Big data center-based equipment health index analysis method and system

Abstract

The present application relates to the technical field of machine learning, solves the technical problem that the existing technology threshold setting often depends on historical extreme value, ignores its change trend and fluctuation characteristics over time, and is extremely insensitive to early and slowly changing faults, and particularly relates to a device health index analysis method and system based on a big data center, the steps of the method being: collecting original device time series data of a target device, obtaining normalized time series data through preprocessing according to the original device time series data, the present application can capture abnormal patterns earlier and more comprehensively through separating and quantifying the trend and fluctuation of the health index, threshold calculation is adaptive, and the problem that a fixed threshold value cannot adapt to individual differences of devices, working condition changes and natural performance degradation is solved, risks can be captured sensitively from any dimension, and transient interference is effectively filtered through a duration condition, so that the detection sensitivity is improved while the false alarm rate is greatly reduced.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a method and system for analyzing device health index based on big data centers. Background Technology

[0002] Machine learning, a branch of artificial intelligence, enables computer systems to improve performance by automatically learning patterns and rules from data. Big data centers provide machine learning with massive storage resources, efficient computing platforms, and integrated data pipelines, enabling machine learning models to more accurately predict and diagnose equipment conditions. Current technologies mainly achieve this through traditional linear regression, which extracts statistical features from historical equipment data, then uses these features to train a regression model, mapping the input features to category labels representing the health status of the equipment, thereby achieving the assessment and prediction of equipment health.

[0003] The threshold settings of existing technologies often rely on historical extreme values, making it difficult to adjust with changes in equipment status. This can easily lead to false alarms or slow degradation that is difficult to detect. They ignore the trend and fluctuation characteristics of the equipment over time and are extremely insensitive to early and slow-changing faults. In actual testing, when the equipment undergoes periodic load adjustments, its health index will reach a steady state at a new level. Traditional static thresholds will continuously misjudge this new steady state as abnormal, leading to abnormal signals from the equipment and seriously affecting the normal operation of the equipment after adjustment. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for analyzing the health index of equipment based on big data centers. This solves the problems of existing technologies, which often rely on historical extreme values ​​for threshold settings, leading to false alarms or slow degradation that is difficult to detect. Furthermore, it ignores the changing trends and fluctuation characteristics over time, making it extremely insensitive to early and slowly changing faults.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a device health index analysis method based on a big data center, the steps of which are: collecting the original device time series data of the target device, and obtaining normalized time series data by preprocessing the original device time series data; Based on normalized time series data, feature decomposition is performed to obtain multidimensional feature vectors. Multimodal feature extraction and fusion are then performed on the multidimensional feature vectors to generate fused feature vectors. Based on the fused feature vector, health status is predicted, generating a first predicted health status index and a second predicted health status index. The first predicted health status index and the second predicted health status index are then converted to health status to obtain the device health status index value. The equipment health index value is pre-detected and processed to obtain the health index trend vector and health index fluctuation characteristics. Based on the health index trend vector and health index fluctuation characteristics, an early warning signal for equipment abnormality is generated through early warning detection processing. A maintenance plan is generated based on the equipment abnormality warning signal, resulting in an equipment maintenance plan.

[0006] Preferably, the original equipment time series data is preprocessed, including: data cleaning of the original equipment time series data to obtain cleaned time series data; Perform data alignment on the cleaning time-series data to generate time-series aligned data; Normalized time-series data is obtained by normalizing the time-aligned data.

[0007] Preferably, feature decomposition based on normalized time-series data includes: performing time-domain feature processing on normalized time-series data to obtain time-domain feature vectors; Frequency domain features are extracted based on normalized time series data to generate frequency domain feature vectors; Wavelet packet decomposition is performed on the normalized time series data to obtain wavelet energy feature vectors.

[0008] Preferably, multimodal feature extraction and fusion of multidimensional feature vectors includes: concatenating features based on multidimensional feature vectors to generate concatenated feature vectors; The encoder performs dimensionality reduction based on the concatenated feature vectors to generate encoded feature vectors. The encoded feature vector is subjected to feature fusion processing to obtain the fused feature vector.

[0009] Preferably, health prediction based on the fused feature vector includes: dividing the fused feature vector into time series segments to generate an input vector, the input vector including a training feature vector and a test feature vector; The first predicted health index is obtained by performing support vector regression processing based on the trained feature vectors. A second predicted health index is generated by processing the training feature vectors through Gaussian regression.

[0010] Preferably, the health conversion of the first predicted health index and the second predicted health index includes: fusing the prediction results based on the first predicted health index and the second predicted health index to obtain a fused predicted health index. Based on the fusion-predicted health index, a standardized health index is generated through standardization processing; in this step, two key boundary values ​​are determined based on historical or reference data; The standardized health index is processed to obtain the equipment health index value.

[0011] Preferably, the equipment health index value is subjected to pre-detection processing, including: performing time series data processing based on the equipment health index value to obtain a health index time series; A weighted moving average is applied to the health index time series to generate a health index trend vector. Standard deviation processing is applied to the time series of the health index to obtain the fluctuation characteristics of the health index.

[0012] Preferably, the early warning detection process is based on the health index trend vector and health index fluctuation characteristics, including: adaptive threshold processing based on the health index trend vector and health index fluctuation characteristics to generate multi-dimensional dynamic abnormal thresholds; Threshold labeling is performed on the health index trend vector, health index fluctuation characteristics, and multi-dimensional dynamic anomaly thresholds to obtain multi-dimensional anomaly labels; Early warning signals are obtained by processing multi-dimensional anomaly markers to obtain equipment anomaly early warning signals.

[0013] Preferably, the maintenance plan is generated based on the equipment anomaly warning signal, including: matching the fault type based on the equipment anomaly warning signal to generate a fault type; Knowledge graph retrieval based on fault type generates a set of candidate maintenance strategies. The candidate maintenance strategy set is optimized to generate an equipment maintenance plan.

[0014] The technical solution also provides a device health index analysis system based on a big data center. The system includes: a preprocessing module, which is used to collect the raw device time series data of the target device and obtain normalized time series data through preprocessing based on the raw device time series data; The feature fusion module is used to perform feature decomposition based on normalized time series data to obtain multi-dimensional feature vectors, and to perform multi-modal feature extraction and fusion on the multi-dimensional feature vectors to generate fused feature vectors. The health module is used to predict health based on the fused feature vector, generate a first predicted health index and a second predicted health index, and perform health conversion on the first predicted health index and the second predicted health index to obtain the device health index value. The anomaly warning module is used to perform pre-detection processing on the equipment health index value to obtain the health index trend vector and health index fluctuation characteristics. Based on the health index trend vector and health index fluctuation characteristics, the module generates an equipment anomaly warning signal through warning detection processing. The maintenance plan module is used to generate maintenance plans based on equipment anomaly warning signals, thus obtaining the equipment maintenance plan.

[0015] By employing the above technical solution, the present invention provides a method and system for analyzing device health index based on big data centers, which has at least the following beneficial effects: 1. This invention employs multi-dimensional parallel analysis using time domain, frequency domain, and wavelet packet decomposition. It can simultaneously capture the macroscopic statistical laws, periodic frequency components, and transient time-frequency local features of equipment operation, achieving lossless and complementary mining of signal information. It avoids information omissions in single-domain analysis and can automatically learn from high-dimensional spliced ​​features and compress the most core and discriminative low-dimensional representations, effectively eliminating redundancy and noise, and improving the information density and quality of features. Through standardized steps, the scale of the dimensionality-reduced features is uniformly adjusted, which greatly enhances the training stability, convergence speed, and generalization ability of the model, and significantly improves the sensitivity and accuracy of early fault identification.

[0016] 2. This invention adopts a combination strategy of support vector regression and Gaussian process regression, which is robust under small sample conditions and can provide a measure of prediction uncertainty. By using weighted average fusion, it not only improves the accuracy of point prediction, but also smooths out the random error of a single model. It converts the original prediction value into a standardized index, which solves the problem that the results of different devices and different models cannot be compared horizontally and monitored uniformly. It can also ensure the unbiasedness of model evaluation and engineering practicality.

[0017] 3. This invention constructs two complementary dimensions describing the direction and stability of equipment status changes by separating and quantifying the health index into trends and fluctuations. Compared with the traditional method of only monitoring the absolute value of the index, it can capture abnormal patterns earlier and more comprehensively. The threshold calculation is adaptive, solving the problem that fixed thresholds cannot adapt to individual differences in equipment, changes in operating conditions, and natural performance degradation. The early warning generation integrates logical combination and time persistence rules, which can not only keenly capture risks from any dimension, but also effectively filter instantaneous interference through duration conditions, thereby improving detection sensitivity while significantly reducing the false alarm rate.

[0018] 4. This invention accurately matches warning signals with fault types through a rule base, avoiding the subjectivity and inconsistency of manual diagnosis, and significantly improving the accuracy and speed of fault location. By using a knowledge graph for maintenance strategy retrieval, it can systematically associate and call historical best practices, solving the problems of isolated maintenance knowledge, easy loss, or low retrieval efficiency in traditional methods. Through a multi-dimensional strategy ranking model, it outputs optimized solutions, comprehensively considering multiple constraints such as cost, time, and resources, making the final decision recommendations not only technically feasible but also in line with the economic efficiency of operation and maintenance and the actual situation on site, realizing scientific and refined recommendations for maintenance. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of the device health index analysis method based on big data centers according to the present invention; Figure 2 This is a structural block diagram of the device health index analysis system based on a big data center according to the present invention. Detailed Implementation

[0020] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This will allow for a full understanding of how the present application uses technical means to solve technical problems and achieve technical effects, and to facilitate its implementation.

[0021] Example 1: Because existing technology threshold settings often rely on historical extreme values, they are prone to false alarms or slow degradation that is difficult to detect. They ignore the changing trends and fluctuations over time, making them extremely insensitive to early, slowly changing faults. Please refer to [reference needed]. Figure 1 This embodiment provides a device health index analysis method based on big data centers, which can separate and quantify the trend and fluctuation of the health index, capture abnormal patterns earlier and more comprehensively, and has adaptive threshold calculation, solving the problem that fixed thresholds cannot adapt to individual differences in equipment, changes in operating conditions, and natural performance degradation. The method includes the following steps: S1. Collect the raw device time series data of the target device, and obtain normalized time series data through preprocessing based on the raw device time series data. Existing technologies often rely on experience to set global thresholds, which can easily lead to a large number of misjudgments or fail to effectively filter complex noise. They do not consider the actual distribution of data within the time window, which may introduce false values ​​or distort instantaneous features. When the device operating mode changes or the new data exceeds the original range, the normalization results may become unstable or even fail, affecting the generalization ability of the model. To solve the above problems, the specific implementation steps are as follows: S11. Perform data cleaning on the original equipment time-series data to obtain cleaned time-series data. The original equipment time-series data includes time-series parameters continuously collected during equipment operation, such as temperature, pressure, vibration amplitude, current, voltage, and rotational speed. In this step, a fixed-size sliding window is first defined, for example, a sliding window of size B with a sliding step size of C. The window covers the current data point and multiple neighboring data points. For each window, the average value of all data points within the window is calculated by summing the values ​​of all data points and then dividing by the total number of data points. Next, the standard deviation of the data points within the window is calculated by first squaring the difference between each data point and the average value, then summing all the squared values, dividing by the number of data points, and finally taking the square root of the result. Then, for the current data point... The absolute value of the difference between the current data point and the window mean is calculated and compared with three times the standard deviation. If the absolute value is less than or equal to three times the standard deviation, the original value of the current data point is retained; otherwise, the current data point is replaced with the window mean. The entire time series data is traversed through a sliding window, and each data point is processed sequentially. The cleaned time series data eliminates abnormal fluctuations caused by measurement errors or transient faults, enhancing the stability and reliability of the data. In the equipment health index analysis method based on big data centers, the statistical characteristics of the data, such as trends, periodicity, and stability, are truly reflected by the outlier removal through sliding window, avoiding the distortion of the mean, variance, and correlation analysis by outliers. This ensures that the health index calculation is based on high-quality data, thereby improving the accuracy of equipment condition monitoring.

[0022] S12. Align the cleaning time series data to generate time-aligned data. In this step, the target alignment time series is first defined, that is, a fixed and uniform time interval grid is set. For each target time point on the grid, a time tolerance range centered on that point is set. Then, in the input cleaning time series data, all original data points whose timestamps fall within this tolerance range are found. Next, the sum of the values ​​of these captured data points is calculated and divided by the total number of valid data points within the tolerance range to obtain the new data value corresponding to the target time point. This calculation process is essentially the arithmetic mean of the values ​​of all data points within the tolerance window, which represents the typical operating state of the equipment within the target time period. By traversing all preset uniform time points and repeating this operation, a complete time alignment sequence, i.e., time-aligned data, is finally generated. Its characteristic is that all data points strictly correspond to a set of predefined equally spaced timestamps, and each data value is the average value of the original data within the corresponding time window.

[0023] S13. Normalize the time-series aligned data to obtain normalized time-series data. In this step, for each device operation feature represented by the aligned time-series data to be processed, firstly, traverse all data points of the feature in the entire time series to find the global maximum and global minimum values. Then, for each original data point in the sequence, subtract the found global minimum value from the value of the point to obtain a difference. At the same time, calculate another difference between the global maximum and global minimum values. Finally, divide the first difference by the second difference, and the quotient is the new normalized value of the data point. This process is applied independently to the time-series data of each feature dimension. After processing, the normalized time-series data output has all values ​​in each feature dimension strictly scaled to the range of zero to one, and the relative distribution relationship of the values ​​is preserved.

[0024] This invention employs a sliding window-based local statistical method to remove outliers, dynamically adapting to non-stationary changes in equipment operating status. It effectively identifies and processes local transient noise and real fault symptoms, avoiding information loss or erroneous retention. Data alignment utilizes resampling and averaging within a time tolerance window, better integrating information from asynchronously acquired data near the target time and smoothing errors caused by minute time fluctuations. Normalization unifies various feature scales, enabling indicators with different physical meanings, such as temperature, vibration, and current, to be summarized and unified in data distribution, thereby progressively improving data quality and consistency.

[0025] S2. Based on normalized time-series data, feature decomposition is performed to obtain multi-dimensional feature vectors. Multi-modal feature extraction and fusion are then performed on the multi-dimensional feature vectors to generate fused feature vectors. Existing technologies have single or fragmented feature extraction methods and crude feature fusion methods, lacking effective dimensionality reduction and redundancy removal mechanisms. This results in high feature dimensionality and large redundancy. Furthermore, in actual operation, under strong non-steady-state conditions such as equipment speed change and load change, the frequency domain features based on steady-state assumptions, such as the amplitude of fixed frequency bands, will be severely distorted or fail. Manually set features are difficult to dynamically adapt to such changes, leading to inaccurate equipment health status analysis and even serious consequences. To solve the above problems, the specific implementation steps are as follows: S21. Perform time-domain feature processing on the normalized time-series data to obtain time-domain feature vectors. The multi-dimensional feature vectors include time-domain feature vectors, frequency-domain feature vectors, and wavelet energy feature vectors. In this step, the arithmetic mean of all data points in the sequence is first calculated, i.e., the sum of all data point values ​​divided by the total number of data points. This reflects the overall level of the operating state during that time period. Next, the standard deviation is calculated by squaring the difference between each data point and the mean, summing all the squared values, dividing by the total number of data points, and then taking the square root. Then, skewness is calculated by summing the cubes of the differences between each data point and the mean, dividing by the total number of data points, and then dividing by the cube of the standard deviation. This describes the degree of asymmetry in the data distribution shape relative to the mean. Finally, kurtosis is calculated. The process involves summing the fourth power of the difference between each data point and the mean, dividing by the total number of data points, and then dividing by the fourth power of the standard deviation. A baseline constant is usually subtracted to characterize the sharpness or flatness of the data distribution. In addition, the root mean square value is calculated by first squaring the value of each data point, then calculating the average of these squared values, taking the square root of the average, and finally extracting the maximum value in the time series segment as the peak feature. Combining these six statistics in a predetermined order constitutes a time-domain feature vector representing the behavior of the feature variable in this period. The predetermined order is generally the mean, standard deviation, skewness, kurtosis, root mean square value, and peak value. This order usually follows a logical arrangement from basic central tendency and dispersion measures to distribution shape characteristics, and then to energy and extreme value characteristics.

[0026] S22. Frequency domain feature extraction based on normalized time series data to generate frequency domain feature vectors. In this step, for a given length of normalized time series data, the efficient algorithm Fast Fourier Transform is used for calculation. The essence of this algorithm is to decompose the original time series and represent it as a superposition of a series of standard sine and cosine waveforms of different frequencies. The calculation process involves a large number of complex number operations, including multiplying each data point in the sequence by a complex rotation factor of different frequencies, and then recursively summing and recombining all the product results. Through this series of specific multiplication and addition operations, a complex result is finally calculated for each preset frequency unit. The square of the modulus of this result represents the energy or power of the original signal corresponding to that frequency component. Arranging the energy values ​​corresponding to all frequency units in order of frequency from low to high constitutes a spectrum describing the frequency energy distribution of the signal. The key frequency band energy or peak frequency and other parameters with engineering significance in this spectrum are extracted and arranged in order to obtain the required frequency domain feature vector.

[0027] S23. Perform wavelet packet decomposition on the normalized time series data to obtain the wavelet energy feature vector. In this step, the input time series data is first fed into the wavelet packet decomposition algorithm. This algorithm uses a set of mathematical tools called high-pass and low-pass filters to decompose the original signal layer by layer. The first layer of decomposition divides the signal into two parts, obtaining two sub-band signals representing the high-frequency and low-frequency components of the signal, respectively. Subsequently, unlike ordinary wavelet decomposition, the algorithm does not only decompose the low-frequency component, but performs the same decomposition on both the high-frequency and low-frequency sub-band signals generated in the current layer in the next layer. This process is repeated recursively. After several layers of decomposition, the original... The signal is divided into a large number of sub-band signal sequences covering different frequency ranges and with time positioning capabilities. For each final sub-band obtained by decomposition, its corresponding wavelet packet coefficient sequence is calculated. Then, the values ​​of all wavelet packet coefficients in the sub-band are squared one by one, and all the squared values ​​are summed. The sum represents the energy of the original signal within the specific time and frequency window. This square summation operation is repeated for all final sub-bands to obtain a set of energy values ​​that correspond one-to-one with each sub-band. Arranging these energy values ​​according to their corresponding frequency sub-band order constitutes the wavelet energy feature vector describing the time-frequency energy distribution of the signal.

[0028] S24. Based on multi-dimensional feature vectors, feature concatenation is performed to generate concatenated feature vectors. In this step, the time-domain feature vector, frequency-domain feature vector, and wavelet energy feature vector are each regarded as an ordered numerical sequence. Then, according to the fixed order of time domain, frequency domain, and wavelet energy, the three sequences are connected end to end to form a new, longer single sequence. This connection process does not involve multiplication, division, addition, or subtraction between values, but only the merging and arrangement of physical order. All values ​​of the first vector are placed at the beginning of the new vector in the original order. Then, all values ​​of the second vector are sequentially followed. Finally, all values ​​of the third vector are continued at the end, thus forming a new vector with a higher dimension that contains all the original feature information. In the device health index analysis method based on big data centers, time-domain feature vectors describe the macroscopic statistical laws and overall fluctuation levels of device operation, frequency-domain feature vectors reveal the frequency structure information corresponding to internal rotating parts or periodic actions, and wavelet energy feature vectors characterize the local details and transient energy distribution of signals in the joint time-frequency domain. If any one dimension of features is used alone, the key state fingerprints contained in other dimensions may be lost. By concatenating vectors, these heterogeneous but complementary features are integrated into a structured data object, providing the richest and most complete input information foundation for subsequent health assessment models. This enables data-driven analysis models to simultaneously consider the statistical characteristics, frequency components, and local time-frequency patterns of device status, thereby making more comprehensive and robust integrated judgments and decisions, significantly improving the identification accuracy of complex degradation modes and compound faults and the reliability of health prediction.

[0029] S25. Perform encoder dimensionality reduction processing based on the concatenated feature vector to generate the encoded feature vector. In this step, the input concatenated feature vector is first operated on with an encoded weight matrix obtained through training. This operation is essentially multiplying each value in the input vector by all values ​​in the corresponding row of the weight matrix in turn, and then adding all the product results in each row to obtain a new set of intermediate values. Subsequently, a preset sequence of values ​​called the encoded bias vector is added to the corresponding positions of this set of intermediate values. Then, each result obtained by the above addition is fed into a specific nonlinear activation function for transformation. This function usually performs mathematical rules on the input value, such as mapping negative numbers to zero, retaining the original value of positive numbers, or compressing them to the interval between zero and one. Through this series of pre-set matrix multiplication, vector addition, and nonlinear function transformation combined calculations, the original high-dimensional features are mapped to a new space with significantly reduced dimensions, forming the encoded feature vector.

[0030] S26. Perform feature fusion processing on the encoded feature vector to obtain the fused feature vector. In this step, firstly, calculate the arithmetic mean of all values ​​of this dimension across all samples. Then, for each specific value under this dimension, subtract the calculated mean from the value to obtain a centered difference. At the same time, calculate the standard deviation of all values ​​of this dimension, which is the square root of the average of the squares of the differences between each value and the mean. Finally, divide the previously obtained centered difference by the standard deviation of this dimension so that the value distribution of all samples of the new data after the transformation of each feature dimension satisfies the normal state of zero mean and one standard deviation. After all dimensions have undergone this standardization process, recombine them in the original order to form the final fused feature vector.

[0031] This invention employs multi-dimensional parallel analysis using time domain, frequency domain, and wavelet packet decomposition. It can simultaneously capture the macroscopic statistical laws, periodic frequency components, and transient time-frequency local features of equipment operation, achieving lossless and complementary mining of signal information. This avoids information omissions in single-domain analysis and can automatically learn from high-dimensional spliced ​​features and compress the most core and discriminative low-dimensional representations, effectively eliminating redundancy and noise, and improving the information density and quality of features. By standardizing the steps to uniformly adjust the scale of the dimensionality-reduced features, the model's training stability, convergence speed, and generalization ability are greatly enhanced, significantly improving the sensitivity and accuracy of early fault identification.

[0032] S3. Based on the fused feature vector, predict the health status and generate a first predicted health status index and a second predicted health status index. Perform health status conversion on the first predicted health status index and the second predicted health status index to obtain the device health status index value. Existing technologies generally rely on a single prediction model, which is prone to failure when the data distribution changes and cannot provide prediction confidence, resulting in high risk of operation and maintenance decisions. They also lack a standardized output system, and the original values ​​output by the model have inconsistent dimensions and uncertain ranges, making them unsuitable for direct use in automated early warning and cross-system integration. For example, in real-world mixed management big data systems with multiple devices and models, existing technologies build separate models for each device and generate different results, making it impossible for the system to use a unified standard to globally monitor the health level of all devices. To solve the above problems, the specific steps are as follows: S31. Perform time series partitioning on the fused feature vectors to generate input vectors, which include training feature vectors and test feature vectors. The input vectors can be divided into training set, validation set, and test set. In this step, all fused feature vectors are first strictly sorted according to the chronological order of their corresponding timestamps to ensure the temporal continuity of the sequence. Then, a predetermined partitioning ratio is set, such as dividing the first 80% and the last 20% of the sequence. Based on the index position of the data points on the time axis, the selected and grouped feature vectors are selected. The continuous feature vectors located at the beginning of the time series, accounting for a predetermined proportion of the total data volume, are extracted as a whole and classified into the training feature vector set, while the continuous feature vectors located at the end of the time series, accounting for the remaining proportion, are extracted as a whole and classified into the test feature vector set. This process ensures that the training set and the test set are completely separated in time and do not overlap.

[0033] S32. Based on the training feature vectors, support vector regression is performed to obtain the first predicted health index. In this step, the input training feature vectors are mapped to a potentially higher-dimensional feature space to handle complex nonlinear relationships in the data. This mapping process usually involves specific operations on the distance or similarity between vectors. For example, the sum of squares of the differences between all corresponding dimensions of two vectors is calculated, and then this sum is exponentially multiplied by a negative coefficient to obtain a new measurement value. The core of model training is to find an optimal linear function such that the deviation between the predicted value of this function for most training samples in the mapped high-dimensional space and its true health index does not exceed a certain threshold. A preset tolerance range is set while ensuring the function's complexity is minimized. This search process is achieved by solving a convex optimization problem, which ultimately determines two sets of key Lagrange multiplier parameters and a bias constant. For a new input feature vector, the method for calculating its predicted health index is as follows: iterate through all support vector samples, i.e., training samples whose corresponding Lagrange multipliers are not zero, subtract the two sets of Lagrange multiplier values ​​corresponding to each support vector, multiply the difference by the similarity measure value calculated by the kernel function between the support vector feature and the new input vector, sum all these product results, and finally add the determined bias constant to obtain the first predicted health index. In the equipment health index analysis method based on big data centers, the output predicted health index is a continuous value, which can delicately depict the gradual deterioration process of equipment from good to faulty. It is superior to simple binary classification judgment of normal or abnormal. The structural risk minimization principle of this model enables it to still show good generalization ability and robustness on small samples or data with complex nonlinear patterns. Through training, the model can automatically capture the key patterns most related to equipment performance degradation from the fused feature vectors and nonlinearly map these high-dimensional features into an intuitive and monitorable health index. This index constitutes a direct and quantitative basis for big data centers to conduct real-time monitoring of equipment status, trend prediction and preventive maintenance decisions.

[0034] S33. Based on the training feature vectors, Gaussian regression is used to generate a second predicted health index. In this step, a mean function is first defined as the basic trend, and a kernel function is selected to characterize the covariance relationship between the health indices corresponding to different feature vectors. This kernel function typically involves summing the squares of the differences between the values ​​of each dimension of two vectors, multiplying by a negative coefficient, and then taking the exponent. During the model training phase, the internal parameters of the kernel function are optimized by maximizing the marginal likelihood function. This process involves inverting the matrix formed by the covariance values ​​calculated pairwise between all training samples using the kernel function and calculating the logarithmic determinant. In the prediction phase, for a new input feature vector, the covariance between this vector and all training feature vectors is first calculated using the kernel function. The covariance values ​​obtained from the function form a covariance vector. This covariance vector is then multiplied by the inverse of the covariance matrix calculated during the training phase, and then multiplied by the vector centered after training the health index, resulting in a weighted sum. Finally, the contribution of the mean function is added to obtain the second predicted health index. Simultaneously, the weighted product of this new vector and its kernel function value with the training data covariance vector is subtracted to obtain the predicted variance. The predicted variance provides an estimate of the uncertainty of the predicted value. The second predicted health index and the predicted variance make the final health assessment not just a point estimate, but a probabilistic judgment containing confidence information, which greatly improves the reliability of the state assessment, the refinement of decision support, and the economy and safety of operation and maintenance actions.

[0035] S34. The prediction results are fused based on the first and second predicted health indices to obtain the fused predicted health index. In this step, a preset weight coefficient is first assigned to the first and second predicted health indices. Both weight coefficients are decimals between zero and one, and their sum is a complete one. For example, the weight coefficients are a and b, and a plus b equals 1. The weight coefficients can be obtained through the analytic hierarchy process. Then, the specific value of the first predicted health index is multiplied by its corresponding weight coefficient to obtain a weighted result. At the same time, the specific value of the second predicted health index is also multiplied by its corresponding weight coefficient to obtain another weighted result. Finally, the two weighted results are added together, and the sum is the final fused predicted health index.

[0036] S35. Based on the fusion predicted health index, a standardized health index is generated through standardization processing. In this step, two key boundary values ​​are determined based on historical or reference data: the possible theoretical or observed minimum value of the health index, and the theoretical or observed maximum value. For each fusion predicted health index to be processed, the determined minimum value is first subtracted from the index value to obtain a difference indicating its deviation from the minimum benchmark. At the same time, the range difference between the maximum and minimum values ​​is calculated. Then, the first difference is divided by the range difference to obtain a proportional quotient value between zero and one. This proportional quotient value reflects the relative position of the current health status in the theoretical complete degradation spectrum. Finally, this proportional quotient value is multiplied by one hundred to linearly amplify the relative proportion and map it onto a standard scale of zero to one hundred. The result is the standardized health index.

[0037] S36. Process the standardized health index to obtain the device health index value. The standardized health index value input in this step is automatically obtained along with the current timestamp information and the unique identification code of the device corresponding to the index. Then, this information, namely the health index value, timestamp, and device identifier, is combined and packaged according to a predefined structured format, such as a specific data field order, to form a complete data record. This process does not involve multiplication, division, addition, or subtraction of the health index value itself, but rather the association and assembly of information. Finally, through the set data interface, this data record containing the complete context is output to the target system, such as a real-time monitoring database, historical data warehouse, or visualization early warning platform, to complete its delivery and persistent storage as the final analysis result.

[0038] This invention employs a combined strategy of support vector regression and Gaussian process regression, which is robust with small samples and provides a measure of prediction uncertainty. By using weighted average fusion, it not only improves the accuracy of point prediction but also smooths out the random errors of a single model. It converts the original predicted values ​​into standardized indices, solving the problem of the inability to compare and uniformly monitor results from different devices and models. It also ensures the unbiasedness of model evaluation and its practicality in engineering.

[0039] S4. Perform pre-detection processing on the equipment health index value to obtain the health index trend vector and health index fluctuation characteristics. Based on the health index trend vector and health index fluctuation characteristics, generate equipment abnormality warning signals through early warning detection processing. Existing technology threshold settings heavily rely on historical extreme values, are highly subjective, and are difficult to adjust with changes in equipment status. This can easily lead to false alarms or undetected slow degradation. It completely ignores the trend and fluctuation characteristics of its changes over time, cannot distinguish between short-term interference and long-term degradation, and is extremely insensitive to early and slowly changing faults. In actual detection, when the equipment undergoes planned maintenance or periodic load adjustment, its health index will reach a steady state at a new level. Traditional static thresholds will continuously misjudge this new steady state as abnormal, leading to the appearance of equipment abnormal signals and seriously affecting the normal operation of the equipment after adjustment. To solve the above problems, the specific steps are as follows: S41. Perform time-series data processing based on the equipment health index values ​​to obtain a health index time series. In this step, firstly, based on the timestamp information carried by each equipment health index value, strictly sort all index values ​​belonging to the same equipment according to the chronological order of their occurrence. This process does not involve addition, subtraction, multiplication, or division of the index values ​​themselves, but rather pure logical organization and sequential linking. Then, these sorted values ​​are paired with their corresponding timestamps, recorded in a structured manner according to the fixed format of the index values ​​at time points, and connected sequentially to form a continuous and ordered numerical sequence in the time dimension. For cases where the timestamp intervals may be irregular, interpolation or marking can be performed according to a preset strategy to maintain the continuity of the sequence.

[0040] S42. Based on the health index time series, a weighted moving average is processed to generate a health index trend vector. In this step, a sliding window of fixed time length is first set. This window covers the time point where the trend value to be calculated is currently and several consecutive historical time points before it. For each historical time point in the window, a weight coefficient is assigned according to its distance from the current time. The closer the time point is, the larger its weight coefficient is. During the calculation, the health index value corresponding to each historical time point in the window is multiplied by its assigned weight coefficient to obtain a series of weighted index values. Then, all weighted index values ​​are summed to obtain a weighted sum. At the same time, all weight coefficients used are summed separately to obtain the total weight. Finally, the weighted sum is divided by the total weight, and the quotient is the health index trend value at the current time point. By moving this sliding window forward in sequence and repeating the above operations of multiplying by weight, summing and dividing for each time point, a corresponding trend value can be calculated for each valid time point in the original sequence. These trend values ​​are arranged in order to form the output trend vector. In the device health index analysis method based on big data centers, the output health index trend vector effectively isolates short-term interference and clearly reveals whether the device performance is stable, slowly deteriorating, or accelerating. A continuously downward trend vector indicates potential failure risks. Even if the current absolute index value is still within the safe range, if the trend changes from declining to stable or rising, it may indicate that the previous maintenance measures have had a positive effect. It can realize the automatic identification of the degradation mode of massive devices, early warning, and intelligent ranking of maintenance priorities, thereby improving health monitoring from static point status assessment to dynamic linear trend prediction, significantly enhancing the foresight of operation and maintenance and the scientific nature of resource allocation.

[0041] S43. Standard deviation processing is performed on the health index time series to obtain the health index volatility characteristics. In this step, a sliding window of fixed time length is first set. This window covers the time point where the volatility value to be calculated is currently located and several consecutive historical time points before it. For all health index values ​​within the window, their arithmetic mean is calculated first, that is, the sum of all values ​​is divided by the number of data points. Then, the difference between the value of each data point in the window and the average value is calculated. Then, each difference is squared to eliminate the influence of positive and negative directions and amplify the degree of deviation. Then, all the squared differences are summed, and this sum of squares is divided by the number of data points in the window to obtain the mean squared deviation. Finally, the square root operation is performed on this mean squared deviation. The result is the health index volatility value at the current time point. By moving this sliding window forward in sequence and repeating the above operations of calculating the average, difference, square, sum, average and square root for each time point, a corresponding volatility value can be calculated for each valid time point in the original sequence. These volatility values ​​are arranged in order to form the output volatility characteristic sequence.

[0042] S44. Adaptive thresholding is performed based on the health index trend vector and health index fluctuation characteristics to generate a multi-dimensional dynamic anomaly threshold. This step is achieved through two parallel calculation paths. For the trend vector, the arithmetic mean of all values ​​in the sequence is calculated first, i.e., the sum of all trend values ​​is divided by the total number of data points. At the same time, its standard deviation is calculated, i.e., the sum of the squares of the differences between each trend value and the average value is divided by the total number of data points and then the square root is taken. Then, the calculated average value is subtracted from the product of a preset adjustment coefficient and the standard deviation. The result is the trend anomaly threshold, which defines the lower limit of the normal range of trend values. For the fluctuation characteristic sequence, the average value and standard deviation of all volatility values ​​are calculated first. Then, the average value of the volatility is added to the product of another preset adjustment coefficient and the standard deviation. The result is the volatility anomaly threshold, which defines the upper limit of the normal range of volatility. In the device health index analysis method based on big data centers, the output multi-dimensional dynamic anomaly threshold is not a fixed value, but is adaptively generated based on the recent actual operating status of the device and represented by statistical measures of trends and fluctuations. When the health trend value of the device is lower than the calculated trend anomaly threshold, it indicates that its performance has shown a statistically significant deterioration trend; when the volatility value is higher than the calculated volatility anomaly threshold, it indicates that its operating status has become abnormally unstable. By comparing the current index with these two dynamic thresholds in real time, the big data center can achieve earlier, more accurate, and less false alarm composite anomaly detection and early warning from two independent and complementary dimensions: level decline and stability loss.

[0043] S45. Threshold marking processing is performed on the health index trend vector, health index volatility characteristics, and multi-dimensional dynamic anomaly thresholds to obtain multi-dimensional anomaly labels. In this step, for each identical time point, the system synchronously reads four data points: the health index trend value, the health index volatility value, the trend anomaly threshold, and the volatility anomaly threshold. First, the current trend value is compared with the trend anomaly threshold to determine if the former is less than the latter. Simultaneously, the current volatility value is compared with the volatility anomaly threshold to determine if the former is greater than the latter. These two comparisons are performed independently, each producing a true or false logical judgment result. Then, these two logical results are ORed. That is, if either comparison result is true (i.e., the trend value is too low or the volatility is too high), it is ultimately judged as an anomaly. Only when both comparison results are false is it considered an anomaly. Only when a certain condition is met is the system considered normal. The system assigns a flag value of 1 to abnormal situations and a flag value of 0 to normal situations. This process is executed sequentially along the time axis for each point, generating a sequence of 0s and 1s, which is the output multi-dimensional anomaly flag. Among them, the health index trend value reveals the true direction of change of the index over a longer period. A continuously declining trend value means that the overall performance or health status of the equipment is irreversibly deteriorating. Even if the current instantaneous absolute value of the health index has not yet fallen below a certain absolute threshold, this clear downward trend is a strong signal that early failures, progressive wear and tear, or performance degradation are occurring. The core value of this trend-based early warning lies in its foresight. It changes the passive mode of traditional monitoring, which only alarms when the indicator exceeds the standard, and enables maintenance strategies to prevent problems in advance.

[0044] S46. Process the warning signal based on the multi-dimensional anomaly markers to obtain the equipment anomaly warning signal. In this step, firstly, count the consecutive anomaly markers, that is, sum the number of consecutive 1s in the sequence, and record the specific time point corresponding to the marker. Then, compare the count result of this consecutive anomaly with a preset time length threshold to determine whether the former is greater than or equal to the latter, thereby determining the warning level: if the duration of the consecutive anomaly reaches or exceeds the threshold, a high-level warning signal is generated; if it is only a short-term or single-point anomaly, a lower-level warning signal is generated. In addition, the system will trace back the original detection dimension corresponding to the anomaly marker that triggered this warning, that is, determine whether it is a trend anomaly, a fluctuation anomaly, or both, thereby determining the warning type. The warning types mainly include trend anomaly warning, fluctuation anomaly warning, and composite anomaly warning. Finally, the determined warning level, warning type, and the start time of the anomaly are combined and packaged into a complete warning signal data packet.

[0045] This invention constructs two complementary dimensions describing the direction and stability of equipment status changes by separating and quantifying the health index into trends and fluctuations. Compared with the traditional method of only monitoring the absolute value of the index, it can capture abnormal patterns earlier and more comprehensively. The threshold calculation is adaptive, solving the problem that fixed thresholds cannot adapt to individual differences in equipment, changes in operating conditions, and natural performance degradation. The early warning generation integrates logical combination and time persistence rules, which can not only keenly capture risks from any dimension, but also effectively filter instantaneous interference through duration conditions, thereby improving detection sensitivity while significantly reducing the false alarm rate.

[0046] S5. Generate a maintenance plan based on the equipment abnormality warning signal to obtain the equipment maintenance plan; Existing fault diagnosis technology is mostly based on the personal memory of maintenance personnel or scattered document records, which is prone to misjudgment or loss of experience. It lacks systematic optimization and multi-dimensional comprehensive consideration, and it is difficult to quickly integrate knowledge across equipment and fault types. It often gives conservative but inefficient solutions or one-sided handling opinions, resulting in over-maintenance or problem recurrence. To solve the above problems, the specific implementation steps are as follows: S51. Based on the equipment abnormality warning signal, perform fault type matching to generate fault type; this step is based on logical condition judgment and pattern comparison. The system maintains a pre-built fault logic rule library. Each logical rule defines a series of logical conditions that must be met to diagnose a specific fault. For example, it may require the warning type to be equal to the abnormal trend and the warning level to be high. It may also require that the duration of the signal be greater than a certain threshold, such as more than 3 days. During matching, each attribute field of the currently input warning signal is compared with the conditions set by each rule in the rule library one by one using equal, greater than, or logical AND operations. When all the conditions of a rule are met by the current warning signal, the match is considered successful, and the fault type description corresponding to the rule is output. The fault type description includes the faulty component, the fault mode, and the possible main causes or typical characteristics. For example, wear of the front bearing of the drive motor is characterized by a continuous and slow downward trend in the health index, accompanied by an intermittent increase in the vibration fluctuation rate. Here, the faulty component is the front bearing of the drive motor, the fault mode is wear, and the main causes or typical characteristics are directly linked to the warning signal and data analysis results. That is, the continuous and slow downward trend of the health index corresponds to the abnormal trend warning, and the intermittent increase in the vibration fluctuation rate corresponds to the abnormal fluctuation warning.

[0047] S52. Based on the fault type, perform knowledge graph retrieval to generate a set of candidate maintenance strategies. In this step, the input fault type description is used as a query request to search and match in the equipment knowledge graph. The knowledge graph is a network database composed of massive nodes and connections. Nodes represent entity concepts such as specific faults, parts, maintenance operations, tools and materials, while connections represent various semantic associations such as causing, needing, applicable, and prerequisites. The system first locates the fault node in the graph that highly matches the input fault type. Then, it traverses and finds all maintenance operation nodes directly or indirectly connected to the fault node along a preset relationship path centered on countermeasures or maintenance plans. This process is executed through a specific query language of the graph database, involving the comparison of node attributes and the exploration of relationship paths, rather than numerical calculations. All relevant maintenance operation nodes found and their associated attribute information, such as step descriptions, estimated time, and required resources, are collected together to form a structured list as the output set of candidate maintenance strategies.

[0048] S53. Perform strategy optimization processing on the candidate maintenance strategy set to generate equipment maintenance plan. In this step, each candidate maintenance strategy is first scored according to multiple preset evaluation dimensions, such as estimated maintenance time, required spare parts cost, implementation complexity, and historical success probability. The scoring of each dimension may involve the quantitative conversion of strategy attribute values, such as mapping high, medium, and low qualitative descriptions to specific numerical scores. Then, a weight coefficient is assigned to each evaluation dimension to reflect its importance in the current decision. Next, for each strategy, its score on each dimension is multiplied by the corresponding dimension's weight coefficient. Then, all the product results are added to obtain a weighted total score representing the overall priority of the strategy. The system sorts all strategies from high to low according to the weighted total score, and finally selects the top three ranked strategies. The detailed information of these strategies, including specific operation steps, required resources, estimated time, safety precautions, etc., is organized and formatted according to a standard template to generate a complete maintenance plan. For example, firstly, multi-source raw time-series data such as vibration and temperature of a centrifugal fan over several months are collected. After cleaning, alignment, and normalization, time-domain, frequency-domain, and wavelet packet features are extracted and fused for dimensionality reduction. Then, its standardized health index sequence is calculated through dual-model fusion. Assuming the current value is 72, analysis shows that the index has shown a continuous and slow downward trend over the past two weeks, with the trend value below the adaptive threshold. Moreover, the high-frequency volatility of vibration has increased significantly recently, with the volatility exceeding the adaptive threshold, triggering a composite anomaly warning. Through rule base matching, this warning pattern is diagnosed as early wear of the front bearing of the drive motor. Subsequently, the system retrieves candidate strategies for this fault from the knowledge graph, including vibration spectrum confirmation, bearing clearance inspection, lubricant replenishment, and bearing replacement preparation. After comprehensive scoring and ranking, the system prioritizes immediately performing vibration spectrum analysis to confirm the wear characteristics and arranges bearing opening inspection and special grease replenishment within the next 24-hour downtime window as the core solution. The system also includes inspection steps, required grease type, estimated time of 2 hours, and corresponding safety lockout procedures, forming a report that is directly sent to the equipment maintenance data center.

[0049] This invention uses a rule base to accurately match early warning signals with fault types, avoiding the subjectivity and inconsistency of manual diagnosis, and significantly improving the accuracy and speed of fault location. By using a knowledge graph for maintenance strategy retrieval, it can systematically associate and call historical best practices, solving the problems of isolated maintenance knowledge, easy loss, or low retrieval efficiency in traditional methods. Through a multi-dimensional strategy ranking model, it outputs optimized solutions, comprehensively considering multiple constraints such as cost, time, and resources, making the final decision recommendations not only technically feasible but also in line with the economic efficiency of operation and maintenance and the actual situation on site, realizing scientific and refined recommendations for maintenance.

[0050] Example 2: Because existing technology threshold settings often rely on historical extreme values, they are prone to false alarms or slow degradation that is difficult to detect. They ignore the changing trends and fluctuations over time, making them highly insensitive to early, slowly changing faults. Please refer to [link to relevant documentation]. Figure 2 The diagram shown is a structural block diagram of the device health index analysis system based on a big data center provided in this embodiment. The system includes a preprocessing module, a feature fusion module, a health module, an anomaly warning module, and a maintenance plan module. The preprocessing module is used to collect the raw device timing data of the target device and obtain normalized timing data through preprocessing based on the raw device timing data. The feature fusion module is used to perform feature decomposition based on normalized time series data to obtain multi-dimensional feature vectors, and to perform multi-modal feature extraction and fusion on the multi-dimensional feature vectors to generate fused feature vectors. The health module is used to predict health based on the fused feature vector, generate a first predicted health index and a second predicted health index, and perform health conversion on the first predicted health index and the second predicted health index to obtain the device health index value. The anomaly warning module is used to perform pre-detection processing on the equipment health index value to obtain the health index trend vector and health index fluctuation characteristics. Based on the health index trend vector and health index fluctuation characteristics, the module generates an equipment anomaly warning signal through warning detection processing. The maintenance plan module is used to generate maintenance plans based on equipment anomaly warning signals, thus obtaining the equipment maintenance plan.

[0051] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code, including but not limited to disk storage, CD-ROM, optical storage, etc.

[0052] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for analyzing equipment health index based on big data centers, characterized in that, The steps of this method are as follows: collect the raw device timing data of the target device, and obtain normalized timing data by preprocessing the raw device timing data; Based on normalized time series data, feature decomposition is performed to obtain multidimensional feature vectors. Multimodal feature extraction and fusion are then performed on the multidimensional feature vectors to generate fused feature vectors. Based on the fused feature vector, health status is predicted, generating a first predicted health status index and a second predicted health status index. The first predicted health status index and the second predicted health status index are then converted to health status to obtain the device health status index value. The equipment health index value is pre-detected and processed to obtain the health index trend vector and health index fluctuation characteristics. Based on the health index trend vector and health index fluctuation characteristics, an early warning signal for equipment abnormality is generated through early warning detection processing. A maintenance plan is generated based on the equipment abnormality warning signal, resulting in an equipment maintenance plan.

2. The device health index analysis method based on big data centers according to claim 1, characterized in that, Based on the original equipment time series data, preprocessing is performed, including: data cleaning of the original equipment time series data to obtain cleaned time series data; Perform data alignment on the cleaning time-series data to generate time-series aligned data; Normalized time-series data is obtained by normalizing the time-aligned data.

3. The device health index analysis method based on big data centers according to claim 1, characterized in that, Feature decomposition based on normalized time series data includes: performing time-domain feature processing on normalized time series data to obtain time-domain feature vectors; Frequency domain features are extracted based on normalized time series data to generate frequency domain feature vectors; Wavelet packet decomposition is performed on the normalized time series data to obtain wavelet energy feature vectors.

4. The device health index analysis method based on big data centers according to claim 1, characterized in that, Multimodal feature extraction and fusion of multidimensional feature vectors includes: feature concatenation based on multidimensional feature vectors to generate concatenated feature vectors; The encoder performs dimensionality reduction based on the concatenated feature vectors to generate encoded feature vectors. The encoded feature vector is subjected to feature fusion processing to obtain the fused feature vector.

5. The device health index analysis method based on big data centers according to claim 1, characterized in that, Health prediction based on fused feature vectors includes: dividing the fused feature vectors into time series segments to generate an input vector, which includes training feature vectors and test feature vectors; The first predicted health index is obtained by performing support vector regression processing based on the trained feature vectors. A second predicted health index is generated by processing the training feature vectors through Gaussian regression.

6. The device health index analysis method based on big data centers according to claim 1, characterized in that, The first predicted health index and the second predicted health index are converted into a health index, including: fusing the prediction results based on the first predicted health index and the second predicted health index to obtain a fused predicted health index. Based on the fusion-predicted health index, a standardized health index is generated through standardization processing; in this step, two key boundary values ​​are determined based on historical or reference data; The standardized health index is processed to obtain the equipment health index value.

7. The device health index analysis method based on big data centers according to claim 1, characterized in that, Pre-detection processing of equipment health index values ​​includes: processing time-series data based on equipment health index values ​​to obtain a health index time series; A weighted moving average is applied to the health index time series to generate a health index trend vector. Standard deviation processing is applied to the time series of the health index to obtain the fluctuation characteristics of the health index.

8. The method for analyzing equipment health index based on big data centers according to claim 1, characterized in that, Based on the health index trend vector and health index fluctuation characteristics, early warning detection processing is performed, including: adaptive threshold processing based on the health index trend vector and health index fluctuation characteristics to generate multi-dimensional dynamic abnormal thresholds; Threshold labeling is performed on the health index trend vector, health index fluctuation characteristics, and multi-dimensional dynamic anomaly thresholds to obtain multi-dimensional anomaly labels; Early warning signals are obtained by processing multi-dimensional anomaly markers to obtain equipment anomaly early warning signals.

9. The device health index analysis method based on big data centers according to claim 1, characterized in that, The maintenance plan is generated based on the equipment abnormality warning signal, including: matching the fault type based on the equipment abnormality warning signal and generating the fault type; Knowledge graph retrieval based on fault type generates a set of candidate maintenance strategies. The candidate maintenance strategy set is optimized to generate an equipment maintenance plan.

10. A system applied to the device health index analysis method based on big data centers as described in any one of claims 1-9, characterized in that, The system includes: The preprocessing module is used to collect the raw device timing data of the target device and obtain normalized timing data through preprocessing based on the raw device timing data; The feature fusion module is used to perform feature decomposition based on normalized time series data to obtain multi-dimensional feature vectors, and to perform multi-modal feature extraction and fusion on the multi-dimensional feature vectors to generate fused feature vectors. The health module is used to predict health based on the fused feature vector, generate a first predicted health index and a second predicted health index, and perform health conversion on the first predicted health index and the second predicted health index to obtain the device health index value. The anomaly warning module is used to perform pre-detection processing on the equipment health index value to obtain the health index trend vector and health index fluctuation characteristics. Based on the health index trend vector and health index fluctuation characteristics, the module generates an equipment anomaly warning signal through warning detection processing. The maintenance plan module is used to generate maintenance plans based on equipment abnormality warning signals, thus obtaining the equipment maintenance plan.

Images