Device predictive maintenance method with exception prediction capability

Abstract

The invention discloses an equipment predictive maintenance method with exception prediction capability, which relates to the technical field of equipment predictive maintenance, and comprises the following steps of: constructing a direction continuous factor based on a guiding change scalar and a historical statistical fluctuation value, and controlling a numerical value recalibration function in a normalization process by using the direction continuous factor, the trend characteristics of the original data are reserved through normalization processing; the normalized data is converted into a continuous curvature space, trend potential energy is formed by analyzing local slope continuity of a data change path, and path integration is executed based on the trend potential energy to generate a trend energy expression vector; and performing multi-order positive and negative axis disassembly on the trend energy expression vector, converting each axial change into an independent distribution curve, and generating a trend offset index through probability increment superposition convolution. The problem that weak trend degradation is difficult to recognize is solved, and early warning and accurate maintenance response of equipment state change are achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of equipment predictive maintenance, in particular to an equipment predictive maintenance method with abnormality prediction capability. BACKGROUND

[0002] The equipment predictive maintenance with abnormality prediction capability refers to, in an industrial scene, using multi-source operation data (such as temperature, vibration, current, voltage, pressure, etc.) collected from various industrial equipment, introducing advanced algorithms such as machine learning, deep learning or time series modeling, and constructing a prediction model capable of identifying potential operation abnormality trends in real time, so as to predict the abnormal state that may occur before the actual failure of the equipment, and formulate or trigger the corresponding maintenance strategy based on the prediction result. This method is different from the traditional periodic preventive maintenance and post-failure repair maintenance, and it emphasizes more on "data-driven active sensing". It can not only significantly reduce the unplanned downtime, but also minimize the waste of maintenance resources and improve the overall availability and production continuity of the equipment. In the industrial environment, the structure of the equipment system is complex, and the operation environment is harsh. Any single point failure may cause the whole line to stop production or even safety accidents. Therefore, by deploying the predictive maintenance mechanism with abnormality prediction capability, intelligent, refined and efficient equipment operation and maintenance management can be achieved, which is an important support means for realizing industrial intelligent upgrading and cost reduction and efficiency improvement.

[0003] In the industrial scene, the existing equipment predictive maintenance technology with abnormality prediction capability usually realizes the intelligent maintenance of the equipment through the cooperation of the following five key links: data acquisition, data preprocessing, state modeling, abnormality detection and prediction, and maintenance decision. First, the system continuously collects multi-dimensional data in the operation process of the equipment through various sensors (such as vibration sensors, temperature sensors, current and voltage detection devices, etc.) deployed on the equipment. Then, the collected data is preprocessed through filtering, normalization, missing data completion and other steps to ensure that it can be used for modeling analysis. In the state modeling stage, the system introduces machine learning algorithms (such as random forest, support vector machine) or deep learning models (such as LSTM, GRU) to establish a dynamic prediction model of the equipment operation state, and learns the boundary features between the normal and abnormal operation of the equipment in real time. Then, by trend prediction or deviation analysis, the system can identify potential operation abnormalities or degradation trends, and even predict the specific time window when the failure may occur. Finally, combined with the abnormal prediction result and the importance level of the equipment, the system will automatically generate or recommend maintenance strategies, such as replacing parts in advance, adjusting the operation condition, arranging planned maintenance, etc., so as to avoid the loss caused by sudden failure. Overall, this technical system is driven by data, centered on model, and oriented by scene, and provides an efficient, controllable and sustainable intelligent maintenance method for industrial equipment.

[0004] The existing technology has the following deficiencies: In the state data of the equipment operation, some key parameters may have a continuous and stable one-way micro-acceleration trend in a local time window. Although the amplitude is extremely small, such a trend often represents early signs of performance degradation inside the equipment, such as load accumulation, structural deformation or precision loss. In this case, since such a trend signal does not show a significant mutation in numerical value, the existing equipment predictive maintenance technology with abnormal prediction capability usually adopts fixed interval scaling or Z-score standardization method when performing data normalization processing, resulting in that the low-amplitude gradual change feature is compressed into an approximately constant value, so that its real trend change is flattened or lost in the modeling input, and cannot be perceived and utilized as an effective feature by the model. The existing technology cannot accurately determine the trend deviation behavior of the equipment according to such state data trajectory, because it does not consider the identification mechanism of such "low-amplitude trend" degradation feature, thereby causing the system to "see nothing" in the early degradation process, making the predictive maintenance means lose the ability to give early warning, and eventually may cause the equipment to run to the critical point to trigger maintenance response, increasing the risk of failure and maintenance cost.

[0005] The above information disclosed in the background section is only intended to strengthen the understanding of the background of the present disclosure, and therefore it can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0006] The purpose of the present application is to provide an equipment predictive maintenance method with abnormal prediction capability to solve the problems in the background.

[0007] In order to achieve the above-mentioned purpose, the present application provides the following technical solution: an equipment predictive maintenance method with abnormal prediction capability, specifically comprising the following steps: S1, constructing a multi-time hierarchical data acquisition window, generating a plurality of non-overlapping time intervals for the equipment state parameters, performing difference calculation in each time interval, and generating a directed change scalar according to the time series change direction; S2, constructing a direction continuity factor based on the directed change scalar and the historical statistical fluctuation value, and using the direction continuity factor to control the numerical rescaling function in the normalization process, so that the normalization processing retains the trend characteristics of the original data; S3, converting the normalized data into a continuous curvature space, forming a trend potential by analyzing the local slope continuity of the data change path, and generating a trend energy expression vector based on the trend potential by performing path integration; S4, decomposing the trend energy expression vector into multiple positive and negative axes, converting each axis change into an independent distribution curve, and generating a trend deviation index by probability increment convolution; S5, establish a mapping interval of the trend deviation index, use a critical change range in a historical sample to construct an asymmetric distribution model, set a three-section state classification interval, and map and classify the current trend deviation index; S6, according to the continuous change rate of the trend deviation index, deduce the local structure of the directional change scalar, dynamically update the time level weight and the normalization function parameter through the difference inversion mechanism, and realize online correction of the trend determination parameter.

[0008] Preferably, S1 specifically comprises: Set a basic sampling period and construct a plurality of time window sets with different lengths and mutually non-overlapping in time dimension, each time window is divided into a plurality of sampling points according to the basic sampling period, forming a data hierarchical structure with short, sub-medium and medium-long term scales; In each time window, the adjacent data difference value calculation is performed on the corresponding state parameter data sequence point by point, and a complete difference sequence is generated to represent the change direction and amplitude information in the time period; Based on the difference sequence, the positive and negative direction number ratio of each difference value is calculated, and the absolute average amplitude of the difference value is combined to generate a directional change scalar, which represents the direction consistency and trend intensity of the state change in the time window.

[0009] Preferably, S2 specifically comprises: Collect historical state parameter data of the equipment in a fault-free state, construct a plurality of sliding time windows based on a fixed sampling period, calculate the maximum value, minimum value, mean value and standard deviation in each sliding time window, and form a historical statistical fluctuation value set for reference; The directional change scalar in the current time period is compared with the corresponding historical standard deviation to obtain a direction continuity factor representing the trend direction and relative change amplitude; According to the positive and negative values of the direction continuity factor, the maximum or minimum boundary in the normalization function is adjusted respectively, so that the normalized state parameter data retains the original change direction feature; The normalized state parameter data is compared for consistency before and after normalization, and when the consistency ratio is lower than a set threshold, the direction continuity factor is adjusted and normalized again.

[0010] Preferably, S3 specifically comprises: The normalized state parameter sequence is regarded as a continuous change path, the change direction and bending strength between data segments are analyzed in time sequence, and a curvature space structure with time traceability is constructed; Based on the constructed curvature structure, it is judged whether the data change direction is continuous in each continuous time section, a trend continuity label is formed, and a trend direction corresponding weight is assigned to each trend section; According to the curvature characteristics and trend continuity weight of each time point, a trend potential sequence is generated, and the trend potential value is used to represent the trend direction and change amplitude of the point in the state evolution process. The trend potential sequence is accumulated and integrated according to the preset time section to form a trend energy expression vector containing multiple dimensions, which is used for subsequent state trend analysis and classification identification.

[0011] Preferably, S4 specifically includes: The trend energy expression vector is disassembled into positive and negative sub-vectors according to the numerical symbol of each dimension, respectively corresponding to the enhancement trend and weakening trend expression path of the state parameter in each time interval; Based on the positive and negative sub-vectors, trend probability density distribution curves are respectively constructed to reflect the occurrence probability and trend amplitude distribution form of the trend in the time structure in each direction; The positive and negative trend probability density curves are respectively converted into trend increment functions, and weighted convolution superposition is performed at the corresponding time position to form a trend shift index, which is used to represent the trend net change direction and intensity.

[0012] Preferably, S5 specifically includes: Collecting device historical operation data, and generating a trend shift index for each historical time point, comparing and labeling the state label corresponding to the time point, and constructing the corresponding relationship between the trend shift index and the device running state; Based on the corresponding relationship, the distribution characteristics of the trend shift index under each type of state are counted, an asymmetric distribution model is established, and the trend shift index is divided into three state classification intervals of central stable interval, negative degradation interval and positive enhancement interval; The trend shift index obtained by the current collection is mapped to the classification interval in real time, and the device running state is determined according to the mapping result, and the predictive maintenance decision is executed in linkage, including risk warning, maintenance suggestion pushing or monitoring frequency adjustment operation.

[0013] Preferably, S6 specifically includes: Based on the continuous change rate of the trend shift index, a trend shift abnormal window is identified, and a shift tracing mechanism is initialized to locate the time period of potential trend expression loss; The guided change scalar structure in the trend shift abnormal window is compared and analyzed in each time period, the ideal guiding structure of the corresponding time period is reconstructed, and the trend expression shift difference atlas is generated based on the structural deviation; According to the trend expression shift difference atlas, the weight distribution of the normalization function boundary parameter and the time level collection window is dynamically adjusted, and the parameter evolution process of each round of adjustment is recorded to realize the non-interrupted online correction of the trend determination parameter.

[0014] In the above technical solution, the present application provides technical effects and advantages: 1、The present application can effectively solve the problem of inaccurate identification of low amplitude trend degradation in the prior art, leading to early warning failure. By constructing a multi-time level data acquisition structure, combining the guided change scalar and the direction continuity factor, the system not only realizes the structural expression of the time sequence micro-change trend of the equipment operation parameter, but also ensures that the micro but continuous trend signal will not be compressed and disappeared in the normalization process through the dynamic reparameterization mechanism of the normalization function, and improves the identifiable of the trend characteristics in the subsequent modeling input. In addition, by mapping the normalization result to the continuous curvature space and constructing the trend potential and trend energy expression vector, the system realizes the dynamic modeling and structural quantization of the data evolution path, effectively enhancing the overall description ability of the trend direction, strength and continuity.

[0015] 2、The present application realizes accurate classification of the equipment running state in different trend evolution stages through the construction of trend deviation index and the state mapping mechanism of asymmetric distribution model; especially in identifying mild degradation or weak structural imbalance trend. The trend deviation index not only integrates the directionality, probability distribution characteristics and incremental rate information, but also can dynamically adjust the normalization function parameters and time acquisition weights through the identification and guided expression inversion mechanism of the trend deviation abnormal window, realize online correction of the model and continuous optimization of the trend judgment accuracy. The overall scheme has high trend sensitivity, model adaptability and deployment stability, can effectively support early risk discovery, maintenance decision response in advance and optimization control of overall operation cost of the equipment in long-term operation, and has significant technical popularization value. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments or the prior art, the drawings needed in the embodiments will be briefly introduced. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0017] Figure 1 The flowchart of the device predictive maintenance method with abnormal prediction capability provided by the present application. DETAILED DESCRIPTION

[0018] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations may, however, be implemented in many different forms and should not be construed as limited to the examples set forth herein; rather, these example implementations are provided so that this disclosure will be thorough and complete, and will fully convey the inventive aspects to those skilled in the art.

[0019] The present application provides a device predictive maintenance method with abnormal prediction capability, comprising the steps of: Figure 1The device predictive maintenance method with abnormal prediction capability shown specifically comprises the following steps: S1, a multi-time level data acquisition window is constructed, a plurality of non-overlapping time intervals are generated for a device state parameter, a difference calculation is performed in each time interval, and a change direction guide scalar is generated according to a time sequence change direction; In this embodiment, in order to realize high-precision identification of trend changes in the device state parameter, especially to identify a one-way micro-acceleration trend change in the form of extremely low amplitude in the early stage, a multi-time level data acquisition window is constructed in the following manner, and a data change direction scalar for representing trend directionality is generated in each layer window. The process comprises the following steps: First, a basic sampling period is set and a multi-time level data acquisition window set with a clear time span is constructed. The device generates a raw data stream of a plurality of state parameters in real time during operation, such as temperature, current, speed, vibration signal, etc. For these state parameters, a uniform basic sampling period needs to be defined first as the smallest analysis unit in the time domain, for example, data is collected once every second. Based on the period, a plurality of non-overlapping time windows of different lengths are constructed in the time dimension, for example: the first layer window has a length of 10 sampling points, the second layer window has a length of 30 sampling points, and the third layer window has a length of 60 sampling points. The windows are arranged in sequence on the time line, forming a time-non-intersecting layered window set. By setting the windows in a layered manner, the system can capture short-term, sub-medium-term and medium-long-term trend information at different time scales, providing sufficient time structure support for identifying micro-amplitude trends.

[0020] Secondly, in each data acquisition window, the state parameters in the window are sequentially subjected to time sequence difference calculation. Specifically, for the state data sequence in each window, starting from the first sampling point, the numerical difference between the current sampling point and the previous sampling point is calculated, and the difference calculation process in the entire window is completed. The difference value of each pair of adjacent data points is used to reflect the change direction and change amplitude in the time period. In the process of performing the difference, the complete difference value sequence is retained and is not compressed or threshold filtered, so as to ensure that any small but continuous change trend can be accurately recorded. The completion of this step provides original gradient information for subsequent trend directionality judgment, ensuring that the trend formation process will not be concealed due to preprocessing.

[0021] Taking the temperature parameter in the running process of the device as an example, assuming that the temperature data sequence collected in a non-overlapping time interval is [65.01, 65.03, 65.06, 65.10, 65.15, 65.21], then according to the difference calculation method, the difference between adjacent sampling points is calculated in turn, and the difference sequence is [+0.02, +0.03, +0.04, +0.05, +0.06]. This sequence reflects that the temperature in this time interval presents a stable, one-way, and gradually increasing upward trend. Although the change amplitude is very small at each step, the difference direction is consistent and the amplitude is increasing, which is exactly the early performance of the accumulation of device load or the gradual attenuation of system heat dissipation performance. In this implementation step, the difference sequence will be kept intact as the basis information for subsequent calculation of the directional change scalar, ensuring that the system has the ability to identify such continuous and small amplitude trends.

[0022] Thirdly, in each difference sequence, the directional feature of the time sequence is extracted, and the corresponding directional change scalar is generated. The extraction of the directional feature is based on the positive and negative direction statistics of all difference items in the window. The sign of each difference value is marked, the number of positive difference values and the number of negative difference values are counted, and the proportional difference is calculated. Then a directional consistency factor is constructed by combining the absolute average amplitude of the difference value, which is used to measure whether the change in the window has a continuous and biased feature. On this basis, a specific directional change scalar is calculated by multiplying the directional consistency factor and the average difference amplitude, which is used to quantify the directional trend and its stability of the state change in the window. The directional change scalar will be used as a key data feature for maintaining the trend characteristics in the next stage of data normalization and trend identification.

[0023] Taking the difference sequence [+0.02, +0.03, +0.04, +0.05, +0.06] as an example, all the difference values in the sequence are positive, indicating that the state parameter presents a continuous upward trend in the time window. When extracting the directional feature, first count the number of positive difference values as 5 and the number of negative difference values as 0, and calculate the directional proportional difference as (5−0) / (5+0) = 1.0, indicating that the directional consistency is very high; then calculate the absolute average amplitude of the difference value as (0.02+0.03+0.04+0.05+0.06) / 5 = 0.04. Multiply the directional proportional difference (1.0) and the average amplitude (0.04) to obtain the directional change scalar as 0.04. This value not only reflects the singleness of the change direction (all positive), but also reflects the relative intensity of the trend (stable upward), which can be used for the next stage of normalization processing and trend maintenance judgment, and is a quantitative description of the micro-acceleration change trend in the time window.

[0024] Finally, the calculated trend change scalar in each non-overlapping time interval is saved as a scalar result representing the directionality of the change trend in that interval, and is saved in a time-sequential data structure for use in the next step without trend combination or induction processing. In this step, the final output of the data processing is a set of trend change scalars with time sequence order, which does not constitute a trend expression or trend sequence structure, and such a structure is completed by the subsequent steps.

[0025] S2, based on the trend change scalar and the historical statistical fluctuation value, a direction continuity factor is constructed, and the direction continuity factor is used to control the numerical rescaling function in the normalization process, so that the normalization processing retains the trend characteristics of the original data; In this embodiment, in order to solve the technical problem that the traditional normalization method is prone to information compression and trend loss when processing low-amplitude one-way trend changes, a method is proposed for constructing a direction continuity factor based on a trend change scalar and a historical statistical fluctuation value, and using the factor to control the numerical rescaling function in the normalization process, so as to realize trend-preserving normalization processing. The embodiment specifically includes: Calculate the historical statistical fluctuation value set of the target state parameter: by selecting the state parameter historical data collected by the device in the fault-free running state for a long period of time (for example, the past one week data collected by a certain temperature sensor under stable working conditions), a sliding time window is constructed in fixed sampling period units, and the maximum value, minimum value, mean value and standard deviation of the parameter are extracted in each window, and then a summary interval of multiple window statistical results is constructed to represent the natural fluctuation interval of the state parameter under normal conditions. The statistical fluctuation value set reflects the variation range of the state parameter under non-abnormal running conditions, and will be used as a reference basis for subsequent trend comparison and interval rescaling.

[0026] Based on the trend change scalar and the historical statistical fluctuation value, a direction continuity factor is calculated: the historical standard deviation value obtained in the previous step is taken as a reference quantity of fluctuation intensity, and a normalized ratio operation is performed with the trend change scalar calculated in the current time interval, that is, the trend change scalar is divided by the corresponding standard deviation value to obtain a trend intensity ratio. The ratio is used to measure the relative intensity of the current trend in the background fluctuation. In addition, if the trend change scalar is positive, the direction continuity factor is positive; if the trend change scalar is negative, the direction continuity factor is negative. The finally formed direction continuity factor is a signed real number, and the absolute value represents the trend significance degree, and the sign represents the trend direction. The factor will be used as a control weight in the normalization process.

[0027] Taking the vibration acceleration parameter of the device as an example, it is assumed that the calculated guide change scalar is +0.015 in a certain time window, indicating that the overall data in this segment presents a stable upward trend. At the same time, the standard deviation of the parameter of the device extracted under the historical stable running condition is 0.005. Then, according to the normalized ratio operation, the trend intensity ratio is obtained by dividing the guide change scalar by the historical standard deviation, which is 0.015 ÷ 0.005 = 3.0. Since the guide change scalar is positive, it indicates that the current trend is positive, and the direction continuity factor is +3.0. The positive sign of the direction continuity factor indicates that the trend direction is continuously enhanced, and the absolute value 3.0 indicates that the intensity of the trend is significantly enhanced in the historical background fluctuation range. In the subsequent normalization processing, the direction continuity factor will be used as a dynamic scaling factor to adjust the upper limit of the normalization interval, so as to ensure that the slight but continuous trend change will not be distorted or lost due to normalization compression, effectively preserving its trend characteristics for further identification and learning by the model.

[0028] The direction continuity factor is used to control the numerical rescaling function in the normalization process: when normalizing the current state parameter, the linear scaling formula commonly used is: (X-min) / (max-min), where min and max are statically set upper and lower boundaries. In this embodiment, in order to avoid the trend signal being flattened in the normalization process, the boundary values of max or min need to be dynamically adjusted according to the value of the direction continuity factor: if the direction continuity factor is positive, it indicates that there is a positive trend, then max is pulled up to enhance the significance of the upward change; if the direction continuity factor is negative, then min is moved down to preserve the fine movement of the negative trend. The boundary adjustment ratio is controlled by the absolute value of the direction continuity factor, and a reasonable limit range is set to prevent abnormal fluctuations from being amplified. In this way, it can be ensured that the trend change will not be compressed to near zero invalid value due to boundary mismatch in the normalization process.

[0029] For example, assume that the original observation value of a certain state parameter in the current time interval is 5.12, and the minimum value of the parameter in the historical statistical fluctuation interval is 4.90 and the maximum value is 5.20, i.e., the default normalization interval is [4.90, 5.20]. In this time interval, the guiding change scalar obtained in the previous step is +0.015, and the historical standard deviation is 0.005, so the direction continuity factor calculated is +3.0, indicating that the parameter has a strong positive trend in this time period. In order to avoid the positive trend being compressed into a weak signal close to a constant value due to the narrowing of the normalization interval, the upper boundary of the normalization interval is dynamically adjusted according to the absolute value of the direction continuity factor. Multiply the absolute value of the direction continuity factor by the proportional adjustment coefficient α=0.005 to obtain the upper boundary correction amount 0.015, and the new upper normalization boundary is 5.215. The normalization formula is updated to: (X-4.90) ÷ (5.215-4.90), at this time the original value 5.12 is normalized to about 0.954, and if the boundary is not adjusted, the normalized value is only about 0.733. This way effectively improves the normalized expression strength of the current trend signal and avoids it being considered as a non-significant state and being weakened in weight in subsequent modeling.

[0030] To avoid the abnormal extreme value of the direction continuity factor causing the normalization boundary to be deformed too much, the system sets an upper threshold for the direction continuity factor when performing the rescaling function, for example, limiting its absolute value to no more than 5.0. When the direction continuity factor exceeds this upper limit, the truncated value is taken as the basis for calculation to prevent the boundary from expanding too quickly and affecting the stability of the data distribution. In the boundary adjustment strategy, the update of the upper and lower normalization limits uses an asymmetric mechanism: if the direction continuity factor is positive, only the upper normalization limit is adjusted; if the direction continuity factor is negative, only the lower normalization limit is adjusted; if the direction continuity factor is close to zero, no boundary adjustment is performed. This processing logic can enhance the sensitivity of the trend direction while avoiding overall value domain shift. In addition, the system can introduce a sliding window mechanism to smooth the direction continuity factor in multiple time intervals to generate a stable boundary correction factor, so that the normalization boundary has time extension and robustness, ensuring consistency in multi-time interval trend tracking, thereby ensuring the continuity and discriminability of the trend in the entire prediction process.

[0031] The state parameter sequence after the trend-preserving normalization processing is output, and the trend retention effect is tested: the state parameter data processed by the numerical rescaling function is subjected to consistency test, that is, the change direction (positive or negative) of the data before and after normalization is compared to see whether the consistency is maintained, if the retention rate exceeds the set threshold (such as 95%), it is considered that the normalization process does not destroy the original trend; if the change direction is reversed or the change rate is sharply weakened, the adjustment direction continuity factor is returned, and the normalization processing is performed again until the trend retention performance meets the standard. The verification mechanism makes the trend retention not only stay at the method design level, but also realize in the perceptible results of the normalized data output, ensuring that the slight trend can continue to participate in the prediction process in the subsequent modeling stage.

[0032] To ensure that the trend-preserving normalization processing not only has theoretical feasibility in method structure, but also has verifiability in actual output results, the embodiment further introduces a quantifiable verification mechanism of consistency test. Specifically, the original state parameter sequence before normalization and the processed result after normalization are respectively subjected to difference operation, forming two groups of difference sequences. Then, by comparing the change direction (i.e. difference sign) of each pair of adjacent data points in the two difference sequences, the consistency ratio of the direction is calculated, that is, the ratio of the number of positive and negative sign consistent point pairs to the total number of point pairs. If the ratio is greater than or equal to the set threshold (for example, 95%), it is considered that the normalization process performs well in retaining the trend direction, and can enter the subsequent trend modeling process; if the ratio is lower than the threshold, the system will automatically trigger the correction mechanism of the normalization process, re-evaluate the weight contribution of the direction continuity factor according to the time period where the difference direction reversal point is located, adjust the rescaling boundary and perform normalization processing again. Through this closed-loop verification and correction mechanism, the embodiment ensures that the trend characteristics have continuity, responsiveness and perceptibility in the whole data processing link, thereby improving the early identification ability of the system to the weak trend degeneration behavior.

[0033] S3, converting the normalized data into a continuous curvature space, forming a trend potential by analyzing the local slope continuity of the data change path, and generating a trend energy expression vector based on the trend potential by performing path integration; In this embodiment, the normalized data is converted into a continuous curvature space, a trend potential is formed by analyzing the local slope continuity of the data change path, and a trend energy expression vector is generated based on the trend potential by performing path integration, specifically including: The normalized state parameter data is mapped to a curvature space structure with continuous expression capability. After completing the trend-preserving normalization processing, the system obtains a sequence of state parameter values arranged at equal time intervals. To further identify the change structure of the trend, the sequence of values is first regarded as a dynamic trajectory arranged in time sequence. By continuously analyzing the morphological changes of adjacent data points in the trajectory, a data structure space with curvature concept is constructed. Specifically, in each local data interval formed by a group of consecutive sampling points, the overall bending direction and bending strength of the values in the interval are identified, which are used as the basic geometric depiction of the trend of state change in the interval. In the process of forming the curvature structure of the entire data sequence, the time sequence is always maintained, so that each curvature change can be traced back to its original time point. This mapping operation enables the system to not only focus on the change amplitude of a single point when analyzing the value trend, but also observe the continuous structure between adjacent data, thereby giving the value sequence an expression capability with direction perception and structural characteristics. After completing the mapping, the system promotes the normalized state data to a set of continuous change paths with time sequence morphological information, laying a foundation for further calculation of the trend direction and trend inertia.

[0034] In further implementation, to make the curvature space mapping process more engineering reproducible, the system processes the normalized state parameter data in a sliding interval manner. Specifically, a sliding window of fixed length is set, and the window is slid from the starting position of the sequence to the rear, containing a number of consecutive equally spaced sampling points. In each sliding window, the system observes the morphological changes of the local data segment and determines whether there is a significant arching, concave or direction turning behavior. For example, if the state values in a sliding window are 0.712, 0.716, 0.721, 0.727, and 0.734 in sequence, the data segment shows a stable upward trend, and the value changes gradually increase, indicating that the trajectory has a consistent upward bending structure, which is marked as a "positive curvature enhancement segment" in the curvature space. Conversely, if a turning point appears in a data sequence, i.e., the value first increases and then decreases or vice versa, the system identifies it as a "bending turning segment" and marks it as a "curvature reversal zone" in the curvature space. Through the above processing, the system not only constructs the trend change path of the data as a whole, but also identifies the geometric structure characteristics and their time sequence positions in the trend, providing a spatiotemporal locatable structure basis for subsequent identification of the inertia continuity, deviation strength and abnormal breakthrough point of the trend, realizing the extension of the trend recognition capability from one-dimensional value sequence to two-dimensional structure expression.

[0035] The trend continuity evaluation mechanism is formed by identifying the slope change trend of each section in the continuous time path based on the curvature space structure. After obtaining the data path with curvature characteristics, the system performs sliding analysis based on time sequence to determine whether the change trend of the data path in the continuous time period is consistent. The specific operation is as follows: in each continuous time section, the system extracts the change direction between adjacent data points and determines whether the direction remains continuous in the entire section, that is, whether the change direction frequently reverses. If the data change direction remains consistent at continuous multiple time points, it can be considered that the data in this section has obvious trend inertia. The system records these continuous time periods with consistent trends and marks their direction attributes, such as continuous rise or continuous fall. Further, the system can also assign different weight values to each time period in combination with the stability degree of the trend direction, to reflect the contribution strength of the data in this section to the overall trend. The weight not only represents the continuity of the trend, but also can express the significant degree of the trend in the numerical structure. The final trend continuity label will be one of the driving factors in the subsequent trend potential generation process, so that the trend potential not only contains the amplitude information of the numerical change, but also covers the trend inertia strength embodied by the direction consistency.

[0036] To make the trend continuity evaluation mechanism realizable and stable, the system uses the way of overlapping sliding intervals to analyze the whole time series in sections when dealing with the curvature space structure. Within each sliding interval, the system extracts the change direction of all adjacent data points in the interval and counts the consistency of these directions. For example, if the normalized state data in a sliding interval are 0.703, 0.707, 0.712, 0.718, and 0.725 in turn, four groups of adjacent data difference directions are all positive, indicating that the data in this section presents a stable upward trend. The system accordingly gives this section a "positive high consistency" label and further assigns a higher trend continuity weight, such as 0.9 or above, to this trend section. On the contrary, if the data change direction in an interval is frequently alternating, such as 0.712, 0.708, 0.715, 0.710, and 0.718, the system identifies it as a "directionally unstable section" and assigns a lower weight, such as 0.2. In addition, the system can also combine the time length of trend continuity in each interval, the trend direction retention rate, the trend bending degree, and other dimensions to evaluate each trend interval, and establish a trend inertia strength grade. For example, for a section that maintains consistent direction for 5 time points and the change amplitude shows an increasing relationship, it is rated as a "strong trend inertia section" to drive the construction of subsequent potential energy; while for a consistent change section that only maintains for two time points, it is marked as a "weak trend section" and its influence is weakened in subsequent processing. In this way, the system can construct a trend continuity label system containing both direction and strength for each position in the trend path, so that the generation of trend potential energy not only depends on the numerical form feature, but also strengthens the consistent feature recognition of trend direction, effectively improving the response ability and expression accuracy of weak trends.

[0037] According to the trend continuity and curvature characteristics, a trend potential expression sequence is constructed to quantify the trend intensity and direction of each position in the data path. The generation process of the trend potential is based on the above-mentioned continuous curvature expression and trend continuity identification. The system simultaneously refers to the path curvature characteristics and the trend continuity weight of the position where the data point is located, and combines the two to evaluate whether the data point has a significant trend performance. If the path bending characteristics of a certain point are obvious, and the trend direction in the section where the point is located has high consistency, it means that the point plays a role in promoting the trend in the current state sequence. The system assigns a higher trend potential value to such points; otherwise, in the area where the trend direction frequently reverses or the data change bending is not obvious, a lower potential value is given, indicating that the trend is not significant. In the process of forming the complete trend potential expression sequence, the system always maintains the continuity of the time sequence structure, ensuring that each trend potential value corresponds to a specific time position. The trend potential expression sequence can be regarded as a trend evolution description with directionality, structure and dynamics, which converts the original state parameter from a single numerical sequence to a trend expression object with "behavior attributes", laying a foundation for further integrating trend potential and forming a feature vector for discrimination.

[0038] To further ensure the stability and interpretability of the trend potential expression sequence, the system uses a combined scoring strategy when constructing the trend potential, which combines the path curvature characteristics and the trend continuity weight obtained in the previous step as input factors for joint evaluation. Taking a certain time point as an example, if the data change curve at the point has a significant upward trend, and it is in a rising interval with consistent direction for 6 consecutive time points, the system determines that the point has a significant trend promoting attribute in the current sequence and assigns it a higher trend potential value, such as 0.87. On the contrary, if the curvature at a certain data point is approximately zero or there is no continuous direction, and the change direction frequently repeats in the time period, the system considers the point as a non-trend dominant point and only assigns it a lower trend potential value, such as 0.12. The entire trend potential expression sequence is composed of the trend potential values of all time points, maintaining a one-to-one correspondence with the original state parameter sequence, so that it has complete time continuity. In practice, the trend potential sequence can be represented as a trend intensity curve fluctuating over time through visualization, allowing device maintenance personnel to clearly identify which time points or time segments constitute trend-enhanced areas, trend-stable areas, or trend-weakened areas. In this way, the system successfully converts a single numerical change trajectory into a structured trend expression with behavior description attributes, providing a clear, rich, and time-sequentially traceable trend signal basis for subsequent trend aggregation, deviation index generation, and health state identification.

[0039] The trend potential expression sequence is converted into a trend energy expression vector with a representative segment of the trend, and output for subsequent trend analysis and classification. After the construction of the trend potential sequence is completed, the system conducts an overall structural summary of the sequence, and extracts a set of vector features reflecting the overall behavior of the trend. Specifically, the system integrates the trend potential expression sequence in a paragraph form in the time axis order. For each previously set time sub-interval, the system summarizes all the trend potential values in the sub-interval to obtain the cumulative performance of the trend intensity in the time period. The summary results of each time period together form a set of trend energy sub-vectors, each sub-vector has a sequence in the time dimension and a comparison ability in the trend intensity dimension. The system can further perform standardization, weight distribution or classification mapping operation on the vector set as needed to generate the final trend energy expression vector. The expression vector, as a high-dimensional structured input for the trend recognition stage, not only contains the comprehensive representation of the trend directionality, trend amplitude and trend persistence, but also has good model adaptability and classification discrimination ability, providing key data support for high-precision device state trend classification and predictive maintenance.

[0040] In practical applications, to ensure that the trend energy expression vector has sufficient time resolution and trend discrimination ability, the system divides the trend potential expression sequence into multiple time sub-intervals with consistent length, for example, each 10 time sampling points is a sub-interval. Taking a trend potential sequence with a total length of 100 points as an example, the system can divide it into 10 continuous time periods, and sum the trend potential values in each time period to obtain the total performance of the trend intensity in the time period, and then generate 10 trend energy sub-vectors. For example, if the trend potential values in the first time period are mostly concentrated in the high position, such as 0.85, 0.88, 0.82, etc., the trend energy sub-vector of this segment shows a higher value, indicating that there is a significant stable trend in the device state in this time period; while if the trend potential values fluctuate greatly and are low on average in a time period, the corresponding trend energy sub-vector is low, reflecting that the data trend in this segment is not clear or frequent reversal occurs. The system can further standardize these sub-vectors on this basis, for example, map the overall trend energy range to a fixed interval, or introduce a weighting coefficient to emphasize the influence of key time periods, and finally form a unified format of the trend energy expression vector. The expression vector not only retains the trend persistence and directionality features in the original data, but also improves the feature compressibility and input consistency through structural integration, providing a directly callable high-dimensional structured input for subsequent trend state classification, deviation identification and prediction model training, greatly enhancing the system's ability to describe and respond to the evolution trend of the device state.

[0041] S4, the trend energy expression vector is decomposed into multiple positive and negative axes, each axis change is converted into an independent distribution curve, and a trend deviation index is generated by probability increment convolution; the specific process is as follows: The trend energy expression vector is decomposed in dimension and axis to construct positive and negative directional sub-sequences. After obtaining the trend energy expression vector, the system regards it as a structural data vector containing the expression of trend strength in multiple time periods. Each dimension of the vector corresponds to the trend energy value in a fixed period, with a clear time sequence attribute. To further identify the bias and evolution direction of the trend in each period, the system decomposes the vector in axis in this step. Specifically, the system first labels the trend energy value of each dimension as positive or negative direction, judges its trend direction as positive (growth), negative (decline), or neutral (fluctuation without obvious direction), and accordingly splits the overall trend energy expression vector into two independent but structurally corresponding sub-vectors: a positive sub-vector that only retains the energy values of all trend-enhancing segments, and a negative sub-vector that only retains the energy values of all trend-weakening segments. This decomposition not only preserves the direction attribute of the trend, but also forms two trend expression paths that are structurally consistent and directionally opposite. Further, to ensure the stability of subsequent calculations, the system will normalize the positive and negative sub-vectors respectively, making them comparable in amplitude and ensuring that the influence of different direction trends can be analyzed and integrated in parallel.

[0042] Taking an 8-length trend energy expression vector as an example, assuming the vector is: [0.12, 0.25, -0.08, 0.00, -0.15, 0.30, -0.05, 0.10], which represents the trend energy performance of the device state in the last 8 time segments. The system first determines the positive and negative directions of each value, and labels the positive values (such as the 1st, 2nd, 6th, and 8th positions) as "trend enhancement", and the negative values (such as the 3rd, 5th, and 7th positions) as "trend weakening", and the 0 value (4th position) as no obvious trend. Then, the system constructs a positive sub-vector: [0.12, 0.25, 0.00, 0.00, 0.00, 0.30, 0.00, 0.10], which only retains positive values, and the rest are assigned as 0; at the same time, a negative sub-vector is constructed: [0.00, 0.00, -0.08, 0.00, -0.15, 0.00, -0.05, 0.00], which only retains negative values, and the rest are set to 0. Next, to improve the consistency of subsequent analysis, the system normalizes the two sub-vectors respectively, for example, scaling according to the maximum absolute value, so that the positive sub-vector is normalized to [0.4, 0.83, 0.00, 0.00, 0.00, 1.00, 0.00, 0.33], and the negative sub-vector is normalized to [0.00, 0.00, 0.53, 0.00, 1.00, 0.00, 0.33, 0.00]. This processing method not only retains the structural and directional characteristics of positive and negative trends, but also lays a foundation for subsequent trend probability modeling and trend shift index construction.

[0043] A probability density distribution curve is constructed for the positive and negative sub-vectors, and the trend direction change profile is extracted. After separating the positive and negative trend energy, the system regards the positive and negative sub-vectors as a set of continuous data sequences with distribution characteristics, and performs probability density analysis based on the internal data fluctuations and time period distribution patterns. In this step, the system processes each sub-vector using a sliding window: the trend energy values in a fixed length time window are counted, the occurrence frequency and average amplitude of trend enhancement or weakening in this interval are calculated, and the probability density curve is constructed based on the summary results of all windows. The positive trend probability density curve describes the possibility and intensity of the upward evolution of the device operating state in continuous time, while the negative trend probability density curve reflects the frequency and amplitude concentration interval of the downward evolution or degradation of the operating state. Through this step, the system not only obtains the directional distribution of the trend in the time structure, but also constructs a trend description function with probability attributes, which converts the original energy value sequence into a trend behavior distribution. This process realizes the transformation from time series trend intensity description to trend behavior space, and provides a distribution basis for the next step of trend shift quantization.

[0044] Taking a positive sub-vector [0.4, 0.83, 0.00, 0.00, 0.00, 1.00, 0.00, 0.33] as an example, the system sets the sliding window length to 3 time points and performs sliding statistics. The first window covers positions 1-3, the trend value is [0.4, 0.83, 0.00], the non-zero value frequency is 2 / 3, and the average trend amplitude is (0.4+0.83) / 2 = 0.615; the second window covers positions 2-4, the value is [0.83, 0.00, 0.00], the non-zero frequency is 1 / 3, and the average amplitude is 0.83; and so on. The system traverses all sliding windows, records the frequency and average amplitude of trend enhancement in each window, and constructs an "enhanced trend occurrence probability distribution curve" about the time position. Similarly, for the negative sub-vector [0.00, 0.00, 0.53, 0.00, 1.00, 0.00, 0.33, 0.00], the system extracts sub-sequences such as [0.53, 0.00, 1.00] through sliding windows, calculates the frequency of 2 / 3, and the average amplitude of (0.53+1.00) / 2 = 0.765, and gradually generates a "trend weakening probability density curve". Finally, the two probability density curves respectively describe the distribution of trend enhancement and trend weakening in the time dimension, providing input basis with direction differentiation and probability description for subsequent trend shift convolution. This method effectively captures the time structure and amplitude concentration of trend behavior, enabling the system to establish a trend pattern discrimination mechanism based on probability behavior.

[0045] Incremental convolution fusion is performed on the positive and negative trend probability density curves to form a single trend shift index function. In order to unify the bidirectional trend behavior into a single index that can be compared, the system converts the above positive and negative probability density curves into trend increment functions and performs cross-direction convolution. Specifically, the system performs time axis sliding processing on the positive trend probability curve to extract the local trend strengthening degree at each position; and performs reverse sliding processing on the negative trend curve to extract the local trend weakening intensity. Subsequently, at each corresponding time position, the system performs weighted convolution integration of the positive trend strengthening increment and the negative trend weakening increment to form a single numerical output, representing the "net trend shift degree" at that time point. When the positive strengthening at a certain time point is much higher than the negative weakening, the shift index is positive, and vice versa. The system repeats this operation at all time points to finally generate a trend shift index sequence. This sequence has a complete time structure, reflecting the combined performance of the trend evolution direction and intensity of the device state at each time point during continuous operation. Compared with using trend energy value or direction value alone, the trend shift index integrates direction attribute, distribution width and change rate, and has stronger trend discrimination ability and system response ability.

[0046] Taking the previously constructed positive trend probability density curve [0.6, 0.8, 0.2, 0.0, 0.0, 0.9, 0.1, 0.5] and the negative trend probability density curve [0.0, 0.1, 0.7, 0.6, 0.9, 0.0, 0.2, 0.3] as an example, the system performs sliding incremental extraction on the positive curve from left to right, selecting the window [0.2, 0.0, 0.0] at position 3, with an average reinforcement degree of 0.067; and on the negative curve from right to left, extracting the window [0.9, 0.0, 0.2] with an average weakening intensity of 0.367. The system integrates these two values by weighted convolution, such as setting the positive weight to 0.6 and the negative weight to 0.4, then the net trend deviation degree of this position is (0.6x0.067) - (0.4x0.367) ≈ -0.088, indicating that the overall trend at the current position is weak and tends to be degraded. The system repeats this operation for all time positions, finally generating a trend deviation index sequence with a length of 8, for example: [+0.35, +0.41, -0.09, -0.20, -0.28, +0.47, -0.02, +0.19]. This sequence indicates the trend change intensity and direction of the device at each time point, with positive values indicating a trend towards reinforcement, and negative values indicating a degradation or performance decay tendency. Through the trend deviation index, the system not only realizes the fusion expression of bidirectional trend behavior, but also has the ability to dynamically track and respond to the trend evolution process, significantly improving the timeliness and pertinence of predictive maintenance.

[0047] The trend deviation index sequence is enhanced in stability and vectorized as a key indicator for subsequent evaluation. After generating the trend deviation index sequence, the system performs sliding average and outlier smoothing to remove spikes or sharp reversals caused by local transient fluctuations, enhancing the stability of trend discrimination. Subsequently, the system divides the deviation index sequence into several logical segments based on the set time period length, and extracts features such as maximum value, average value, positive and negative fluctuation ratio, and change slope for each segment. Finally, these local deviation features are combined into a trend deviation feature vector with a unified dimension and model input format, which is used for subsequent state classification, health level evaluation or warning strategy formulation. In addition, the system can also construct a risk accumulation index based on the dynamic change speed of the trend deviation index as a whole, to identify "trend turning point precursors" or "long-term deviation deposition sections", thereby further enhancing the forward-looking and precision of predictive maintenance capabilities. The generation of the trend deviation vector marks the completion of the whole process from the original state change path to the structured trend expression, enabling the entire system to have state recognition and decision-making capabilities oriented towards trend evolution.

[0048] S5, mapping interval of trend deviation index is established, critical change range in historical samples is used to construct asymmetric distribution model, three-section state classification interval is set, and current trend deviation index is mapped and classified; the specific process is as follows: The historical equipment operation data are collected and the corresponding trend deviation index is extracted, and a complete sample base is established to support the construction of trend classification mapping rules. In the actual application process, the system first collects the historical data of the equipment covering various running states. These data should come from long-term running monitoring records in real industrial scenes and cover the complete state evolution process from stable equipment performance to mild degradation and then to serious failure. The system processes these historical data one by one through the trend energy expression method constructed previously, and generates the trend deviation index value corresponding to each time point. At the same time, the system also needs to call the manual maintenance records, alarm records or quality control results corresponding to these historical trend deviation indexes, and perform time stamp alignment to identify whether there is human intervention or system abnormality identification when the trend deviation index changes. In this way, the system establishes the correlation mapping between the "trend deviation index change" and the "actual state transfer event", and provides a real basis for the subsequent establishment of classification standards. Further, the system marks the running events, such as "maintaining stable", "mild degradation occurs but no failure", "failure trigger" and the like, and analyzes the trend deviation index interval of these labels, extracts the typical deviation range corresponding to the critical change. The finally formed data set contains not only the deviation value itself, but also the running meaning embodied in the actual operation and maintenance environment, which lays a solid foundation for the construction of distribution model.

[0049] Based on the correspondence between the trend deviation index and the actual state label, an asymmetric distribution model is constructed, and a three-section classification interval for state recognition is divided. After forming the historical trend deviation index samples covering multiple state labels, the system statistically organizes these indexes, and finds that the distribution of the trend deviation index is not a symmetric structure. For example, in most scenarios, slight negative deviation can trigger early warning of device performance degradation, and even if the positive deviation is large, it does not necessarily mean risk, so the system no longer uses the average value symmetric segmentation method, but constructs a set of asymmetric trend deviation distribution models. The system classifies all deviation index samples according to their state, and respectively calculates their distribution boundary, peak frequency band and critical concentration area under different running states. For example, the system may find that the deviation index of slight degradation state is mostly concentrated between-0.10 and-0.25, while the deviation value of severe degradation is mostly concentrated in the interval below-0.30; at the same time, the deviation value of performance enhancement state is mainly distributed between +0.15 and +0.40. Based on this finding, the system divides three non-overlapping intervals: the first type is the central interval, representing stable running state, mainly covering the interval from-0.10 to +0.15; the second type is the negative degradation interval, extending from-0.10 to lower, representing different levels of performance degradation and failure risk; the third type is the positive enhancement interval, extending from +0.15 upwards, reflecting the potential performance improvement or high load running trend of the device. This three-section structure ensures that each type of trend direction has a clear attribution, and the introduction of asymmetric boundary controls the probability of misjudgment, so that the system has actual adaptability.

[0050] The trend deviation index obtained by current acquisition is mapped to the preset classification interval in real time to complete the state level judgment and predictive maintenance response linkage. After completing the interval modeling of the trend deviation index, the system enters the real-time mapping phase. Whenever new state data is generated during device operation, a specific numerical trend deviation index is obtained after data preprocessing, trend identification, normalized retention, trend potential conversion, and trend deviation generation. The system first reads the current deviation index value and compares it with the preset three-section classification interval: if the value falls into the central stable interval, it is determined that the current device operating state is normal and no maintenance intervention is needed, only the standard monitoring frequency is maintained; if the value falls into the negative degradation interval, the system immediately enters the warning state and further judges the depth position of the deviation value in the degradation interval, combined with the duration and historical comparison trend to determine whether to push the maintenance suggestion; if the value falls into the positive enhancement interval, it means that the device may be in a high load or structural stress enhancement stage, the system will judge whether the trend has risk attributes, and if necessary, adjust the normalization scale or time window to improve the model sensitivity. Through such a classification mapping mechanism, the trend deviation index not only has real-time discrimination ability, but also has dynamic regulation ability, and the numerical change directly guides the predictive maintenance system to respond, effectively realizing early identification, early prediction, and early intervention of potential risks of the device.

[0051] S6, according to the continuous change rate of the trend deviation index, the local structure of the directional change scalar is deduced, the time level weight and the normalization function parameter are dynamically updated through the difference inversion mechanism, and the online correction of the trend judgment parameter is realized.

[0052] Based on the continuous change of the trend deviation index, the rate of deviation mutation segment in the key time window is identified, and the deviation tracing mechanism is initialized. In the predictive maintenance scene of actual equipment, the trend deviation index should show a relatively continuous and slowly changing structure in the time dimension. However, in some cases, when potential weak abnormal evolution trends occur inside the equipment, the trend deviation index will show a continuously rising or falling accelerated change feature in a certain period of time. In order to capture such changes, the system first constructs a continuous tracking mechanism, processes the trend deviation index sequence in a time sliding window, and records the change direction and change amplitude of each period. By calculating the direction consistency and trend change intensity between multiple adjacent time segments, the system identifies the time period with significant accelerated evolution in the trend deviation sequence. Such time period is defined as a "trend anomaly window", representing the existence of obvious trend information in the state data that is not fully expressed by the current modeling structure. After completing the identification, the system takes the above "trend anomaly window" as the key section for subsequent reverse structure tracing, and starts the deviation tracing mechanism to compare and analyze all normalized input data in the abnormal window by time period, in order to deduce whether the current trend deviation change is caused by error accumulation or expression attenuation of the guiding change scalar. This step lays the foundation for the target window for subsequent difference inversion and online correction, ensuring that the decision adjustment process has clear boundaries and positioning basis.

[0053] The local structure of the guiding change scalar is established, and the difference inversion mechanism is executed to locate the trend deviation root cause. After completing the positioning of the trend deviation anomaly window, the system performs point-by-point structure analysis on the guiding change scalar structure corresponding to the window. The goal of this implementation step is to identify the missing directional expression that may have been ignored or compressed before normalization behind the trend deviation. The system first retrieves the original state parameter sequence before normalization corresponding to the abnormal window period, and reconstructs the difference sequence and direction marker sequence of this period to obtain the ideal guiding change scalar structure that should be formed. Then, the current guiding change scalar sequence is compared with the reconstructed guiding structure one by one, and the deviation in slope direction, direction consistency factor and continuity expression is analyzed. If it is found that the existing guiding structure has frequent fluctuations in trend direction, insufficient guiding continuity or change amplitude significantly lags behind the change characteristics of the state parameter, it can be determined that the guiding change scalar expression is distorted, leading to the expansion of the subsequent trend deviation judgment error. The system further quantizes the structural difference into a set of trend expression deviation difference signals, and collects them according to time to construct a trend deviation difference atlas, which clearly marks each point where the guiding expression deviation and trend change are inconsistent. This atlas serves as a direct reference for the system to make dynamic corrections, guiding the precision adjustment of the normalization function and time weight allocation parameters, and is a key support point for establishing a trend judgment feedback loop.

[0054] Based on the difference map results, the normalization function parameter correction and time hierarchy adjustment are implemented to realize the online evolution of trend judgment ability. After obtaining the difference map of trend expression deviation, the system enters the dynamic adjustment stage of trend judgment parameters. First, for the boundary adjustment problem of the normalization function, the system performs aggregated analysis on the time period where frequent misjudgments occur in the difference map, and counts the boundary distance distribution of these points in the original normalization process. If it is found that most of these trend expression missing points are concentrated in the upper and lower boundary areas of the normalization function, it means that the original boundary setting is too tight and cannot accommodate weak but continuous trend changes. At this time, the system automatically expands the normalization function boundary in this time period to the outside, and improves the expression fidelity of the trend change amplitude in the normalization process. At the same time, the system performs redistribution operation on the weight distribution strategy of the collection window in the time hierarchy structure, and transfers more sampling weight to the time period corresponding to the trend deviation abnormal window, and reduces the weight of other slow change periods, so as to improve the response ability of the overall modeling to the trend sensitive area. In addition, the system establishes a parameter evolution cache mechanism to record the boundary expansion amplitude and weight adjustment ratio in each adjustment process, and introduces a trend inertia factor to smooth the weight updating process, to prevent excessive adjustment caused by noise or short-term fluctuations. The whole parameter correction process is triggered immediately after the system detects that the trend deviation change rate exceeds the threshold, and has non-interruptive online updating ability, which can continuously adapt to the trend change structure of the device running state, and significantly enhances the stability and prediction sensitivity of the predictive maintenance system in the long-term deployment environment.

[0055] The above embodiments can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented by software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions according to the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another, for example, the computer instructions can be transferred from one website, computer, server or data center to another website, computer, server or data center through wired or wireless (such as infrared, wireless, microwave, etc.) mode. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. containing one or more available medium collections. The available medium can be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a DVD), or a semiconductor medium. The semiconductor medium can be a solid state disk.

[0056] It should be understood that the size of the sequence number of the above processes does not mean the order of execution in various embodiments of the present application, and the execution order of the processes should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0057] Those skilled in the art can appreciate that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether the functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0058] In several embodiments provided in the present application, it should be understood that the disclosed system and method can be implemented in other ways. For example, the above-described embodiments are only illustrative, for example, the division of units is only a logical function division, and actual implementation can have another division manner, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0059] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, they can be located in one place, or they can be distributed on multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the present embodiment.

[0060] In addition, the functional units in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit.

[0061] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method of predictive maintenance of equipment with anomaly prediction capability, characterized in that, Specifically comprising the following steps: S1, constructing a multi-time hierarchical data acquisition window, generating a plurality of non-overlapping time intervals for the device state parameter, performing difference calculation in each time interval, and generating a guiding change scalar according to the time sequence change direction; S2, based on the guiding change scalar and the historical statistical fluctuation value, a direction continuity factor is constructed, and a numerical rescaling function in the normalization process is controlled by using the direction continuity factor, so that the normalization process retains the trend characteristics of the original data; S3, the normalized data is converted into a continuous curvature space, the trend potential is formed by analyzing the local slope continuity of the data change path, and the trend energy expression vector is generated based on the trend potential; S4, the trend energy expression vector is decomposed into a plurality of positive and negative axes, each axis change is converted into an independent distribution curve, and a trend shift index is generated by probability increment convolution; S5, a mapping interval of the trend shift index is established, an asymmetric distribution model is constructed using the critical change range in the historical sample, a three-section state classification interval is set, and the current trend shift index is mapped and classified; S6, the local structure of the guiding change scalar is inversely deduced according to the continuous change rate of the trend shift index, the time level weight and the normalization function parameter are dynamically updated through the difference inversion mechanism, and the online correction of the trend determination parameter is realized.

2. The device predictive maintenance method with abnormality prediction capability according to claim 1, characterized by, S1 specifically includes: Set the basic sampling period and construct a plurality of time windows with different lengths and non-overlapping in time dimension, each time window is divided into a plurality of sampling points according to the basic sampling period, forming a data hierarchical structure with short, medium and long term scales; In each time window, the adjacent data difference value calculation is performed on the corresponding state parameter data sequence point by point, and a complete difference sequence is generated to represent the change direction and amplitude information in the time period; Based on the difference sequence, the positive and negative direction number proportion of each difference value is calculated, and the absolute average amplitude of the difference value is combined to generate a guiding change scalar, which represents the direction consistency and trend intensity of the state change in the time window.

3. The device predictive maintenance method with abnormality prediction capability according to claim 1, characterized by, S2 specifically includes: Collect the historical state parameter data of the device in the fault-free state, construct a plurality of sliding time windows based on the fixed sampling period, calculate the maximum value, minimum value, mean value and standard deviation in each sliding time window, and form a historical statistical fluctuation value set for reference; The guiding change scalar in the current time period is compared with the corresponding historical standard deviation to obtain a direction continuity factor representing the trend direction and relative change amplitude; According to the positive and negative values of the direction continuity factor, the maximum or minimum boundary in the normalization function is adjusted respectively, so that the normalized state parameter data retains the original change direction characteristics; The consistency of the change direction before and after the normalization of the state parameter data is compared, and when the consistency ratio is lower than the set threshold, the direction continuity factor is adjusted and normalized again.

4. The device predictive maintenance method with abnormality prediction capability according to claim 1, characterized by, S3 specifically includes: The normalized state parameter sequence is regarded as a continuous change path, the change direction and bending intensity between each data segment are analyzed in time sequence, and a curvature space structure with time traceability is constructed; Based on the constructed curvature structure, it is judged whether the data change direction is continuous in each continuous time segment, a trend continuity label is formed, and a weight corresponding to the trend direction is assigned to each trend segment; According to the curvature characteristics and trend continuity weight of each time point, a trend potential sequence is generated, and the trend potential value is used to represent the trend direction and change amplitude of the point in the state evolution process; The trend potential sequence is accumulated and integrated according to the preset time segment to form a trend energy expression vector containing multiple dimensions, which is used for subsequent state trend analysis and classification identification.

5. The device predictive maintenance method with abnormality prediction capability according to claim 1, characterized by, S4 specifically includes: The trend energy expression vector is disassembled into positive and negative sub-vectors according to the numerical symbol of each dimension, respectively corresponding to the enhancement trend and weakening trend expression path of the state parameter in each time interval; Based on the positive and negative sub-vectors, trend probability density distribution curves are constructed respectively to reflect the occurrence probability and trend amplitude distribution form of the trend in the time structure in each direction; The positive and negative trend probability density curves are respectively converted into trend increment functions, and weighted convolution superposition is performed at the corresponding time position to form a trend shift index, which is used to represent the trend net change direction and intensity.

6. The device predictive maintenance method with abnormality prediction capability according to claim 1, characterized by, S5 specifically includes: Collecting device historical operation data, and generating a trend shift index for each historical time point, comparing and labeling the state label corresponding to the time point, and constructing the corresponding relationship between the trend shift index and the device running state; Based on the corresponding relationship, the distribution characteristics of the trend shift index under each state are counted, an asymmetric distribution model is established, and the trend shift index is divided into three state classification intervals of central stable interval, negative degradation interval and positive enhancement interval; The current collected trend shift index is mapped to the classification interval in real time, the device running state is determined according to the mapping result, and the predictive maintenance decision is executed in linkage, including risk warning, maintenance suggestion pushing or monitoring frequency adjustment operation.

7. The device predictive maintenance method with abnormality prediction capability according to claim 1, characterized by, S6 specifically includes: Based on the continuous change rate of the trend shift index, the trend shift abnormal window is identified, and the shift traceability mechanism is initialized to locate the time period of potential trend expression loss; The guided change scalar structure in the trend shift abnormal window is compared and analyzed in each time segment, the ideal guiding structure of the corresponding time period is reconstructed, and the trend expression shift difference atlas is generated based on the structural deviation; According to the trend expression shift difference atlas, the boundary parameters of the normalization function and the weight distribution of the time level collection window are dynamically adjusted, and the parameter evolution process of each round of adjustment is recorded to realize the non-interrupted online correction of the trend determination parameters.

Images