Knowledge graph and large model-based device health state evaluation method and device

Abstract

The application provides a device health state evaluation method and device based on a knowledge graph and a large model, relates to the technical field of industrial natural language processing and application, and performs frequency band decomposition and energy modeling on industrial equipment vibration signals, positions energy changes of sensitive frequency bands, and effectively captures transient impact and mutation information. Through a neural network, the dynamic response mechanism of the preset frequency band is used for feature extraction, the above target feature vector has strong distinguishability, can effectively identify the feature change trend closely related to the device health state, and abstracts the high-layer feature with high discrimination ability. In addition, the knowledge graph of the industrial equipment is combined for fusion reasoning of the device health state, so that the accuracy and interpretability of the diagnosis result are improved. The problems of traditional methods, such as difficulty in feature extraction, low classification precision and poor interpretability when facing non-stationary, nonlinear and multi-band coupled vibration signals, are effectively solved.

Description

Technical Field

[0001] The present invention relates to the field of industrial natural language processing and application technology, and in particular to a method and device for evaluating the health status of equipment based on a knowledge graph and a large model. Background Art

[0002] With the increasing intelligence and automation of industrial equipment, equipment operating status monitoring and fault diagnosis have gradually become important links in ensuring the stable, safe, and efficient operation of industrial systems. In particular, in critical equipment such as motors, compressors, and gearboxes, early fault identification and health status assessment are directly related to production efficiency, maintenance costs, and safety risks. Vibration signals, as an important representation of equipment operating status, have been widely used in equipment health monitoring systems due to their sensitivity to fault changes. However, the vibration signals actually collected often have complex characteristics such as non-stationarity, nonlinearity, and multi-band coupling. They are also easily affected by background noise and environmental interference, making it difficult for traditional feature extraction and classification methods to accurately identify fault types. Summary of the Invention

[0003] In view of this, the purpose of the present invention is to provide a method and device for equipment health status assessment based on knowledge graphs and large models, which can effectively solve the problems of traditional methods such as difficulty in feature extraction, low classification accuracy, and poor interpretability when facing non-stationary, nonlinear and multi-band coupled vibration signals.

[0004] In a first aspect, an embodiment of the present invention provides an equipment health status assessment method based on a knowledge graph and a large model, wherein the method includes: obtaining a vibration monitoring signal of industrial equipment during operation; decomposing the vibration monitoring signal into multiple frequency sub-bands, and extracting the energy distribution corresponding to each frequency sub-band; normalizing the vibration monitoring signal according to the energy distribution, and generating a target feature vector corresponding to the vibration monitoring signal; inputting the target feature vector into a preset neural network, performing feature extraction on the target feature vector through the neural network, and determining a feature to be measured corresponding to the target feature vector; wherein the feature to be measured is determined based on the dynamic response of the neural network to the preset frequency band; based on the feature to be measured and the equipment knowledge graph corresponding to the industrial equipment, classifying and evaluating the equipment health status of the industrial equipment, and determining a health status assessment result indicated by the industrial equipment.

[0005] In combination with the first aspect, an embodiment of the present invention provides a first implementation method of the first aspect, wherein the vibration monitoring signal is decomposed into multiple frequency sub-bands, and the step of extracting the energy distribution corresponding to each frequency sub-band includes: performing multi-scale decomposition of the vibration monitoring signal through a wavelet transform function to obtain frequency band local characteristics corresponding to different time scales; wherein each time scale corresponds to a preset frequency sub-band; determining the modulus response of the frequency band local characteristics in the current frequency sub-band; the modulus response is used to characterize the amplitude characteristics of the signal energy density indicated by the frequency band local characteristics; based on the modulus response, determining the energy distribution corresponding to the current frequency sub-band.

[0006] In combination with the first aspect, an embodiment of the present invention provides a second implementation of the first aspect, wherein the step of normalizing the energy distribution to generate a target feature vector corresponding to the vibration monitoring signal includes: calculating an energy normalization value corresponding to the energy distribution based on the energy distribution of each frequency sub-band; performing weighted combination of the energy distribution based on a preset weighting coefficient, and normalizing the result of the weighted combination based on the energy normalization value to generate a target feature vector corresponding to the vibration monitoring signal.

[0007] In combination with the first aspect, an embodiment of the present invention provides a third implementation of the first aspect, wherein the target feature vector is input into a preset neural network, feature extraction is performed on the target feature vector through the neural network, and the step of determining the feature to be measured corresponding to the target feature vector includes: inputting the target feature vector into the preset neural network, performing high-dimensional nonlinear mapping on the target feature vector according to the similarity of the target feature vector in the neighborhood, and determining the high-dimensional feature representation of the target feature vector; performing nonlinear amplitude amplification on the high-dimensional feature representation, and determining the frequency band response change corresponding to the high-dimensional feature representation; based on the frequency band response change, performing frequency band amplitude adjustment on the high-dimensional feature representation; based on the dynamic response of the frequency band amplitude adjustment indication, performing nonlinear conversion on the high-dimensional feature representation, and determining the feature to be measured corresponding to the target feature vector.

[0008] In combination with the first aspect, an embodiment of the present invention provides a fourth implementation of the first aspect, wherein, based on the similarity of the target feature vector in the neighborhood, the target feature vector is subjected to high-dimensional nonlinear mapping to obtain a high-dimensional feature representation of the target feature vector, including: identifying multiple local center vectors corresponding to the target feature vector through a preset clustering algorithm, and determining multiple fault modes indicated by the target feature vector; determining the neighborhood perception kernel width corresponding to the target feature vector based on the local center vector; using a preset exponential kernel function, according to the neighborhood perception kernel width and the local center vector, measuring the feature vector similarity corresponding to each fault mode indicated by the target feature vector; based on the feature vector similarity, performing high-dimensional nonlinear mapping processing on the target feature vector to determine the high-dimensional feature representation corresponding to the target feature vector.

[0009] In combination with the first aspect, an embodiment of the present invention provides a fifth implementation of the first aspect, wherein the equipment knowledge graph includes the fault phenomenon, fault location, fault root cause and fault solution of the industrial equipment; based on the features to be measured and the equipment knowledge graph corresponding to the industrial equipment, the equipment health status of the industrial equipment is classified and evaluated, and the steps of determining the health status evaluation result indicated by the industrial equipment include: classifying and predicting the features to be measured through a preset first prediction model, and outputting the equipment health status category probability indicated by the features to be measured; identifying the fault phenomenon characteristic parameters indicated by the features to be measured through a preset second prediction model, and matching the equipment fault phenomenon indicated by the fault phenomenon characteristic parameters from the equipment knowledge graph; locating the fault location and deriving the fault root cause of the industrial equipment based on the equipment health status category probability and the equipment fault phenomenon, and determining the health status evaluation result indicated by the industrial equipment.

[0010] In combination with the first aspect, an embodiment of the present invention provides a sixth implementation of the first aspect, wherein the first prediction model is constructed based on a preset neural network; the step of classifying and predicting the features to be measured through the preset first prediction model, and outputting the probability of the equipment health status category indicated by the features to be measured, includes: determining the statistical characteristics of the fault mode distribution corresponding to the vibration monitoring signal based on the target feature vector of the vibration monitoring signal; initializing the neural network weights of the first prediction model based on the statistical characteristics of the fault mode distribution; and performing differential constraints on the neural network weights based on the spatiotemporal correlation of the target feature vector; classifying and predicting the features to be measured based on the neural network weights of the first prediction model and the preset multi-objective loss function, and determining the probability of the equipment health status category corresponding to the features to be measured.

[0011] In combination with the first aspect, an embodiment of the present invention provides a seventh implementation method of the first aspect, wherein the method for calculating the multi-objective loss function includes: determining the classification accuracy loss of the neural network based on the training sample set corresponding to the feature to be tested and the category prediction probability corresponding to the training sample set; determining the robustness loss corresponding to the neural network based on the preset perturbation sample; and calculating the multi-objective loss function corresponding to the neural network based on the classification accuracy loss and the robustness loss.

[0012] In combination with the first aspect, an embodiment of the present invention provides an eighth implementation of the first aspect, wherein the above method also includes: determining the current error direction of the neural network corresponding to a preset training sample set based on a multi-objective loss function; calculating the first-order gradient information of the multi-objective loss function to the current weight of the neural network, as well as the second-order gradient information of the weights of each layer of the neural network; based on the first-order gradient information, the second-order gradient information and the preset weight attenuation term, updating the initial weights of the neural network to determine the final weight parameters of the neural network.

[0013] In a second aspect, an embodiment of the present invention provides an equipment health status assessment device based on a knowledge graph and a large model, wherein the device includes: a signal acquisition module for acquiring a vibration monitoring signal of industrial equipment during operation; a signal decomposition module for decomposing the vibration monitoring signal into multiple frequency sub-bands and extracting the energy distribution corresponding to each frequency sub-band; a data processing module for normalizing the vibration monitoring signal according to the energy distribution, and generating a target feature vector corresponding to the vibration monitoring signal; an execution module for inputting the target feature vector into a preset neural network, performing feature extraction on the target feature vector through the neural network, and determining a feature to be measured corresponding to the target feature vector; wherein the feature to be measured is determined based on the dynamic response of the neural network to the preset frequency band; an output module for classifying and assessing the equipment health status of the industrial equipment based on the feature to be measured and the equipment knowledge graph corresponding to the industrial equipment, and determining a health status assessment result indicated by the industrial equipment.

[0014] The embodiments of the present invention bring the following beneficial effects: The embodiments of the present invention provide a method and apparatus for assessing the health status of equipment based on a knowledge graph and a large model. These methods perform frequency band decomposition and energy modeling on industrial equipment vibration signals, extracting their structured energy distribution vectors and locating energy changes in sensitive frequency bands. These methods are capable of effectively capturing transient impacts and mutation information, enhancing robustness to noise and environmental interference. Furthermore, by extracting features through a neural network using its dynamic response mechanism to preset frequency bands, combined with the strong discriminability of the target feature vectors, these methods are capable of effectively identifying feature change trends closely related to the health status of the equipment and abstracting them into high-level features with high discriminative capabilities.

[0015] Furthermore, embodiments of the present invention integrate knowledge graphs of industrial equipment to perform fusion reasoning on the health status of equipment. This allows for precise location of fault locations and root causes based on semantic information such as equipment structure and historical failure patterns, improving the accuracy and interpretability of diagnostic results. In summary, this invention effectively addresses the challenges of traditional methods in extracting features, achieving low classification accuracy, and achieving poor interpretability when dealing with non-stationary, nonlinear, and multi-band coupled vibration signals.

[0016] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or understood by practicing the present invention. The purposes and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0017] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0019] Figure 1 A flowchart of a method for evaluating the health status of equipment based on a knowledge graph and a large model provided by an embodiment of the present invention;

[0020] Figure 2 A flowchart of another device health status assessment method based on knowledge graph and large model provided by an embodiment of the present invention;

[0021] Figure 3 A schematic diagram of a process framework and its implementation steps provided by an embodiment of the present invention;

[0022] Figure 4 A schematic diagram illustrating the effects of different preprocessing methods on feature separability provided by an embodiment of the present invention;

[0023] Figure 5 A schematic diagram showing the effect of different initialization strategies on the model convergence speed provided by an embodiment of the present invention;

[0024] Figure 6 A schematic diagram showing the effects of different activation functions on fault classification accuracy provided by an embodiment of the present invention;

[0025] Figure 7 A schematic diagram showing the comparative effects of different preprocessing methods provided in an embodiment of the present invention on the extraction capabilities of complex fault features;

[0026] Figure 8 Schematic diagram of the separation effect of complex fault modes according to an embodiment of the present invention;

[0027] Figure 9 A schematic diagram of the structure of an equipment health status assessment device based on a knowledge graph and a large model provided by an embodiment of the present invention;

[0028] Figure 10 A schematic structural diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0030] The present invention provides a method and device for equipment health status assessment based on knowledge graphs and large models, which can effectively solve the problems of traditional methods such as difficulty in feature extraction, low classification accuracy, and poor interpretability when facing non-stationary, nonlinear, and multi-band coupled vibration signals.

[0031] To facilitate understanding of this embodiment, firstly, a device health status assessment method based on knowledge graph and large model disclosed in an embodiment of the present invention is introduced in detail, see Figure 1 The flowchart of a method for evaluating the health status of equipment based on a knowledge graph and a large model is shown. The method may include the following steps:

[0032] Step S102: Acquire vibration monitoring signals of industrial equipment during operation.

[0033] Step S104: decompose the vibration monitoring signal into multiple frequency sub-bands, and extract the energy distribution corresponding to each frequency sub-band.

[0034] Step S106 : performing normalization processing on the vibration monitoring signal according to the energy distribution to generate a target feature vector corresponding to the vibration monitoring signal.

[0035] Raw vibration signals contain rich information about the equipment's status. However, due to factors such as changing operating conditions, fault evolution, and environmental interference, the statistical characteristics of the signal fluctuate dramatically over time, resulting in significant data non-stationarity. Furthermore, factors such as imperfect mechanical structure fit, impact friction, and material fatigue contribute to the signal's complex dynamic behavior and strong nonlinearity. These two factors work together to cause the same type of fault to exhibit highly inconsistent signal characteristics at different stages or under different circumstances, posing significant challenges to fault identification and status assessment.

[0036] To address this issue, embodiments of the present invention focus on the fact that different types of faults often correspond to specific frequency components (such as bearing fault frequencies and gear meshing frequencies). They perform frequency band decomposition and energy modeling on vibration monitoring signals from industrial equipment during operation, converting the signals into structured energy distribution vectors. Energy changes in sensitive frequency bands are located, enhancing robustness against noise and environmental interference. This approach can isolate local frequency anomalies caused by faults. Compared to conventional Fourier transform methods, it can effectively capture transient fault information (such as shocks and sudden changes). Furthermore, data normalization based on the corresponding energy distribution effectively suppresses the effects of sensor variations, gain drift, and background noise while preserving the overall energy distribution pattern of the signal. In summary, a highly discriminative target feature vector is constructed, preventing misleading model judgments caused by local anomalies. Embodiments of the present invention can enhance the feature's ability to distinguish different fault modes, improve the model's ability to identify faults early on, and further enhance the model's fault identification accuracy and robustness under complex operating conditions.

[0037] In step S108 , the target feature vector is input into a preset neural network, and feature extraction is performed on the target feature vector through the neural network to determine the feature to be measured corresponding to the target feature vector.

[0038] Step S110 , based on the features to be measured and the equipment knowledge graph corresponding to the industrial equipment, classify and evaluate the equipment health status of the industrial equipment, and determine the health status evaluation result indicated by the industrial equipment.

[0039] The embodiment of the present invention extracts features from a target feature vector through a neural network, and during the process of extracting features from the target feature vector by the neural network, the embodiment of the present invention determines the features to be measured based on the dynamic response of the neural network to a preset frequency band. In a specific implementation, the embodiment of the present invention can adaptively adjust the degree of attention paid to each frequency band based on the energy distribution of each frequency sub-band in different input samples, so as to effectively identify the key frequency bands that are closely related to the health status of the device, and abstract its changing trends into high-level features with strong discriminability, thereby determining the corresponding features to be measured.

[0040] Furthermore, semantic information in the device knowledge graph is combined with fused reasoning to determine the health status assessment results of the industrial equipment. In one embodiment, this invention constructs a knowledge graph for the equipment, associating knowledge about fault phenomena, fault locations, possible root causes, and solutions, thereby enabling fault diagnosis and reasoning. Based on the equipment's historical data, industry experience, and maintenance records, a knowledge graph is constructed that includes equipment components, failure modes, and maintenance solutions. The fault phenomena output by the machine learning model are matched with standard fault phenomena already in the knowledge graph. Furthermore, an inference engine (such as graph database inference based on SPARQL queries) can be used to associate the fault phenomena with the corresponding fault locations, root causes, and solutions. Furthermore, the root cause of the fault, such as "bearing wear" or "insufficient lubrication," can be output based on the model inference results combined with the knowledge graph. Based on the above steps, assume that in a real industrial environment, a motor on a production line is equipped with a vibration sensor. This motor may have a bearing fault or an electrical fault. After the system is started, the vibration signals of the motor can be collected in real time through sensors. These signals can be preprocessed by the device's edge computing device, and the preprocessed feature vectors are input into the trained neural network model. The model calculates the probability of various faults through forward propagation and determines the current health status of the motor as "bearing fault" or "electrical fault".

[0041] In summary, the embodiment of the present invention provides a method for assessing equipment health based on a knowledge graph and a large model. This method performs frequency band decomposition and energy modeling on industrial equipment vibration signals, extracts structured energy distribution vectors, and locates energy changes in sensitive frequency bands. This method can effectively capture transient impacts and mutation information, enhancing robustness to noise and environmental interference. Furthermore, by extracting features through a neural network using its dynamic response mechanism to preset frequency bands, combined with the strong discrimination of the target feature vectors, it can effectively identify feature change trends closely related to the equipment health status and abstract them into high-level features with high discriminative capabilities.

[0042] Furthermore, embodiments of the present invention integrate knowledge graphs of industrial equipment to perform fusion reasoning on the health status of equipment. This allows for precise location of fault locations and root causes based on semantic information such as equipment structure and historical failure patterns, improving the accuracy and interpretability of diagnostic results. In summary, this invention effectively addresses the challenges of traditional methods in extracting features, achieving low classification accuracy, and achieving poor interpretability when dealing with non-stationary, nonlinear, and multi-band coupled vibration signals.

[0043] Furthermore, in conjunction with the above embodiments, technical problems existing in the prior art include: 1. Conventional methods, such as time-domain statistical feature extraction and Fourier transform, cannot effectively process non-stationary, nonlinear, and multi-band coupled signals, resulting in the inability of features to accurately distinguish different fault modes. This is particularly true in complex equipment environments, where signal noise and coupling effects may lead to the loss of fault features. 2. Many neural network training methods rely on random initialization or traditional Gaussian distribution initialization, lacking a mechanism for adaptively adjusting weights based on the data distribution. This initialization method can lead to slow training convergence or trapping in local minima, impacting model training efficiency and ultimate performance. 3. Conventional methods typically use linear dimensionality reduction or simple low-dimensional projections, which cannot effectively capture key differences in high-dimensional nonlinear data. This results in insufficient fault mode discrimination and makes it difficult to adapt to the needs of high-dimensional, complex data. 4. Standard activation functions (such as ReLU and Sigmoid) cannot accurately capture the multi-band coupled features of complex fault signals, resulting in poor model performance when handling complex faults. 5. Traditional L1 or L2 regularization methods cannot flexibly adjust weights when processing data with high local feature similarity, which can cause noise points to significantly interfere with the training process and ignore the temporal and spatial correlations in the data. 6. Traditional methods often use a single loss function for optimization, which can lead to model overfitting or insufficient adaptability to specific fault types. This lacks sufficient consideration of model robustness and generalization capabilities.

[0044] In response to the above problems, an embodiment of the present invention provides another equipment health status assessment method based on knowledge graph and large model on the basis of the above embodiment. Figure 2 A flowchart of another device health status assessment method based on knowledge graph and large model provided by an embodiment of the present invention is shown. Figure 2 , including the following steps:

[0045] Step S202: Acquire vibration monitoring signals of industrial equipment during operation.

[0046] Step S204: decompose the vibration monitoring signal into multiple frequency sub-bands, and extract the energy distribution corresponding to each frequency sub-band.

[0047] Step S206 : performing normalization processing on the vibration monitoring signal according to the energy distribution to generate a target feature vector corresponding to the vibration monitoring signal.

[0048] Collected equipment vibration sensor data typically contains components across various frequency bands and is susceptible to background noise, making it difficult to directly obtain highly discriminative features for fault health status assessment and classification. Conventional techniques often use time-domain or simple frequency-domain statistical features to preprocess vibration signals. However, these methods are often ineffective for non-stationary, nonlinear, and multi-band coupled fault signals, making it difficult to accurately distinguish different types of equipment fault characteristics.

[0049] To address the above issues, the present invention uses multi-band energy analysis and wavelet transform to decompose and normalize the vibration signal in the preprocessing stage, and uses the superposition of multi-band features to improve the ability to distinguish different fault modes. Specifically, the following steps are included:

[0050] 1) Perform multi-scale decomposition of the vibration monitoring signal using a wavelet transform function to obtain local frequency band features corresponding to different time scales. Determine the modulus response of the local frequency band features in the current frequency sub-band. Based on the modulus response, determine the energy distribution corresponding to the current frequency sub-band.

[0051] The embodiment of the present invention uses a wavelet transform function (which can be expressed as ) Decomposes the vibration signal at multiple scales to determine the multi-band characteristics corresponding to the vibration monitoring data. These characteristics include the wavelet modulus, statistical energy, and instantaneous amplitude information for each frequency band, and comprehensively reflect the local abnormal patterns of the signal in the time-frequency domain. This allows the extraction of local characteristics of the signal at different time scales, capturing potential non-stationary and nonlinear components in the vibration signal and improving the ability to identify complex fault modes. In particular, compared to the Fourier transform, the wavelet transform solves the problem of frequency band coupling in non-stationary signals through multi-scale decomposition (high frequencies capture transient shocks, low frequencies reflect steady-state vibrations).

[0052] The modulus response is used to characterize the amplitude characteristics of the signal energy density in the frequency band, reflecting the instantaneous energy concentration (i.e., energy distribution) of the frequency band under the current signal, so that the preprocessing can capture the non-stationary characteristics of the signal from multiple scales. After the vibration signal is subjected to wavelet transform in the time domain sampling sequence, the modulus response in the pth frequency band can be obtained. .

[0053] 2) Based on the energy distribution of each frequency subband, an energy normalization value corresponding to the energy distribution is calculated; the energy distribution is weighted and combined based on a preset weighting coefficient; and the result of the weighted combination is normalized based on the energy normalization value to generate a target feature vector corresponding to the vibration monitoring signal.

[0054] In combination with the above steps, the embodiment of the present invention decomposes the vibration signal into multiple frequency sub-bands with physical significance, extracts and normalizes the energy distribution of each sub-band, and thus performs multi-band energy analysis on the vibration monitoring signal, so as to identify weak disturbances that may exist in the fault signal from different frequency bands and improve the separability of the fault signal in the frequency domain.

[0055] The energy normalization term of each frequency band is the ratio of the sum of the energy of all frequency bands of the signal in the sample to the energy of the maximum frequency band. It is used to eliminate the dimension difference of the signal amplitude. The calculation method is expressed as:

[0056]

[0057] Where, E i is the energy normalization term of the i-th signal in each frequency band, indicating the corresponding normalized value; is the maximum value function. is the above modulus response, p is used to represent the p-th frequency band and is a positive integer; P is the total number of frequency bands.

[0058] Furthermore, the energy, weighting coefficients and other information of each frequency band are combined and calculated to highlight the differences between different fault characteristics while reducing noise, which is expressed as:

[0059]

[0060] Where, is the ith preprocessed vibration signal feature vector, representing the data after filtering, normalization and multi-band energy superposition; i is a positive integer; E i is the energy normalization term of the i-th signal in each frequency band; p is a positive integer; P is the total number of frequency bands; is the weighting coefficient of the pth frequency band, which represents the importance of this frequency band in fault feature extraction; is the sampling sequence of the i-th vibration signal in the time domain. It should be noted that, and The weighting coefficient and modulus response representing the pth frequency band. The combination of the two effectively highlights the energy and characteristics of different frequency bands. Calculation through this weighted combination can amplify the differences between different fault characteristics while reducing noise, thereby improving the accuracy of fault diagnosis.

[0061] In step S208 , the target feature vector is input into a preset neural network, and a high-dimensional nonlinear mapping is performed on the target feature vector according to the similarity of the target feature vector in the neighborhood to determine a high-dimensional feature representation of the target feature vector.

[0062] Equipment vibration signals often exhibit nonlinear distributions in high-dimensional feature spaces. Similar fault modes are not necessarily clustered together in the original space, and direct use of linear dimensionality reduction or mapping methods can easily lead to a loss of fault differentiation. Conventional techniques typically employ linear mapping or simple low-dimensional projections, but these mapping methods are limited in effectiveness when processing high-dimensional, nonlinear fault data and struggle to preserve key differences.

[0063] The present invention uses a nonlinear mapping method to map the pre-processed vibration signal feature vector to a new high-dimensional space in a neighborhood in the form of similarity weighting, thereby retaining more nonlinear structures related to the fault. In specific implementation, the embodiment of the present invention combines the exponential kernel function with the high-dimensional nonlinear mapping function to make similar fault data close to each other in the mapping space, while widening the distance between different fault modes. Among them, in the embodiment of the present invention, the exponential kernel function is combined with the high-dimensional nonlinear mapping function in the form of a multiplication operation between the exponential kernel function and the high-dimensional nonlinear mapping function. Specifically, the following steps are included:

[0064] 1) Through a preset clustering algorithm, multiple local center vectors corresponding to the target feature vector are identified, and multiple fault modes indicated by the target feature vector are determined.

[0065] By using clustering algorithms (such as K-Means, density peak, etc.) on all pre-processed vibration features, K representative centers (i.e., k local center vectors, such as (where k is the kth local center vector). These centers represent the "main distribution regions" of the data in high-dimensional space. The closer to the cluster center, the more distinct the fault characteristics, making classification easier. Each principal component region corresponds to a fault mode and is further used as a sensing point for a predefined kernel function for feature perception.

[0066] 2) Based on the local center vector, determine the neighborhood perception kernel width corresponding to the target feature vector.

[0067] The neighborhood-aware kernel width can be dynamically calculated based on the average distance of samples around the kth local center vector, where the definition is is the kth kernel width, which is adaptively adjusted according to the distance between samples in the cluster. The kernel width is larger in sparse areas (to avoid overfitting) and smaller in dense areas (to capture detail differences). The specific calculation method is ,in, is the jth preprocessed vibration signal feature vector, is a small positive number that avoids division by 0. Preferably, Set to 10 -6 . is the kth local center vector.

[0068] 3) Using the preset exponential kernel function, the similarity of the feature vectors corresponding to each fault mode indicated by the target feature vector is measured according to the neighborhood perception kernel width and the local center vector.

[0069] 4) Based on the similarity of feature vectors, the target feature vector is subjected to high-dimensional nonlinear mapping to determine the high-dimensional feature representation corresponding to the target feature vector.

[0070] The high-dimensional nonlinear mapping function implements nonlinear expansion of input features in a locally weighted manner to enhance the modeling capability of complex boundaries. The calculation method is expressed as:

[0071]

[0072] Where, is a high-dimensional nonlinear mapping function. is the mapping coefficient of the kth kernel center, and is weighted according to the distribution of sample category labels at the kth local center vector to represent the category weight of each local center. This allows the sensitivity to different categories to be adaptively adjusted during the mapping process, improving the ability to distinguish different fault modes. It uses label information weighting to make similar fault samples in the mapping space more similar (for example, when the proportion of similar labels in the neighborhood is high, The specific calculation method is: , where N k represents the sample set that falls into the neighborhood of the k-th kernel center, Indicates N k The number of samples in y i Encode the label of the i-th sample.

[0073] The method of processing the preprocessed vibration signal feature vector using the above exponential kernel function is expressed as ,in, is an exponential function with a natural constant as its base, for The i-th neighborhood data point of is the kernel width parameter, which characterizes the influence range of the neighborhood. Preferably, Set to 0.01. Based on this, the i-th vibration signal feature vector mapped to the new high-dimensional space is expressed as .

[0074] Step S210 : performing nonlinear amplitude amplification on the high-dimensional feature representation to determine a frequency band response change corresponding to the high-dimensional feature representation.

[0075] The nonlinear amplitude amplification factor corresponding to the frequency band can be determined by calculating the variance of the corresponding feature of the i-th vibration signal feature vector mapped to the new high-dimensional space in the entire data set (such as ), emphasizing the response changes in certain frequency bands (such as small anomalies in high frequencies) to perform nonlinear amplification on the features. Frequency bands with large variances are more sensitive to fault changes and should therefore be given greater nonlinear amplification. The specific calculation method is:

[0076]

[0077] in, is the nonlinear amplitude amplification factor of the p-th frequency band, It represents the variance of the corresponding feature of the i-th vibration signal feature vector mapped to the new high-dimensional space in the entire data set. is the mapping eigenvalue on the pth frequency band. ε is a small constant to prevent the denominator from being zero, for example, it is set to 0.000001.

[0078] Step S212: adjusting the frequency band amplitude of the high-dimensional feature representation based on the frequency band response change.

[0079] By normalizing the amplitude of the p-th frequency band of each sample and then taking the average value in all samples, the amplitude adjustment coefficient corresponding to each frequency band of the high-dimensional feature representation is determined (as defined by For the The amplitude adjustment coefficient of each frequency band is used to control the relative influence of each frequency band feature on neuron activation, thereby estimating the overall weight of the frequency band, so that the network can dynamically respond according to the feature importance of different frequency bands to adjust the frequency band amplitude of high-dimensional feature representation.

[0080] Step S214 : Based on the dynamic response of the frequency band amplitude adjustment indication, a nonlinear transformation is performed on the high-dimensional feature representation to determine the feature to be measured corresponding to the target feature vector.

[0081] There are strong nonlinearities and coupling characteristics in fault vibration signals. The use of traditional ReLU or Sigmoid activation functions may not be able to accurately characterize multimodal fault characteristics, and the discrimination of complex signals is limited. When faced with high-order nonlinearities or frequency band coupling, the output of conventional activation functions is prone to saturation or over-suppression, and cannot flexibly adapt to various fault modes. In combination with the above steps, during the forward propagation of the neural network, the present invention combines the different frequency bands and amplitude components of high-dimensional signal characteristics in the activation function to perform nonlinear conversion in a multi-coupling manner to capture more potential fault characteristics, flexibly respond to high-frequency and low-frequency coupling characteristics, and improve nonlinear modeling capabilities. The calculation method is expressed as:

[0082]

[0083] Where, is the activation function of the neural network. is the mapping eigenvalue on the p-th frequency band; For the Amplitude adjustment coefficient for each frequency band.

[0084] Step S216: Based on the features to be measured and the equipment knowledge graph corresponding to the industrial equipment, classify and evaluate the equipment health status of the industrial equipment to determine the health status evaluation result indicated by the industrial equipment.

[0085] In one embodiment, the following steps may be included:

[0086] 1) Classify and predict the features to be measured using a preset first prediction model, and output the probability of the device health status category indicated by the features to be measured.

[0087] The first prediction model of the embodiment of the present invention can be constructed based on a preset neural network. Specifically, it can be constructed through the following steps:

[0088] a- Based on the target feature vector of the vibration monitoring signal, determine the statistical characteristics of the fault mode distribution corresponding to the vibration monitoring signal. Based on the statistical characteristics of the fault mode distribution, initialize the neural network weights of the first prediction model.

[0089] The distribution of vibration sensor data under different equipment failure modes varies greatly. The data distribution of some fault types is concentrated in local areas, while other fault types present more dispersed high-dimensional features. If the initialization is unreasonable, the network is prone to slow training convergence or falling into local extreme values.

[0090] Conventional neural network initialization methods mostly use random initialization or Gaussian distribution initialization, which makes it difficult for the network to identify the differences between different fault types in the initial stage. When faced with high-dimensional nonlinear data, the initial weights are too random, which often leads to low convergence efficiency. The present invention combines the statistical characteristics of the vibration signal distribution and adaptively sets the network initial weights, which allows the model to distinguish the distribution of different fault modes in the initial stage. The statistical quantities such as the mean and square norm of the vibration signal data are used to assign them to the initial weights, which accelerates subsequent training and reduces unnecessary parameter fluctuations. In the specific implementation, the initialization weights are determined with reference to the following formula:

[0091]

[0092] Where, is the initialized neural network weight matrix. is the initialization scaling factor, which represents the amplitude of weight initialization; is a polynomial weighting factor used to incorporate the data square norm information. ne is the dimension of the neural network input feature (preprocessed vibration signal feature vector); me is the number of training data (preprocessed vibration signals); D is the training dataset, which contains all feature vectors and label vectors; y is the label vector, which represents the label of the training data (preprocessed vibration signals); is the L2 norm. Preferably, Set to 0.3, Set to 0.5.

[0093] It should be noted that in the calculation method of the initialized neural network weight matrix, By multiplying the feature mean and the label mean, the data distribution trend is encoded into the initial weights (for example, the vibration energy of a certain type of fault is generally higher), making the network sensitive to the statistical differences between different categories at the initial stage; Using the L2 norm of features and labels, Balance the amplitude influence to avoid the initial weight being biased towards certain categories due to large differences in feature amplitude (for example, the energy of high frequency band is much higher than that of low frequency band).

[0094] Furthermore, embodiments of the present invention differentially constrain neural network weights based on the spatiotemporal correlation of target feature vectors. Collected vibration signals typically exhibit time-varying or spatial correlations. In particular, the characteristics of similar fault points are often quite similar, while noise points may differ significantly from the distribution of the majority of data. Conventional techniques typically employ global regularization constraints, such as L1 or L2 regularization, but do not impose specific constraints on locally similar fault features and are unable to dynamically suppress the impact of noise on some weights.

[0095] The present invention uses a regularization constraint method based on neighborhood differences to differentially constrain the weight parameters of the neural network within a local neighborhood, and adaptively adjusts the regularization term based on spatiotemporal dependencies, so that the parameters corresponding to similar fault modes are more consistent, and noise points cannot dominate the training. Among them, the spatiotemporal correlation considers the similarity of the feature space and the temporal or spatial differences of the sampling, and is used to characterize the dynamic correlation strength between fault data. The calculation method is expressed as:

[0096]

[0097] Where, is the ith vibration signal feature vector mapped to the new high-dimensional space and the jth vibration signal feature vector mapped to the new high-dimensional space The spatiotemporal correlation between the two samples is considered, which takes into account both feature similarity and temporal proximity (e.g., continuously sampled vibration signals have temporal correlation) to avoid misjudgment due to time drift. is the width of the spatiotemporal structure perception kernel; t i is the sampling time of the i-th vibration signal feature vector; t j is the sampling time of the jth vibration signal feature vector. Preferably, Set to 0.04.

[0098] Furthermore, by using the distance between neighbors and the exponential kernel function to bind the weight difference to the similarity of the data, while considering the global structure, we further focus on local correlation to differentially constrain the weights of the neural network. Specifically, the calculation method of the regularization term corresponding to the neural network is referred to the following formula:

[0099]

[0100] Where, It is a regularization term based on neighborhood differences, which represents the constraints on the weight parameters of the neural network, ensuring that the weights learned by the network are more consistent between similar fault modes, and will not have too much impact on noisy data. is the jth weight parameter of the neural network; is the i-th weight parameter of the neural network. is the exponential kernel function mentioned above, for The i-th neighborhood data point; is the eigenvector of the vibration signal after preprocessing of i; is the kernel width parameter. Further, is the regularization coefficient, which controls the regularization strength. Preferably, Set to 0.2. is the covariance adaptive adjustment coefficient, which is used to correct the weight difference in the spatiotemporal structure of neighborhood differences.

[0101] b- Based on the neural network weights of the first prediction model and the preset multi-objective loss function, the features to be measured are classified and predicted to determine the probability of the device health status category corresponding to the features to be measured.

[0102] In this embodiment of the present invention, the neural network weights of the first prediction model can be determined by the following steps:

[0103] i- Based on the multi-objective loss function, determine the current error direction of the neural network corresponding to the preset training sample set.

[0104] ii-Calculate the first-order gradient information of the multi-objective loss function to the current weight of the neural network, as well as the second-order gradient information to the weights of each layer of the neural network.

[0105] iii-Based on the first-order gradient information, the second-order gradient information and the preset weight decay term, the initial weight of the neural network is updated to determine the final weight parameters of the neural network.

[0106] In equipment fault health status assessment and classification, gradients tend to vanish or explode when the number of network layers and input feature dimensionality are high. This is especially true when fault modes are coupled, making it difficult for deep networks to coordinate the learning process of features at different levels. While conventional backpropagation methods can update weights layer by layer, they lack adaptive control over the evolution of gradients across different layers for high-dimensional nonlinear vibration data, resulting in insufficient training efficiency and stability.

[0107] The present invention uses an enhanced multi-layer network learning method to combine first-order and second-order gradient information at each update and recalibrate large weights, so that the network can suppress irrelevant or noise features while maintaining sensitivity to key fault characteristics. Specifically, the embodiment of the present invention combines the current gradient, second-order gradient, and weight decay term to determine the weight update amount of the neural network. The initial weights of the neural network are updated based on the weight update amount to determine the final weight parameters of the neural network, so as to achieve a more stable fault mode training process in high-dimensional vibration signals. The calculation method of the weight update amount is expressed as:

[0108]

[0109] Where, The weight update amount of the neural network, which represents how the weight changes in each training, is added to the weight of the current neural network and used as the neural network weight parameter for the next iteration; is the learning rate, which represents the step size of each update. is the weight attenuation factor, which controls the attenuation speed; int(t) is the value of the number of iterations, which represents the index value of the current number of iterations. Set to 0.005, Set to 0.3, Set to 0.2.

[0110] also, is the number of weight parameters of the neural network; It is the gradient of the loss function of the neural network with respect to the weight (also known as the first-order gradient), which represents the current error direction of the model. is the adaptive gradient coefficient, which represents the impact on the second-order gradient; is the second-order gradient of the loss function of the neural network with respect to the kth weight; k is a positive integer. W represents the matrix of all trainable weight parameters in the entire network, W kCharacterize the kth specific parameter of W. The second-order gradient can utilize the curvature information of the loss function to guide parameter updates to be more stable and in a more reasonable direction, thereby improving the training convergence speed and accuracy of deep neural networks in high-dimensional nonlinear vibration signals.

[0111] Furthermore, the embodiment of the present invention also sets a recalibration coefficient when calculating the weight update amount. , which is used to dynamically compress larger weights. In specific implementations, the recalibration coefficient controls the compression factor by averaging the weights, preventing excessive weights from dominating the learning process while improving the network's stability to small sample perturbations, expressed as:

[0112]

[0113] Where, is the recalibration reference factor; r is a control index greater than 1. Preferably, Set α to 0.1 and r to 2.

[0114] Furthermore, in actual industrial applications, the health status assessment model of equipment is easily affected by different working conditions and environments, so high requirements are placed on the robustness and generalization ability of the model; conventional methods only use a single loss function (such as classification cross entropy), which may result in insufficient robustness when the accuracy is improved, or excellent performance in specific scenarios but difficulty in adapting to new fault scenarios.

[0115] The present invention combines classification accuracy, robustness, and a regularization term based on neighborhood differences by setting a multi-objective optimization loss function to obtain network parameters with more global optimal properties. Specifically, the multi-objective loss function is determined by the following steps:

[0116] i-Determine the classification accuracy loss of the neural network based on the training sample set corresponding to the feature to be tested and the category prediction probability corresponding to the training sample set.

[0117] The classification accuracy loss can be calculated using the cross-entropy loss function based on the difference between the class probabilities output by the neural network and the true labels. The feature vectors output by the neural network can be fed into the Softmax function to convert them into predicted probabilities for each class. The log-likelihood is then constructed with the true labels to determine the degree of penalty for misjudgment of each class, which in turn guides weight updates.

[0118] ii- Based on the preset perturbation samples, determine the robustness loss corresponding to the neural network.

[0119] Robustness loss encourages the model to maintain stable output to input perturbations and prevent overfitting to local anomalies. The calculation method is expressed as:

[0120]

[0121] Where, is the prediction mapping for the i-th input under disturbance or noise.

[0122] The prediction mapping of the i-th input under disturbance or noise conditions adopts the disturbance enhancement mapping method. Specifically, a certain range of disturbance is added to the i-th preprocessed vibration signal feature vector to generate a disturbance sample. The disturbance method is:

[0123]

[0124] Where, For the perturbation samples; is the disturbance amplitude coefficient, simulating measurement noise or local disturbance; is a matrix with mean 0 and variance as the unit Preferably, Set to 0.001.

[0125] Furthermore, the perturbation sample is input into the mapping network with the same number of neural networks as the current training iterations to obtain the perturbation mapping output, which is expressed as:

[0126]

[0127] Where, The mapping network function is the same as the neural network of the current training iteration number. The mapping process is consistent with the original input and maintains structural consistency.

[0128] iii- Based on the classification accuracy loss and robustness loss, calculate the multi-objective loss function corresponding to the neural network.

[0129] In summary, the calculation method of the multi-objective loss function is expressed as:

[0130]

[0131] Where, It is a multi-objective loss function that characterizes the loss calculation of the neural network and comprehensively considers the optimization goals of classification accuracy and model robustness; is the number of batch samples input to the neural network; is the classification accuracy weight coefficient; is the robustness weight coefficient; is the classification accuracy loss; is the robustness loss. Preferably, Set to 0.3, Set to 0.2.

[0132] Furthermore, by continuously updating the weights until the termination condition is met, each round of training performs forward and backward propagation based on the current loss function. Training stops when any of the following conditions are met: 1) the training loss decreases by less than a set threshold over several consecutive rounds, indicating that convergence has been achieved; 2) the model's accuracy on the validation set does not increase for more than a set number of rounds to prevent overfitting; 3) the maximum number of iterations is reached.

[0133] After the neural network model (i.e., the first prediction model) is trained, fault diagnosis is performed on new, unseen equipment vibration signals to determine the probability of the equipment health status category corresponding to the measured features. In conjunction with the above-described embodiment and steps, this embodiment of the present invention uses a fully connected neural network to process the equipment's vibration sensor data. The input vibration signal undergoes the same preprocessing steps described above. The preprocessed signal feature vector is then forward propagated through the trained neural network. Nonlinear mapping is performed using the weighted activations of neurons in each layer, ultimately outputting a prediction of the equipment fault category.

[0134] 2) Identifying the fault phenomenon characteristic parameters indicated by the characteristics to be measured through a preset second prediction model, and matching the device fault phenomenon indicated by the fault phenomenon characteristic parameters from the device knowledge graph.

[0135] 3) Based on the equipment health status category probability and equipment failure phenomenon, the fault location of industrial equipment and the root cause of the failure are located, and the health status assessment results indicated by the industrial equipment are determined.

[0136] In one embodiment, a large model can be used as a second prediction model. Based on the semantic understanding ability of the large model, a detailed fault diagnosis report can be generated according to the user input and the reasoning results of the knowledge graph. In combination with the above steps, the user inputs the current operating status or fault phenomenon of the device. Furthermore, the large model combines the above-mentioned features to be tested, the fault information of the knowledge graph and the user description to generate a detailed diagnostic report. For example, possible causes of the fault and solutions. In one embodiment, the report may output: "According to the analysis of vibration data, the equipment bearings have slight wear and tear, and it is recommended to inspect and lubricate within 30 days."

[0137] in, Figure 3 The following figure shows the process framework and implementation steps corresponding to the embodiment of the present invention. Figure 3 In this example, vibration sensor data is fed into the machine learning module (the first prediction model) to predict the equipment's health status. Furthermore, the knowledge graph is used to infer fault phenomena, and the large model module (the second prediction model) generates a detailed diagnostic report.

[0138] In summary, the present invention provides another equipment health assessment method based on a knowledge graph and a large model. By preprocessing vibration signals using a combination of multi-band energy analysis and wavelet transform, this method effectively decomposes the multi-band components in the signal, improving the ability to distinguish different fault modes. The combination of multi-band energy superposition and wavelet transform improves the separability of fault features. Furthermore, during neural network training, the present invention adaptively sets initial weights based on the statistical characteristics of the vibration signal (such as the mean and square norm). Compared to traditional random initialization methods, this method can better identify the distribution of different fault modes in the early stages of training, significantly speeding up convergence and reducing unnecessary fluctuations during training. Furthermore, the present invention uses a nonlinear mapping method to map the vibration signal feature vectors to a new high-dimensional space. Combining an exponential kernel function with a high-dimensional nonlinear mapping function can better preserve the nonlinear structure in the fault data and enhance sensitivity to complex fault modes. Furthermore, the activation function proposed in the present invention combines the amplitude adjustment coefficients and nonlinear amplification factors of different frequency bands, which can more finely characterize the high-frequency and low-frequency coupling characteristics in the signal and has stronger recognition capabilities for complex multimodal faults.

[0139] Furthermore, by adopting regularization constraints based on neighborhood differences in the neural network, the weights of the neural network can be dynamically adjusted so that the weights of similar fault modes are more consistent, and the influence of noise points on training is minimized, which can better suppress the influence of noise and enhance the recognition of local fault features. Furthermore, the embodiment of the present invention comprehensively considers classification accuracy, robustness and regularization terms based on neighborhood differences to determine a multi-objective loss function, so that the model has stronger robustness and can adapt to different fault modes and environmental conditions. In the process of equipment health status assessment, combined with the reasoning ability of knowledge graphs and large models, intelligent reasoning can be performed based on existing information such as fault phenomena, root causes and solutions to generate detailed diagnostic reports, thereby improving the application effect of the model in actual industrial environments.

[0140] Furthermore, the effectiveness of the embodiment of the present invention is verified by the following experiments:

[0141] First, the embodiment of the present invention also provides a schematic diagram of the influence of different preprocessing methods on feature separability, referring to Figure 4, the embodiment of the present invention verifies the advantages of the preprocessing technology of multi-band energy analysis combined with wavelet transform proposed by the present invention by comparing the effects of different preprocessing methods on feature separability. The traditional time domain statistical features, fast Fourier transform frequency domain analysis, traditional wavelet transform and other methods are compared with the method of the present invention. The results show that the method of the present invention can more effectively decouple the multi-band components in the vibration signal and significantly improve the distinguishability of fault features. The experimental results show that the method of the present invention is significantly superior to other methods in terms of feature separability indicators. The weighted multi-band energy calculation method and adaptive wavelet transform processing can simultaneously consider the energy distribution and weight coefficients of different frequency bands, thereby highlighting the differences in various fault features while reducing noise.

[0142] The embodiment of the present invention also analyzes the neural network initialization method and verifies the impact of different initialization strategies on the model convergence speed. Figure 5 The figure shows the effect of different initialization strategies on the model convergence speed. Figure 5 The experiment compared traditional methods such as random initialization, Xavier initialization, and He initialization with the adaptive initialization method based on data distribution proposed in this invention. The loss curve during the training process shows that the method of this invention sets initial weights by combining the statistical characteristics and distribution patterns of the vibration signal, enabling the network to better distinguish different fault modes early in training. This not only accelerates convergence but also achieves lower final loss values. This demonstrates the effectiveness of encoding statistics such as the data mean and squared norm into the initial weights, avoiding the low training efficiency caused by traditional random initialization.

[0143] The embodiment of the present invention also verifies the superiority of the frequency band coupling activation function designed by the present invention by analyzing the influence of different activation functions on the fault classification accuracy. Figure 6 The figure shows the effect of different activation functions on the fault classification accuracy. Figure 6 The experiment selected common activation functions such as the rectified linear unit, leaky rectified linear unit, and self-gated activation function as comparison baselines and tested them on different types of equipment fault data. Experimental results show that the activation function proposed in this paper, by using a frequency band amplitude adjustment coefficient and a nonlinear amplification factor, can more finely characterize the coupling characteristics of high-frequency and low-frequency components in vibration signals. It is particularly capable of identifying complex fault modes such as gear wear. This activation mechanism, tailored to the characteristics of vibration signals, effectively overcomes the limitations of traditional activation functions in processing multimodal fault characteristics.

[0144] The embodiment of the present invention also provides a schematic diagram of the comparative effect of different preprocessing methods on the extraction ability of complex fault features, referring to Figure 7The embodiment of the present invention compares the ability of different preprocessing methods to extract complex fault features through the visualization of three-dimensional feature space distribution, and compares the differences in feature decoupling effects between multi-band energy analysis and traditional wavelet transform and Fourier transform. The results show that the features extracted by this technology present a clearer and more orderly spiral distribution structure in three-dimensional space, and the overlapping area between various fault feature points is significantly reduced. This superior feature separability intuitively demonstrates that the method of combining multi-band energy normalization with wavelet transform can effectively separate the coupled frequency band components in the mixed fault signal, providing more discriminative feature input for subsequent classifiers.

[0145] The embodiment of the present invention also verifies the improvement of the complex fault mode separation effect by nonlinear mapping through decision boundary visualization in feature space. Figure 8 A schematic diagram showing the separation effect of complex failure modes is shown. Figure 8 The embodiment of the present invention compares the performance of the high-dimensional nonlinear mapping function of the present invention and the traditional linear mapping in forming the decision boundary. It can be clearly observed that the decision boundary generated by this technology has richer curvature changes and larger classification intervals, which can better adapt to the complex distribution of fault characteristics in actual industrial scenarios, proving the effectiveness of the exponential kernel function and the local weighting strategy.

[0146] The embodiment of the present invention also provides a device health status assessment device based on knowledge graph and large model, Figure 9 FIG1 shows a schematic diagram of the structure corresponding to an embodiment of the present invention. Figure 9 The device includes: a signal acquisition module 10, which is used to obtain the vibration monitoring signal of the industrial equipment during operation; a signal decomposition module 20, which is used to decompose the vibration monitoring signal into multiple frequency sub-bands and extract the energy distribution corresponding to each frequency sub-band; a data processing module 30, which is used to normalize the vibration monitoring signal according to the energy distribution and generate a target feature vector corresponding to the vibration monitoring signal; an execution module 40, which is used to input the target feature vector into a preset neural network, extract features of the target feature vector through the neural network, and determine the feature to be measured corresponding to the target feature vector; wherein the feature to be measured is determined based on the dynamic response of the neural network to the preset frequency band; an output module 50, which is used to classify and evaluate the equipment health status of the industrial equipment based on the feature to be measured and the equipment knowledge graph corresponding to the industrial equipment, and determine the health status evaluation result indicated by the industrial equipment.

[0147] An embodiment of the present invention provides an equipment health status assessment device based on a knowledge graph and a large model. Its implementation principle and the technical effects produced are the same as those of the aforementioned method embodiment. For the sake of brief description, for matters not mentioned in the device embodiment, reference can be made to the corresponding content in the aforementioned method embodiment.

[0148] An embodiment of the present invention further provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-mentioned Figures 1 to 2 The embodiment of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to execute the above-mentioned Figures 1 to 2 The embodiment of the present invention also provides a structural diagram of an electronic device, such as Figure 10 FIG. 1 is a schematic diagram of the structure of the electronic device, wherein the electronic device includes a processor 101 and a memory 100, the memory 100 stores computer executable instructions that can be executed by the processor 101, and the processor 101 executes the computer executable instructions to implement the above Figures 1 to 2 Either of the methods shown. Figure 10 In the illustrated embodiment, the electronic device further includes a bus 102 and a communication interface 103. The processor 101, communication interface 103, and memory 100 are connected via bus 102. Memory 100 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk drive. Communication between the system network element and at least one other network element is achieved via at least one communication interface 103 (which may be wired or wireless). This communication interface may utilize the Internet, a wide area network, a local area network, a metropolitan area network, or the like. Bus 102 can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, or an AMBA (Advanced Microcontroller Bus Architecture) bus. AMBA defines three types of buses, including an APB (Advanced Peripheral Bus) bus, an AHB (Advanced High-performance Bus) bus, and an AXI (Advanced eXtensible Interface) bus. Bus 102 can be divided into an address bus, a data bus, a control bus, and the like. For ease of representation, Figure 10 Only one bidirectional arrow is used in the diagram, but this does not mean that there is only one bus or one type of bus.

[0149] The processor 101 may be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method can be completed by an integrated logic circuit of hardware in the processor 101 or by instructions in the form of software. The above-mentioned processor 101 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in conjunction with the embodiments of the present application can be directly embodied as being executed by a hardware decoding processor, or can be executed by a combination of hardware and software modules in the decoding processor. The software module can be located in a storage medium mature in the art, such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor 101 reads the information in the memory and combines its hardware to complete the above Figures 1 to 2 Any of the methods shown.

[0150] The embodiments of the present invention provide a computer program product for a method and apparatus for assessing equipment health based on a knowledge graph and a large model, including a computer-readable storage medium storing program code. The program code includes instructions that can be used to execute the methods described in the aforementioned method embodiments. For detailed implementation, please refer to the method embodiments and will not be described in detail here. Those skilled in the art will clearly understand that, for ease and brevity of description, the specific operating process of the system described above can refer to the corresponding process in the aforementioned method embodiments and will not be described in detail here. If the functions described are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes instructions for causing a computer device (such as a personal computer, server, or network device) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0151] In the description of the present invention, it should be noted that the terms "first", "second" and "third" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance. Finally, it should be noted that the above embodiments are only specific implementation methods of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that any person skilled in the art who is familiar with this technical field can still modify the technical solutions described in the aforementioned embodiments within the technical scope disclosed by the present invention, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A device health status assessment method based on knowledge graph and large model, characterized by: The method comprises: Obtain vibration monitoring signals of industrial equipment during operation; Decomposing the vibration monitoring signal into a plurality of frequency sub-bands, and extracting the energy distribution corresponding to each of the frequency sub-bands; performing normalization processing on the vibration monitoring signal according to the energy distribution to generate a target feature vector corresponding to the vibration monitoring signal; Inputting the target feature vector into a preset neural network, performing feature extraction on the target feature vector through the neural network, and determining a feature to be measured corresponding to the target feature vector; wherein the feature to be measured is determined based on a dynamic response of the neural network to a preset frequency band; Based on the features to be measured and the equipment knowledge graph corresponding to the industrial equipment, classify and evaluate the equipment health status of the industrial equipment, and determine the health status evaluation result indicated by the industrial equipment; The step of decomposing the vibration monitoring signal into a plurality of frequency sub-bands and extracting the energy distribution corresponding to each of the frequency sub-bands comprises: Performing multi-scale decomposition on the vibration monitoring signal through a wavelet transform function to obtain local features of frequency bands corresponding to different time scales; wherein each time scale corresponds to a preset frequency sub-band; Determining a modulus response of the frequency band local characteristic in a current frequency sub-band; the modulus response is used to characterize an amplitude characteristic of a signal energy density indicated by the frequency band local characteristic; determining an energy distribution corresponding to a current frequency sub-band based on the modulus response; The step of inputting the target feature vector into a preset neural network, performing feature extraction on the target feature vector through the neural network, and determining the feature to be measured corresponding to the target feature vector includes: Inputting the target feature vector into a preset neural network, performing high-dimensional nonlinear mapping on the target feature vector according to the similarity of the target feature vector in the neighborhood, and obtaining a high-dimensional feature representation of the target feature vector; performing nonlinear amplitude amplification on the high-dimensional feature representation to determine a frequency band response change corresponding to the high-dimensional feature representation; Based on the frequency band response change, adjusting the frequency band amplitude of the high-dimensional feature representation; Based on the dynamic response of the frequency band amplitude adjustment indication, a nonlinear transformation is performed on the high-dimensional feature representation to obtain a feature to be measured corresponding to the target feature vector.

2. The method according to claim 1, characterized in that The step of normalizing the energy distribution to generate a target feature vector corresponding to the vibration monitoring signal includes: Calculating an energy normalization value corresponding to the energy distribution based on the energy distribution of each frequency sub-band; The energy distribution is weightedly combined based on a preset weighting coefficient, and the result of the weighted combination is normalized based on the energy normalization value to generate a target feature vector corresponding to the vibration monitoring signal.

3. The method according to claim 1, characterized in that The step of performing high-dimensional nonlinear mapping on the target feature vector based on the similarity of the target feature vector in the neighborhood to obtain a high-dimensional feature representation of the target feature vector includes: Identifying multiple local center vectors corresponding to the target feature vector through a preset clustering algorithm, and determining multiple fault modes indicated by the target feature vector; Determining a neighborhood perception kernel width corresponding to the target feature vector based on the local center vector; Using a preset exponential kernel function, based on the neighborhood perception kernel width and the local center vector, the similarity of the feature vectors corresponding to each fault mode indicated by the target feature vector is measured; Based on the feature vector similarity, a high-dimensional nonlinear mapping process is performed on the target feature vector to determine a high-dimensional feature representation corresponding to the target feature vector.

4. The method according to claim 1, wherein The equipment knowledge graph includes the fault phenomenon, fault location, fault root cause and fault solution of the industrial equipment; The steps of classifying and evaluating the health status of the industrial equipment based on the features to be measured and the equipment knowledge graph corresponding to the industrial equipment, and determining the health status evaluation result indicated by the industrial equipment include: Performing classification prediction on the feature to be measured by using a preset first prediction model, and outputting the probability of the device health status category indicated by the feature to be measured; Identifying the fault phenomenon characteristic parameters indicated by the feature to be measured by using a preset second prediction model, and matching the device fault phenomenon indicated by the fault phenomenon characteristic parameters from the device knowledge graph; Based on the equipment health status category probability and the equipment fault phenomenon, the fault location and fault root cause of the industrial equipment are located, and the health status assessment result indicated by the industrial equipment is determined.

5. The method according to claim 4, characterized in that The first prediction model is constructed based on a preset neural network; The step of performing classification prediction on the feature to be measured by using a preset first prediction model and outputting the probability of the device health status category indicated by the feature to be measured includes: Determining, based on the target eigenvector of the vibration monitoring signal, a statistical characteristic of a failure mode distribution corresponding to the vibration monitoring signal; Initializing the neural network weights of the first prediction model based on the statistical characteristics of the fault mode distribution; and performing differential constraints on the neural network weights based on the spatiotemporal correlation of the target feature vector; Based on the neural network weights of the first prediction model and a preset multi-objective loss function, the features to be measured are classified and predicted to determine the probability of the equipment health status category corresponding to the features to be measured.

6. The method according to claim 5, characterized in that The calculation method of the multi-objective loss function includes: Determining the classification accuracy loss of the neural network based on the training sample set corresponding to the feature to be tested and the category prediction probability corresponding to the training sample set; Determining a robustness loss corresponding to the neural network based on a preset perturbation sample; Based on the classification accuracy loss and the robustness loss, a multi-objective loss function corresponding to the neural network is calculated.

7. The method according to claim 5, characterized in that The method further comprises: Determining a current error direction of the neural network corresponding to a preset training sample set based on the multi-objective loss function; Calculating the first-order gradient information of the multi-objective loss function with respect to the current weight of the neural network, and the second-order gradient information with respect to the weight of each layer of the neural network; Based on the first-order gradient information, the second-order gradient information and a preset weight decay term, the initial weight of the neural network is updated to determine the final weight parameter of the neural network.

8. A device for evaluating equipment health status based on knowledge graph and large model, characterized in that: The device comprises: Signal acquisition module, used to obtain vibration monitoring signals of industrial equipment during operation; A signal decomposition module, configured to decompose the vibration monitoring signal into a plurality of frequency sub-bands and extract the energy distribution corresponding to each of the frequency sub-bands; a data processing module, configured to perform normalization processing on the vibration monitoring signal according to the energy distribution, and generate a target feature vector corresponding to the vibration monitoring signal; an execution module, configured to input the target feature vector into a preset neural network, perform feature extraction on the target feature vector through the neural network, and determine a feature to be measured corresponding to the target feature vector; wherein the feature to be measured is determined based on a dynamic response of the neural network to a preset frequency band; An output module is configured to classify and evaluate the health status of the industrial equipment based on the features to be measured and the equipment knowledge graph corresponding to the industrial equipment, and determine a health status evaluation result indicated by the industrial equipment; The signal decomposition module is further used to: Performing multi-scale decomposition of the vibration monitoring signal using a wavelet transform function to obtain frequency band local features corresponding to different time scales; wherein each time scale corresponds to a preset frequency sub-band; determining a modulus response of the frequency band local feature in a current frequency sub-band; the modulus response is used to characterize an amplitude characteristic of signal energy density indicated by the frequency band local feature; and determining an energy distribution corresponding to the current frequency sub-band based on the modulus response; The execution module is further configured to: The target feature vector is input into a preset neural network, and according to the similarity of the target feature vector in the neighborhood, the target feature vector is subjected to high-dimensional nonlinear mapping to obtain a high-dimensional feature representation of the target feature vector; the high-dimensional feature representation is subjected to nonlinear amplitude amplification to determine the frequency band response change corresponding to the high-dimensional feature representation; based on the frequency band response change, the high-dimensional feature representation is subjected to frequency band amplitude adjustment; based on the dynamic response indicated by the frequency band amplitude adjustment, the high-dimensional feature representation is subjected to nonlinear transformation to obtain the feature to be measured corresponding to the target feature vector.

Images