps the original features to a high-dimensional space using kernel functions, extracting nonlinear principal components and revealing hidden patterns when multiple sensor signals undergo subtle joint changes. Furthermore, dynamic time warping is employed to calculate the similarity between the current initial tensor (the current three-dimensional data array) and the baseline tensor (the baseline three-dimensional data array) in a healthy state. This eliminates normal temporal phase differences caused by operational fluctuations, retaining only morphological differences related to inherent equipment degradation. Using similarity as a weighted modulation tensor element achieves the transformation from raw data to a degradation measure, providing a continuous tensor sequence quantifying the degradation magnitude for subsequent analysis.

[0057] S130, construct an offset field based on the degenerate tensor sequence and the reference tensor, calculate the system potential energy function based on the offset field, generate a fault symptom index based on the time rate of change of the system potential energy function, and issue an early warning when the fault symptom index exceeds a threshold.

[0058] In this embodiment, the current degenerate tensor is first subtracted from the reference tensor to obtain a three-dimensional deviation field, which quantifies the offset distribution of the device relative to the healthy baseline in each dimension of time lag, characteristic variables, and nonlinear principal components. Then, a system potential energy function is constructed based on this deviation field. This function integrates an energy well term describing small fluctuations, a nonlinear instability term capturing accelerated degradation, a curvature driving term reflecting abrupt changes in multi-signal correlations, and a long-range interaction term characterizing fault propagation. The sum of these four terms yields the system potential energy value at the current moment. Finally, the potential energy function is numerically differentiated to obtain its first and second derivatives. The two derivatives are combined and mapped to a fault symptom index in the [0,1] interval using a sigmoid function. When this index exceeds the statistical quantile threshold of the historical normal window, an early warning is triggered, and the position of the deviation field element that contributes the most is recorded for subsequent source tracing.

[0059] In one embodiment of this application, constructing an offset field based on the degenerate tensor sequence and the reference tensor includes: subtracting the reference tensor from the degenerate tensor at the current time in the degenerate tensor sequence to obtain the offset field. The generated degenerate tensor at the current time t... With the reference tensor representing the device health benchmark By performing element-wise subtraction, the three-dimensional deviation field is obtained. Its function is to remove the constant components in the original degraded tensor caused by changes in the inherent structure of the equipment or normal operating conditions, and retain only the dynamic deviation components that reflect the performance degradation of the equipment.

[0060] Each element of the off-field The magnitude and direction of the device's current state offset relative to its own health baseline were precisely quantified at the tensor space location determined by the time lag index i, the characteristic variable index j, and the nonlinear principal component index k. Through this subtraction operation, the constant components in the original degraded tensor caused by changes in the device's inherent structure or normal operating conditions were effectively eliminated, retaining only the dynamic deviation components reflecting the device's performance degradation. This provides a clean data input containing only degradation information for the subsequent calculation of the system's potential function.

[0061] In one embodiment of this application, the system potential energy function is calculated based on the off-field, including:

[0062] The energy potential well term is determined based on the square value of each element in the off-field, the nonlinear instability term is determined based on the fourth power value of each element in the off-field, the curvature driving term is determined based on the second-order mixed partial derivative of the off-field in the spacetime dimension, and the long-range interaction term is determined based on the interaction between elements at different positions in the off-field.

[0063] The system potential energy function value at the current moment is obtained by summing the energy potential well term, the nonlinear instability term, the curvature driving term, and the long-range interaction term.

[0064] In this embodiment, the first term A1 of the system potential function is the energy well term, which is determined based on the squared values ​​of each element in the deviation field, and its specific form is as follows:

[0065] ,

[0066] in, The first characteristic parameter of the device is obtained by fitting and calibrating the device nameplate parameters with historical health data. It is used to control the stiffness of the potential well, that is, to determine the strength of the elastic restoring force generated by the device in response to small random deviations.

[0067] Where i, j, and k represent the indices of the three dimensions of the off-field tensor (time lag, feature variables, and principal components), respectively. Specifically, i represents the index of the first dimension of the tensor, ranging from 1 to P, where P is the time lag order set in step two, used to identify the specific position of any element in the off-field in the time lag dimension. j represents the index of the second dimension of the tensor, ranging from 1 to Q, where Q is the total number of feature variables in the time series data matrix in step one, used to identify the specific position of any element in the off-field in the original feature variable dimension. k represents the index of the third dimension of the tensor, ranging from 1 to R, where R is the number of nonlinear principal components extracted by kernel principal component analysis, used to identify the specific position of any element in the off-field in the nonlinear principal component dimension.

[0068] in, Represents the off-field tensor The specific value at index position (i,j,k) represents the deviation of the current time relative to the healthy baseline under a specific combination of time lag order, characteristic variables, and nonlinear principal components.

[0069] In this embodiment, the second term A2 is a nonlinear instability term, which is determined based on the fourth power values ​​of each element in the deviation field, and its specific form is as follows:

[0070] ,

[0071] in, This represents the second characteristic parameter of the device, used to control the strength of the nonlinear instability term, that is, to determine the rate at which the potential energy accelerates as the deviation increases. This indicates a deviation from a specific element in the field. The fourth power of the value has the function of making the contribution of this term almost zero when the deviation is small, and when the deviation exceeds the critical value, it makes this term rapidly dominate the total potential energy at a fourth power speed, thereby accurately simulating the process of the potential barrier drop after the equipment enters the accelerated degradation stage.

[0072] The second term, A2, has high-order nonlinear characteristics. Its design motivation is to simulate the physical behavior of the equipment after it enters the degradation acceleration stage. When the deviation is small, this term is much smaller than the energy potential well term and has almost no impact on the total potential energy. Once the deviation exceeds a certain critical value, this term increases rapidly at a fourth power rate, causing the system potential barrier to drop sharply, reflecting the phase transition process of the equipment from slow wear to rapid deterioration. The two terms together constitute a dual potential well structure that reflects local stability.

[0073] In this embodiment, the third term A3 is a curvature-driven term, which is determined based on the second-order mixed partial derivative of the deviation field in the spacetime dimension, and its specific form is as follows:

[0074] ,

[0075] in, This represents the third device characteristic parameter, used to control the weight ratio of the curvature drive term in the total potential energy of the system, and determines the sensitivity of the system to the distortion of the correlation mode between sensor signals. The symbol represents the partial differential operator, used to calculate the partial derivative of a multivariable function with respect to a specific variable. In this formula, it is used to calculate the degree of local curvature of the off-field tensor in the spacetime dimension. The second-order mixed partial derivative of a specific element in the deviation field with respect to the time lag dimension index i and the characteristic variable dimension index j is obtained by taking the limit after calculating the difference along the two directions i and j in the high-dispersion tensor field. Its value reflects the degree of local curvature of the tensor field in the spatiotemporal intersection direction.

[0076] The second-order mixed partial derivative in the third item A3 calculates the local curvature of the tensor field in the direction of the intersection of the time lag index dimension and the feature variable index dimension. When the synchronous correlation mode between different sensor signals remains stable, the curvature value is low; once a sudden change occurs in the multimodal coupling relationship (such as vibration phase shift accompanied by temperature response delay), the curvature value will increase significantly. This item is weighted by the deviation value itself, so as to accurately capture the distortion of the correlation mode between signals.

[0077] In this embodiment, the fourth term A4 is a long-range interaction term, which is determined based on the interaction between elements at different positions in the offset field, and its specific form is as follows:

[0078] ,

[0079] in, The fourth device characteristic parameter is used to control the overall contribution of the long-range interaction term to the total potential energy of the system, and to determine the degree of influence of the fault propagation effect on the potential energy value. p, q, and r represent the second set of tensor indices from the same source as i, j, and k, corresponding to the time lag order dimension, the feature variable dimension, and the nonlinear principal component dimension, respectively. They are used to traverse another position (p, q, r) in the tensor space that is different from (i, j, k), and to calculate the interaction between any two elements at different positions.

[0080] This represents a double summation operator. The outer and inner layers traverse all possible index positions in the tensor space, respectively. Its function is to enumerate all ordered position pairs and accumulate the interaction contribution between any two different elements in the tensor field.

[0081] This indicates the first off-field element involved in the interaction calculation, whose position is determined by the index (i,j,k), providing the first multiplier factor in the long-range interaction formula; This represents the second off-field element involved in the interaction calculation. Its position is determined by the index (p,q,r). It provides the second multiplier factor in the long-range interaction formula and together with the previous element, forms the product term of the pairwise interactions.

[0082] This represents the squared distance between two tensor element positions in the time lag dimension; This represents the squared distance between two tensor elements along the dimension of the feature variable. It represents the squared distance between the positions of two tensor elements along the nonlinear principal component dimension.

[0083] The fourth item, A4, borrows the mathematical form of Coulomb interaction in field theory of physics to calculate the weighted product sum between any two elements at different positions in the tensor space. The denominator is the Euclidean distance between the two positions, which makes the elements that are closer to each other contribute more and the elements that are farther apart contribute less, but the scope of the effect covers the entire tensor space. This effectively captures the spatiotemporal correlation effect of fault symptoms spreading from one component or feature dimension to other components or dimensions.

[0084] Then, the energy well term, nonlinear instability term, curvature driving term, and long-range interaction term are summed at all tensor index positions (i,j,k) to obtain the system potential energy function value Φ(t) at the current moment:

[0085] ,

[0086] This potential energy function value is not simply an accumulation of errors, but rather a comprehensive characterization of the overall potential energy of the device's current degradation state from four complementary dimensions: system stability, nonlinear instability trend, signal correlation distortion, and fault propagation range. When the device is in a healthy state, all elements of the off-field are close to zero, the energy well term and curvature driving term both contribute small positive values, and the contributions of the nonlinear instability term and long-range interaction term are negligible; the system potential energy function value is stable near the baseline. When the device begins to degrade, the coordinated changes of various factors cause the system potential energy function value to increase monotonically, and its rate of increase and acceleration directly reflect the severity of the degradation process. This potential energy function provides a physically clear and continuously differentiable input for the subsequent calculation of the rate of change over time.

[0087] In one embodiment of this application, a fault symptom index is generated based on the time rate of change of the system potential energy function, and an early warning is issued when the fault symptom index exceeds a threshold, including:

[0088] By numerically differentiating the system potential energy function, we obtain the first and second derivatives of the system potential energy function with respect to time.

[0089] The first and second derivatives are combined and mapped to a preset interval to obtain the fault symptom index;

[0090] Using the statistical quantile of the fault symptom index in the historical normal window as a threshold, if the current fault symptom index exceeds the threshold, an early warning is issued, and the position of the tensor element that contributes the most in the deviation field is recorded.

[0091] Specifically, after obtaining the continuous-time system potential function value Φ(t), this step first performs a numerical differentiation operation on it with respect to time. Specifically, the central difference method is used to process Φ(t) on the discrete-time series to obtain the first and second derivatives of the system potential function with respect to time. The first derivative characterizes the rate of change of the potential function with time at the current moment; its continuously increasing positive value indicates that the device is deviating from a healthy steady state at an increasingly faster rate. The second derivative characterizes the changing trend of the first derivative itself, i.e., the acceleration of the potential energy change; its significant positive value indicates that the degradation process has shifted from slow drift to accelerated runaway, an important characteristic of the system approaching the phase transition critical point. The two derivatives reflect the dynamic evolution information of the device degradation process at different time scales.

[0092] The first and second derivatives of the system's potential energy function are nonlinearly combined or weighted and summed to generate the fault symptom index for the current moment. Specifically, the product of the first derivative and the first weight is added to the product of the second derivative and the second weight to obtain the combined input value. This combination method takes into account both the instantaneous rate of potential energy change and the acceleration effect. In the stable degradation phase, the first derivative is positive but the second derivative is close to zero, and the combined value is mainly contributed by the first derivative term. When the equipment enters the rapid degradation phase, the second derivative increases rapidly and is superimposed on the first derivative, forming a multiplier effect, causing the combined input value to jump rapidly. Subsequently, the combined input value is mapped to the preset interval [0,1] through a sigmoid function to obtain the fault symptom index for the current moment. This index intuitively represents the probability that the current equipment is in a pre-fault state in a probabilistic form.

[0093] Regarding threshold setting, a dynamic threshold strategy is adopted instead of a fixed threshold. Specifically, a certain quantile (e.g., the 95th quantile) in the statistical distribution of fault symptom indicators from the most recent preset number of historical normal windows (i.e., periods where no warnings were triggered and subsequent maintenance confirmed no faults) is used as the dynamic judgment threshold for the current moment. This threshold adaptively drifts with the statistical characteristics of the equipment's normal operating state, effectively avoiding the failure of fixed thresholds due to individual equipment differences or slow changes in the operating environment. If the fault symptom indicator at the current moment exceeds this dynamic threshold, a warning signal is triggered, and the position index of the tensor element in the current deviation field that contributes the most to the fault symptom indicator exceeding the threshold is recorded. This contribution is determined by calculating the partial derivative of the fault symptom indicator with respect to each tensor element. The position of the element with the largest absolute value of the partial derivative is the core abnormal source causing this warning, providing a precise mathematical basis for causal tracing and fault location in subsequent steps.

[0094] The above process subtracts the reference tensor from the current degradation tensor to obtain the deviation field, stripping away the constant component brought about by the inherent structure of the equipment and retaining only the dynamic degradation deviation. The system potential function contains four terms: the energy potential well term imposes a squared penalty on small-amplitude random fluctuations, avoiding oversensitivity to normal noise; the nonlinear instability term grows at a fourth-power acceleration, accurately reflecting the sudden drop in the potential barrier after the equipment enters accelerated degradation; the curvature-driven term captures abrupt changes in the correlation patterns between multi-sensor signals, compensating for the limitations of single-channel analysis; and the long-range interaction term quantifies the diffusion effect of the fault across different dimensions. Summing the four terms yields the system potential function, whose first-order time derivative reflects the degradation rate, and its second-order time derivative reflects the degradation acceleration. The combination of these two terms, after nonlinear mapping, generates a fault symptom index that considers both the current degree of deviation and the rapid changes in the degradation trend, exhibiting stronger early sensitivity compared to the fixed threshold method. The dynamic threshold strategy is adaptively adjusted based on the statistical distribution of historical normal windows, further reducing the false alarm rate.

[0095] S140, in response to the warning, trace back to the time series data matrix based on the position of the tensor element that contributes the most in the deviation field, determine the abnormal features, map the abnormal features to the fault prediction result, and output maintenance suggestions based on the time rate of change of the system potential energy function.

[0096] In this embodiment, in response to the warning signal, starting from the position of the tensor element that contributes the most to the deviation field, the process traces back to the original time-series data matrix along the third dimension. The partial derivative values ​​of this position with respect to each original sensor feature and operating condition variable are calculated, and the top-ranked feature combination is selected as the abnormal feature in descending order of absolute value. Then, the abnormal feature combination is input into a preset fault mapping rule base for matching. This rule base contains the correspondence between abnormal features and physical components and fault types. The fault prediction type is output based on the matching results. Finally, based on the maintenance urgency level corresponding to the fault prediction type and the magnitude of the first derivative of the system potential function, the matching level is selected from four maintenance levels: continue observation, adjust process parameters, arrange shutdown for maintenance, and prepare for spare parts replacement. The maintenance suggestion, which includes the fault prediction location, fault prediction type, confidence level, and suggested maintenance time window, is output in the form of a structured chemical order.

[0097] In one embodiment of this application, in response to the warning, the time-series data matrix is ​​traced back to the position of the tensor element contributing the most in the deviation field to determine the abnormal features, the abnormal features are mapped to the fault prediction result, and maintenance suggestions are output based on the time rate of change of the system potential function, including:

[0098] Calculate the partial derivative values ​​of the tensor element positions with respect to each original feature, and select the abnormal features according to the partial derivative values;

[0099] The abnormal features are matched with preset fault mapping rules, and the fault prediction type is output based on the matching results.

[0100] Based on the fault prediction type and the value of the first derivative of the system potential function, the maintenance level is selected, and a maintenance work order containing the fault prediction type, maintenance level, and maintenance recommendations is output.

[0101] Specifically, in this embodiment, in response to the triggered warning signal, the position index of the tensor element that contributes the most to the recorded deviation field is first obtained, denoted as (i, j, k*). This position points to a specific unit in the degenerate tensor space, corresponding to a specific time lag order, a specific feature variable, and a specific nonlinear principal component direction, and is the core source of the entire abnormal deviation. Based on this, the tensor position is traced back to the original time-series data matrix. By calculating the partial derivative of the fault symptom index with respect to the element value at this tensor position with respect to each original feature (including various sensor statistical features and operating parameters) in the time-series data matrix, the contribution of each original feature to the abnormal deviation of the tensor element is obtained. The partial derivative values ​​are sorted in descending order of absolute value, and several of the top-ranked original features are selected as the abnormal feature combination. This process essentially completes the inverse mapping from the mathematical anomaly in the high-dimensional tensor space to the original physical signal.

[0102] After obtaining the abnormal feature combinations, they are input into a pre-established fault mapping rule base for matching. This fault mapping rule base is a pre-built structured knowledge base, the core of which is the probabilistic correspondence between abnormal feature combination patterns and physical component-level fault types. The rule base is constructed based on equipment physical mechanism model analysis and data mining and expert annotation of historical fault cases. The matching process specifically involves: ranking the contribution of each feature in the current abnormal feature combination and calculating the similarity between this ranking and the feature combination patterns stored in the rule base; the fault type associated with the rule entry with the highest similarity is selected as the output result. For example, when the high-frequency kurtosis contribution value of the vibration signal ranks first and the first-order difference contribution value of the temperature feature ranks second in the abnormal feature combination, and there is a temporal co-offset between the two, the matching result output is early lubrication failure of the rolling bearing. Finally, this matching step outputs a structured fault prediction type and its corresponding confidence level.

[0103] After obtaining the fault prediction type, based on the preset maintenance urgency level corresponding to that fault type and the specific value of the first derivative of the current system potential function, a matching maintenance level is selected from multiple preset maintenance levels as a maintenance recommendation. These multiple maintenance levels include at least the "continue observation" level, the "adjust process parameters" level, the "schedule shutdown for maintenance" level, and the "prepare for spare parts replacement" level. The value of the first derivative reflects the rate of the current degradation process; a larger value indicates faster equipment deterioration and a more urgent maintenance level. The fault prediction type determines the specific direction of maintenance and resource preparation requirements. The selected maintenance level, along with the fault prediction type, the key evidence chain in the anomaly feature combination, the fault prediction confidence level, and the recommended maintenance time window calculated based on the first derivative, are integrated and output in a structured format as a maintenance work order for on-site maintenance personnel to refer to and execute. This work order ensures complete traceability and interpretability of the entire decision-making chain from the warning signal to the specific maintenance action.

[0104] The above process, starting from the position of the tensor element with the largest contribution recorded at the time of warning, calculates its partial derivative with respect to each original feature, realizing reverse attribution from abstract anomalies in high-dimensional tensor space to original physical signals. The located anomaly feature combination is matched with preset fault mapping rules, transforming mathematical anomalies into specific fault types at the equipment engineering level, thus overcoming the inability of purely data-driven models to provide interpretation. Furthermore, by combining the degradation rate reflected by the first derivative of the system potential function, a matching maintenance level (from observation to emergency replacement) is selected, ensuring that the urgency of maintenance actions corresponds to the actual rate of equipment degradation, avoiding premature downtime or delayed processing.

[0105] S150, Obtain feedback information after actual maintenance, and use the feedback information to update the adjustable parameters in the system potential energy function.

[0106] In this embodiment, after the maintenance work order is completed by the on-site maintenance personnel, the maintenance personnel enter the actual feedback information of the maintenance through the workshop terminal or handheld device.

[0107] In this embodiment, the feedback information includes at least: the actual location of the equipment failure (such as bearings, gearboxes, or seals), the actual cause of the failure, whether the failure is consistent with the prediction, and the time interval from the warning to the actual maintenance.

[0108] Optionally, cases where no abnormalities are found after an early warning is issued are recorded as false alarms; cases where a fault occurs without an early warning are recorded as missed alarms. All feedback information is uploaded to the cloud server via the Industrial Internet.

[0109] After receiving feedback information, the cloud server compares it with historical early warning records to generate sample data labeled with accurate, false, or missed warnings. With the goal of comprehensively improving early warning accuracy, a Bayesian hyperparameter optimization algorithm is used to asynchronously update the adjustable parameters in the system's potential function. These adjustable parameters include the energy well coefficient, nonlinear instability coefficient, curvature driving coefficient, long-range interaction coefficient, first derivative weight, and second derivative weight, resulting in optimized parameters with better predictive performance under current equipment operating conditions.

[0110] After the update is complete, the new parameters are distributed to the edge server via the Industrial Internet. The edge server loads the new parameters to replace the old model in the next operating cycle. At the same time, based on the statistical results of recent false alarms and missed alarms, the warning thresholds of fault symptom indicators are adjusted accordingly, enabling the system to continuously adapt to the long-term aging trend of equipment and changes in operating conditions.

[0111] The above process establishes a closed-loop verification channel between the prediction system and the physical world through feedback information from actual maintenance. Adjustable parameters in the system's potential function are updated using Bayesian hyperparameter optimization, allowing the model to continuously adapt to individual device differences and long-term aging drift, preventing static model performance from decaying over time. Simultaneously, thresholds are adjusted based on false alarm and missed alarm statistics, forming a complete perception-decision-execution-feedback-optimization evolutionary loop, achieving adaptive and continuous improvement in prediction performance.

[0112] This technical solution eliminates the time misalignment problem of heterogeneous data from different frequencies in industrial sites by collecting multi-source data and aligning it according to the operating cycle to generate a time-series matrix. The matrix is ​​then upgraded to a tensor, and nonlinear principal components are extracted. A degradation tensor sequence is generated using similarity modulation, enabling continuous quantitative characterization of the multi-dimensional nonlinear degradation process of equipment. Based on this, an off-field is constructed, and the system potential energy function is calculated. The rate of change of potential energy over time is used to generate fault symptom indicators, allowing accurate identification of equipment before it enters the rapid degradation stage. Dynamic thresholds effectively reduce false alarms. After triggering an early warning, the abnormal tensor position is traced back to the original feature space and mapped to a specific fault type, transforming mathematical anomalies into actionable maintenance suggestions, thus solving the problem of the uninterpretable nature of black-box models. Finally, maintenance feedback is used to update the potential energy function parameters, forming a self-evolving closed loop. This ensures that the predictive performance remains accurate as the equipment is used, thereby achieving early, interpretable, and adaptive predictive maintenance without the need for fault labels, improving the efficiency and reliability of industrial production.

[0113] The following describes embodiments of the equipment fault prediction and maintenance apparatus based on the Industrial Internet of Things (IIoT) of this application, which can be used to execute the equipment fault prediction and maintenance method based on the Industrial Internet of Things (IIoT) in the above embodiments of this application. It is understood that the equipment fault prediction and maintenance apparatus based on the Industrial Internet of Things (IIoT) can be a computer program (including program code) running on a computer device. For example, the apparatus can install industrial application software or industrial control management software to achieve industrial cloud computing of industrial big data generated during the production process through an industrial cloud platform. The apparatus can be used to execute the corresponding steps in the method provided in the embodiments of this application. For details not disclosed in the embodiments of the equipment fault prediction and maintenance apparatus based on the Industrial Internet of Things (IIoT) of this application, please refer to the embodiments of the equipment fault prediction and maintenance method based on the Industrial Internet of Things (IIoT) described above.

[0114] Figure 3 A block diagram of an apparatus for predictive maintenance of equipment failure based on the Industrial Internet of Things is shown according to one embodiment of this application.

[0115] Reference Figure 3 As shown, an apparatus for predictive maintenance of equipment faults based on the Industrial Internet, according to an embodiment of this application, includes:

[0116] The acquisition module 310 is used to collect multi-source monitoring data during the operation of the equipment, and to slice and align the multi-source monitoring data according to the operating cycle of the equipment to generate a time-series data matrix.

[0117] The conversion module 320 is used to convert the time-series data matrix into a three-dimensional data array, extract the nonlinear principal components of the three-dimensional data array, calculate the similarity between the current three-dimensional data array and the reference tensor, and modulate the three-dimensional data array according to the similarity to generate a degenerate tensor sequence.

[0118] The early warning module 330 is used to construct an deviation field based on the degenerate tensor sequence and the reference tensor, calculate the system potential energy function based on the deviation field, generate a fault symptom index based on the time rate of change of the system potential energy function, and issue an early warning when the fault symptom index exceeds a threshold.

[0119] The response module 340 is used to respond to the warning, trace back to the time series data matrix based on the position of the tensor element that contributes the most in the deviation field, determine the abnormal features, map the abnormal features to the fault prediction result, and output maintenance suggestions based on the time rate of change of the system potential energy function.

[0120] The update module 350 is used to obtain feedback information after actual maintenance and to update the adjustable parameters in the system potential energy function using the feedback information.

[0121] In this application, based on the aforementioned scheme, the multi-source monitoring data includes at least sensor signals, operating parameters, and operation records; slicing and aligning the multi-source monitoring data according to the equipment's operating cycle to generate a time-series data matrix includes: determining the start and end trigger signals of the slicing window according to the equipment's operating cycle; slicing the sensor signals of each type in the multi-source monitoring data according to the slicing window to generate slice signals; extracting time-domain statistical features from the slice signals; and arranging the extracted time-domain statistical features, the operating parameters, and the operation record events according to a unified time axis to form a time-series data matrix.

[0122] In this application, based on the aforementioned scheme, the step of converting the time-series data matrix into a three-dimensional data array, extracting the nonlinear principal components of the three-dimensional data array, calculating the similarity between the current three-dimensional data array and the reference tensor, and modulating the three-dimensional data array according to the similarity to generate a degenerate tensor sequence includes: taking a preset number of historical lag time points forward from the current time as the center, using the feature vector at each historical lag time point as a time slice of the three-dimensional data array, stacking all time slices along the time dimension to construct an initial three-dimensional data array, wherein the first dimension of the initial three-dimensional data array is the time lag order, the second dimension is the original feature variables, and the third dimension is the nonlinear principal components to be extracted; mapping the original feature variables in the second dimension of the initial three-dimensional data array to a high-dimensional space using a kernel function, and in the high-dimensional... The first few nonlinear principal component directions are extracted from the space, and the original feature variable dimension is compressed into the nonlinear principal component dimension to obtain a compressed three-dimensional data array. The original feature variable part of the initial three-dimensional data array that has not been extracted by nonlinear principal components is merged with the compressed three-dimensional data array along the third dimension to form a three-dimensional data array to be matched. The average three-dimensional data array after the same processing of the historical data of the device in a healthy state is used as the three-dimensional data array. The similarity between the three-dimensional data array to be matched at the current moment and the reference tensor is measured by dynamic time warping distance. The similarity is used as a weighting factor to weight each element in the three-dimensional data array to be matched to generate the degenerate three-dimensional data array at the current moment. The degenerate three-dimensional data arrays at each moment are arranged in chronological order to obtain the degenerate tensor sequence.

[0123] In this application, based on the aforementioned scheme, the step of constructing the deviation field according to the degenerate tensor sequence and the reference tensor includes: subtracting the reference tensor from the degenerate tensor at the current moment in the degenerate tensor sequence to obtain the deviation field.

[0124] In this application, based on the aforementioned scheme, the calculation of the system potential function based on the off-field includes: determining an energy potential well term based on the square value of each element in the off-field, determining a nonlinear instability term based on the fourth power value of each element in the off-field, determining a curvature driving term based on the second-order mixed partial derivative of the off-field in the spatiotemporal dimension, and determining a long-range interaction term based on the interaction between elements at different positions in the off-field; summing the energy potential well term, the nonlinear instability term, the curvature driving term, and the long-range interaction term to obtain the system potential function value at the current moment.

[0125] In this application, based on the aforementioned scheme, the step of generating a fault symptom index according to the time rate of change of the system potential energy function, and issuing an early warning when the fault symptom index exceeds a threshold, includes: performing numerical differentiation on the system potential energy function to obtain the first and second derivatives of the system potential energy function with respect to time; combining the first and second derivatives and mapping them to a preset interval to obtain the fault symptom index; using the statistical quantile of the fault symptom index in the historical normal window as the threshold; issuing an early warning when the current fault symptom index exceeds the threshold, and recording the position of the tensor element that contributes the most in the deviation field.

[0126] In this application, based on the aforementioned scheme, the response to the warning, tracing back to the time-series data matrix based on the position of the tensor element contributing the most in the deviation field, determining the abnormal features, mapping the abnormal features to the fault prediction result, and outputting maintenance suggestions based on the time rate of change of the system potential function, includes: calculating the partial derivative values ​​of the tensor element positions with respect to each original feature, selecting abnormal features according to the partial derivative values; matching the abnormal features with preset fault mapping rules, outputting the fault prediction type based on the matching result; selecting the maintenance level based on the fault prediction type and the value of the first derivative of the system potential function, and outputting a maintenance work order containing the fault prediction type, maintenance level, and maintenance suggestions.

[0127] This technical solution eliminates the time misalignment problem of heterogeneous data from different frequencies in industrial sites by collecting multi-source data and aligning it according to the operating cycle to generate a time-series matrix. The matrix is ​​then upgraded to a tensor, and nonlinear principal components are extracted. A degradation tensor sequence is generated using similarity modulation, enabling continuous quantitative characterization of the multi-dimensional nonlinear degradation process of equipment. Based on this, an off-field is constructed, and the system potential energy function is calculated. The rate of change of potential energy over time is used to generate fault symptom indicators, allowing accurate identification of equipment before it enters the rapid degradation stage. Dynamic thresholds effectively reduce false alarms. After triggering an early warning, the abnormal tensor position is traced back to the original feature space and mapped to a specific fault type, transforming mathematical anomalies into actionable maintenance suggestions, thus solving the problem of the uninterpretable nature of black-box models. Finally, maintenance feedback is used to update the potential energy function parameters, forming a self-evolving closed loop. This ensures that the predictive performance remains accurate as the equipment is used, thereby achieving early, interpretable, and adaptive predictive maintenance without the need for fault labels, improving the efficiency and reliability of industrial production.

[0128] Figure 4 A schematic diagram of the structure of a computer system suitable for implementing the computer device of the present application is shown.

[0129] It should be noted that the computer system of the computer device in this embodiment is only an example and should not impose any limitations on the function and scope of use of the embodiments of this application.

[0130] In this embodiment, the computer system includes a central processing unit 401, which can perform various appropriate actions and processes based on programs stored in read-only memory 402 or programs loaded from storage section 408 into random access memory 403, such as executing the equipment fault prediction and maintenance method based on the Industrial Internet described in the above embodiment. The random access memory 403 also stores various programs and data required for system operation, thereby realizing big data storage and big data management. The central processing unit 401, read-only memory 402, and random access memory 403 are interconnected via bus 404. Input / output interface 405 is also connected to bus 404.

[0131] The following components are connected to the input / output interface 405: an input section 406 including a keyboard, mouse, etc.; an output section 407 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 408 including a hard disk, etc.; and a communication section 409 including a network interface card such as a LAN (Local Area Network) card, modem, etc. The communication section 409 performs communication processing via a network such as the Internet. A drive 410 is also connected to the input / output interface 405 as needed. A removable medium 411, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 410 as needed so that computer programs read from it can be installed into the storage section 408 as needed.

[0132] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable storage medium, the computer program including a computer program for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 409, and / or installed from removable medium 411. When the computer program is executed by central processing unit 401, it performs various functions defined in the system of this application.

[0133] It should be noted that the computer-readable storage medium shown in the embodiments of this application can be a computer-readable signal medium, a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying a computer-readable computer program. The transmitted data signal can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium can also be any computer-readable storage medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The computer program contained on the computer-readable storage medium can be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination thereof.

[0134] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. Each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0135] The units described in the embodiments of this application can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.

[0136] One aspect of this application provides a computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in the various optional implementations described above.

[0137] On the other hand, this application also provides a computer-readable storage medium, which may be included in the computer device described in the above embodiments; or it may exist independently and not assembled into the computer device. The computer-readable storage medium carries one or more programs, which, when executed by the computer device, cause the computer device to implement the equipment fault prediction and maintenance method based on the Industrial Internet described in the above embodiments.

[0138] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to the embodiments of this application, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0139] Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, touch terminal, or network device, etc.) to execute the methods according to the embodiments of this application.

[0140] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the embodiments disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0141] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for predictive maintenance of equipment faults based on the Industrial Internet, characterized in that, include: Collect multi-source monitoring data during equipment operation, slice and align the multi-source monitoring data according to the equipment's operating cycle, and generate a time-series data matrix; The time-series data matrix is ​​converted into a three-dimensional data array, the nonlinear principal components of the three-dimensional data array are extracted, and the similarity between the current three-dimensional data array and the reference tensor is calculated. The three-dimensional data array is modulated according to the similarity to generate a degenerate tensor sequence. An deviation field is constructed based on the degenerate tensor sequence and the reference tensor. The system potential energy function is calculated based on the deviation field. A fault symptom index is generated based on the time rate of change of the system potential energy function. An early warning is issued when the fault symptom index exceeds a threshold. In response to the warning, the position of the tensor element that contributes the most in the deviation field is traced back to the time series data matrix to determine the abnormal features. The abnormal features are mapped to the fault prediction results, and maintenance suggestions are output based on the time rate of change of the system potential energy function. Obtain feedback information after actual maintenance, and use the feedback information to update the adjustable parameters in the system potential energy function.

2. The method for predictive maintenance of equipment faults based on the Industrial Internet according to claim 1, characterized in that, Multi-source monitoring data should include at least sensor signals, operating parameters, and operation records; The multi-source monitoring data is sliced ​​and aligned according to the equipment's operating cycle to generate a time-series data matrix, including: The start and end trigger signals of the slicing window are determined according to the equipment's operating cycle. The sensor signals of each type in the multi-source monitoring data are sliced ​​according to the slicing window to generate slice signals. Extract time-domain statistical features from the sliced ​​signal; The extracted time-domain statistical features, the operating parameters, and the running record events are arranged according to a unified time axis to form a time-series data matrix.

3. The method for predictive maintenance of equipment faults based on the Industrial Internet according to claim 1, characterized in that, The time-series data matrix is ​​converted into a three-dimensional data array, the nonlinear principal components of the three-dimensional data array are extracted, and the similarity between the current three-dimensional data array and the reference tensor is calculated. The three-dimensional data array is then modulated based on the similarity to generate a degenerate tensor sequence, including: Taking the current moment as the center, a preset number of historical lag time points are extracted. The feature vector at each historical lag time point is used as a time slice of the three-dimensional data array. All time slices are stacked along the time dimension to construct an initial three-dimensional data array. The first dimension of the initial three-dimensional data array is the time lag order, the second dimension is the original feature variable, and the third dimension is the nonlinear principal component to be extracted. The original feature variables in the second dimension of the initial three-dimensional data array are mapped to a high-dimensional space by a kernel function. The first few nonlinear principal component directions are extracted in the high-dimensional space, and the original feature variable dimension is compressed into the nonlinear principal component dimension to obtain the compressed three-dimensional data array. The original feature variables that have not been extracted by nonlinear principal components in the initial three-dimensional data array are merged with the compressed three-dimensional data array along the third dimension to form a three-dimensional data array to be matched. The average three-dimensional data array of the historical data of the device in a healthy state after the same processing is used as the three-dimensional data array. The dynamic time warping distance is used to measure the similarity between the three-dimensional data array to be matched at the current moment and the reference tensor. Using the similarity as a weighting factor, each element in the three-dimensional data array to be matched is weighted to generate the degenerate three-dimensional data array at the current time. The degenerate three-dimensional data arrays at each time are arranged in chronological order to obtain the degenerate tensor sequence.

4. The method for predictive maintenance of equipment faults based on the Industrial Internet according to claim 1, characterized in that, Constructing an offset field based on the degenerate tensor sequence and the reference tensor includes: The deviation field is obtained by subtracting the reference tensor from the current degenerate tensor in the degenerate tensor sequence.

5. The method for predictive maintenance of equipment faults based on the Industrial Internet according to claim 1, characterized in that, The system potential energy function is calculated based on the aforementioned off-field, including: The energy potential well term is determined based on the square value of each element in the off-field, the nonlinear instability term is determined based on the fourth power value of each element in the off-field, the curvature driving term is determined based on the second-order mixed partial derivative of the off-field in the spacetime dimension, and the long-range interaction term is determined based on the interaction between elements at different positions in the off-field. The system potential energy function value at the current moment is obtained by summing the energy potential well term, the nonlinear instability term, the curvature driving term, and the long-range interaction term.

6. The method for predictive maintenance of equipment faults based on the Industrial Internet according to claim 1, characterized in that, Fault symptom indicators are generated based on the time rate of change of the system potential energy function. When the fault symptom indicators exceed a threshold, an early warning is issued, including: By numerically differentiating the system potential energy function, we obtain the first and second derivatives of the system potential energy function with respect to time. The first and second derivatives are combined and mapped to a preset interval to obtain the fault symptom index; Using the statistical quantile of the fault symptom index in the historical normal window as a threshold, if the current fault symptom index exceeds the threshold, an early warning is issued, and the position of the tensor element that contributes the most in the deviation field is recorded.

7. The method for predictive maintenance of equipment faults based on the Industrial Internet according to claim 1, characterized in that, In response to the warning, the system traces back to the time-series data matrix based on the position of the tensor element contributing the most to the deviation field to determine the abnormal features. These abnormal features are then mapped to fault prediction results, and maintenance suggestions are output based on the time rate of change of the system potential function, including: Calculate the partial derivative values ​​of the tensor element positions with respect to each original feature, and select the abnormal features according to the partial derivative values; The abnormal features are matched with preset fault mapping rules, and the fault prediction type is output based on the matching results. Based on the fault prediction type and the value of the first derivative of the system potential function, the maintenance level is selected, and a maintenance work order containing the fault prediction type, maintenance level, and maintenance recommendations is output.

8. A device for predictive maintenance of equipment faults based on the Industrial Internet, characterized in that, include: The acquisition module is used to collect multi-source monitoring data during the operation of the equipment, and to slice and align the multi-source monitoring data according to the operating cycle of the equipment to generate a time-series data matrix. The conversion module is used to convert the time-series data matrix into a three-dimensional data array, extract the nonlinear principal components of the three-dimensional data array, calculate the similarity between the current three-dimensional data array and the reference tensor, and modulate the three-dimensional data array according to the similarity to generate a degenerate tensor sequence. The early warning module is used to construct an deviation field based on the degenerate tensor sequence and the reference tensor, calculate the system potential energy function based on the deviation field, generate a fault symptom index based on the time rate of change of the system potential energy function, and issue an early warning when the fault symptom index exceeds a threshold. The response module is used to respond to the warning, trace back to the time series data matrix based on the position of the tensor element that contributes the most in the deviation field, determine the abnormal features, map the abnormal features to the fault prediction result, and output maintenance suggestions based on the time rate of change of the system potential energy function. The update module is used to obtain feedback information after actual maintenance and to update the adjustable parameters in the system potential energy function using the feedback information.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When a computer program is executed by a processor, it implements the method for predictive maintenance of equipment faults based on the Industrial Internet as described in any one of claims 1 to 7.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method for predictive maintenance of equipment faults based on the Industrial Internet, as described in any one of claims 1 to 7.

Images