Cement production equipment fault early warning method and system

Abstract

The invention discloses a cement production equipment fault early warning method and system, and relates to the technical field of cement production equipment health monitoring, and the method comprises the steps: building a cylindrical coordinate system on the surface of a rotary kiln cylinder, and dividing the cylindrical coordinate system into a plurality of independent visual monitoring surface elements; collecting infrared and visible light image sequences in multiple working conditions and multiple continuous rotation periods for each visual monitoring surface element in a healthy state, and constructing a healthy visual fingerprint reference library representing stable state characteristics of the visual monitoring surface element; according to the method, the rotary kiln cylinder visual monitoring surface elements are divided, the healthy visual fingerprint reference library is constructed, real-time accurate comparison is achieved in combination with the twin neural network, cylinder state monitoring can be comprehensively covered, and limitation of local monitoring is avoided; and through time sequence deviation degree calculation and spatial correlation analysis, the deterioration trend and spatial correlation characteristics of the fault precursor can be captured in an early stage, and the early warning timeliness is greatly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cement production equipment health monitoring, in particular to a cement production equipment fault early warning method and system. BACKGROUND

[0002] In the cement production process, the rotary kiln is the core equipment for completing the calcination of raw materials, and its running stability directly affects the production efficiency, product quality and production cost. The rotary kiln is in a harsh working condition of high temperature, heavy load and continuous rotation for a long time, and the cylinder is prone to local overheating, wear, cracking, and peeling of refractory materials. If such hidden dangers are not discovered and handled in time, they will gradually develop into serious faults, causing equipment downtime, production interruption and significant economic losses.

[0003] The existing rotary kiln fault monitoring methods have obvious limitations: The method of relying on periodic manual inspection is limited by the inspection cycle, personnel experience and field of view, and it is difficult to capture subtle changes in equipment in real time, the early warning is lagging, and the fault is often discovered after it appears; traditional single-point sensor monitoring can only cover a local area and cannot fully reflect the overall state of the cylinder, and is easily disturbed by high temperature, dust and other complex working conditions, with high false and missed alarm rates. At the same time, the existing technology lacks a systematic way to represent the health status of the equipment, and does not establish a time evolution and spatial correlation analysis mechanism for abnormal states, making it difficult to effectively identify fault precursor characteristics and unable to meet the actual needs of early fault warning of the rotary kiln. SUMMARY

[0004] To solve the above technical problems, the present application provides a cement production equipment fault early warning method and system. The following technical solutions are adopted: A cement production equipment fault early warning method, comprising the following steps: Step 1: Establish a cylindrical coordinate system on the surface of the rotary kiln cylinder and divide it into a plurality of independent visual monitoring surface elements; for each visual monitoring surface element, collect infrared and visible light image sequences under healthy state in multiple working conditions and multiple continuous rotation periods, and construct a health visual fingerprint reference library representing its stable state characteristics; Step 2: During online operation, real-time acquisition of the current images of each visual monitoring surface element is performed, and the current images are compared with the health visual fingerprints of the corresponding working conditions and phases in the health visual fingerprint reference library to calculate the similarity between the current state and the healthy state; based on the similarity change trend of each surface element in multiple continuous rotation periods, the time sequence deviation degree representing the state deterioration degree is calculated; Step 3: Real-time analysis of the spatial distribution of the time sequence deviation degrees of all visual monitoring surface elements is performed to identify abnormal regions with correlation; the spatial integral of the deviation degrees in the correlated regions is performed to calculate the abnormal energy accumulation value and accumulation speed of the region; Step 4, according to the individual threshold of the time sequence deviation, the cumulative value and the cumulative speed of the abnormal energy, combined with the spatial form of the abnormal area, trigger the corresponding pre-warning of different levels, and output the position information and suspected fault type of the associated visual monitoring facet.

[0005] Optionally, it further includes step 5, according to the review results of the pre-warning on the spot, supplement the confirmed normal variant data to the health visual fingerprint reference library.

[0006] Optionally, in step 1, the division of the visual monitoring facet is specifically: dividing the axial section along the axial direction of the rotary kiln cylinder at a fixed length interval, dividing the circumferential piece along the circumferential direction of the cylinder at a fixed angle interval, and the intersection of the axial section and the circumferential piece defines a facet.

[0007] Optionally, in step 1, the construction of the health visual fingerprint includes: for each facet, extracting the temperature mean value distribution, temperature standard deviation distribution and temperature time sequence stability index of the infrared image sequence; extracting the depth texture feature mean value, color mean value and texture cross-cycle change tolerance of the visible light image sequence.

[0008] Optionally, in step 2, the comparison is realized by a twin neural network, one branch of the twin neural network inputs the current image block, the other branch inputs the reference image block generated from the health visual fingerprint, and the output of the twin neural network is the similarity score.

[0009] Optionally, in step 2, the calculation of the time sequence deviation is to take the similarity score sequence of the corresponding facet in the last N continuous rotation periods, and calculate the weighted sum of the negative change slope and the number of periods below the health threshold.

[0010] Optionally, in step 3, the spatial correlation identification includes: identifying the facets that are continuously arranged along the axial direction and have high deviation, and determining as line correlation; identifying the facets that are adjacent to each other in the same axial section and have high deviation, and determining as surface correlation.

[0011] Optionally, in step 4, the dynamic grading pre-warning is specifically: When the time sequence deviation of a single facet first exceeds the first threshold, a first-level observation pre-warning is triggered; When line correlation or surface correlation is formed, and the regional abnormal energy exceeds the second threshold, a second-level warning pre-warning is triggered, and the fault type is prompted; When the cumulative speed of the regional abnormal energy exceeds the third threshold, a third-level action alarm is triggered; The calculation formula of the abnormal energy cumulative value is: ; wherein R represents a potential abnormal region identified by spatial correlation analysis, the region R is composed of one or more spatially adjacent or related visual monitoring surface elements; i represents the i-th visual monitoring surface element in the region R; represents the current time series deviation degree of surface element i, quantifies the degree of state deterioration, represents the change rate of the time series deviation degree of surface element i in the latest unit time, and is used to represent the acceleration of the deterioration trend; and are state weight coefficients and trend weight coefficients, respectively, represents the normalized coefficient of the surface area of the cylinder represented by surface element i; represents a spatial position weight function, and the value depends on the spatial position s of surface element i, and the distance from the key component, the closer the distance, the higher the weight.

[0012] Optionally, in step 4, in step 5, the optimization comparison model is to take the fault precursor data as a training sample, and perform supervised fine-tuning training on the twin neural network, maintain the health state discrimination ability while improving the identification sensitivity to specific precursor patterns.

[0013] A cement production equipment fault early warning system for realizing a cement production equipment fault early warning method, the system comprising a visual acquisition unit, a data processing and storage unit, an intelligent analysis unit, a warning and interaction unit and a closed-loop learning module; The visual acquisition unit is used for a pan-tilt control system for controlling the shooting angle and rhythm, and synchronizing it with the rotation phase of the rotary kiln; The data processing and storage unit is used to store the rotary kiln cylinder space grid model, the health visual fingerprint reference library of each surface element, and perform online image acquisition, label binding and data preprocessing; The intelligent analysis unit is built-in twin neural network model, used to perform time series comparison and time series deviation calculation of surface element image, and run spatial correlation analysis and abnormal energy evaluation algorithm; The warning and interaction unit is used to generate graded warning information according to the analysis result, and highlight the abnormal region on the three-dimensional visualization model, and provide a historical data query and feedback input interface; The closed-loop learning module is used to receive field feedback information, manage incremental updates of the reference library and trigger optimization training tasks of the comparison model.

[0014] In summary, the present application includes at least one of the following beneficial technical effects: This invention provides a method and system for early warning of faults in cement production equipment. By dividing the rotary kiln body into visual monitoring elements and constructing a healthy visual fingerprint benchmark library, and combining it with a twin neural network to achieve real-time and accurate comparison, it can comprehensively cover the monitoring of the kiln body status and avoid the limitations of local monitoring. Through time-series deviation calculation and spatial correlation analysis, it can capture the deterioration trend and spatial correlation characteristics of fault precursors at an early stage, greatly improving the timeliness of early warning. The tiered early warning mechanism, combined with the cumulative assessment of abnormal energy, can accurately trigger early warnings based on risk levels, facilitating targeted on-site handling. The closed-loop learning module enables incremental updates to the benchmark library and model optimization, continuously improving the accuracy of early warnings.

[0015] The 3D visualization of abnormal areas can quickly locate the fault location, effectively reduce equipment downtime, ensure continuous and stable operation of cement production, and reduce economic losses. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a method for early warning of faults in cement production equipment according to the present invention. Detailed Implementation

[0017] The present invention will be further described in detail below with reference to the accompanying drawings.

[0018] This invention discloses a method and system for early warning of faults in cement production equipment.

[0019] Reference Figure 1 Example 1: A method for early warning of faults in cement production equipment, comprising the following steps: Step 1: Establish a cylindrical coordinate system on the surface of the rotary kiln shell and divide it into several independent visual monitoring elements; for each visual monitoring element in a healthy state, collect infrared and visible light image sequences in multiple working conditions and multiple continuous rotation cycles to construct a health visual fingerprint benchmark library that characterizes its stable state features. Step 2: During online operation, the current image of each visual monitoring element is acquired in real time and compared with the health visual fingerprint of the corresponding working condition and phase in the health visual fingerprint benchmark library to calculate the similarity between the current state and the health state; based on the similarity change trend of each element in multiple consecutive rotation cycles, the temporal deviation degree representing the degree of continuous deterioration of the state is calculated. Step 3: Analyze the spatial distribution of temporal deviation of all visual monitoring elements in real time, identify related abnormal regions; perform spatial integration on the deviation within the related regions, and calculate the cumulative value and velocity of abnormal energy in the region. Step 4: Based on the individual threshold of temporal deviation, the cumulative value and accumulation rate of abnormal energy, and the spatial morphology of the abnormal area, trigger warnings of different levels and output the location information of the associated visual monitoring elements and the suspected fault type.

[0020] Example 2 also includes step 5, which involves supplementing the confirmed normal variant data into the health visual fingerprint benchmark database based on the on-site verification results of the early warning.

[0021] By adopting the above technical solution, the rotary kiln shell is divided into independent visual monitoring elements using a cylindrical coordinate system. Essentially, this transforms the complex shell surface into a precisely positioned and independently monitored unit, solving the problem that traditional overall monitoring cannot locate local anomalies. Acquiring infrared and visible light image sequences under multiple operating conditions and multiple continuous rotation cycles in a healthy state is to cover the normal state fluctuations of the rotary kiln under different production loads (such as full load and start-up / shutdown stages) and different rotation phases. This ensures that the health visual fingerprint can comprehensively characterize the stable operation characteristics of the equipment, providing an accurate "health reference standard" for subsequent comparisons.

[0022] The real-time comparison and temporal deviation calculation principle involves online acquisition of the current image and comparison with the health visual fingerprint, focusing on dual matching of operating conditions and phase to avoid misjudgments caused by changes in operating conditions or phase differences. The comparison is achieved using a Siamese neural network, which utilizes its accurate ability to distinguish similar features to output a quantified similarity score, intuitively reflecting the difference between the current state and the health state. The temporal deviation calculation takes the similarity sequence of the most recent N rotation cycles, and through the weighted sum of the negative change slope and the number of cycles below the health threshold, it transforms the single similarity difference into a quantitative indicator of the degree of continuous deterioration, capturing the cumulative process of fault precursors.

[0023] Spatial correlation analysis and anomaly energy calculation principles: Faults on rotary kiln shells often exhibit spatial correlation. Therefore, by identifying line correlation and surface correlation, it is possible to distinguish between occasional anomalies of a single surface element and regional fault precursors, reducing false alarms caused by isolated anomalies. The calculation of anomaly energy accumulation values ​​introduces state weight, trend weight, area normalization coefficient, and spatial location weight. This is to comprehensively consider the degree of deterioration, deterioration acceleration, coverage area, and correlation degree of key components of each surface element within the region, thereby achieving accurate quantification of the severity of anomalies and avoiding the limitations of single-indicator evaluation.

[0024] The dynamic hierarchical early warning principle is based on the progressive logic of individual anomaly - regional correlation - deterioration rate to set up hierarchical early warning: the first-level observation warning is for the first time a single surface element exceeds the limit, so as to achieve early detection; Level 2 warnings are issued when spatial correlations are formed and abnormal energy levels are met, clearly indicating that early signs of a fault have appeared and suggesting the type of fault, thus enabling early diagnosis. The Level 3 action alert targets abnormal energy accumulation rates exceeding limits, predicting that the fault will rapidly worsen and enabling early intervention. The tiered mechanism can match different response strategies according to the severity of the anomaly, avoiding both excessive warnings that disrupt production and serious anomalies that are ignored.

[0025] A closed-loop learning mechanism of early warning, review, and optimization is constructed to continuously improve the accuracy and sensitivity of the early warning system. The specific principle is as follows: The optimization principle of the health visual fingerprint benchmark library is to verify the authenticity of the early warning results through on-site verification. The confirmed normal variant data is essentially a special operating condition or phase feature that was not covered by the initial benchmark library under the health state. Supplementing such data to the health visual fingerprint benchmark library can expand the coverage of the benchmark library, reduce misjudgments caused by incomplete health state representation, make the health reference standard more in line with the dynamic changes of actual equipment operation, and improve the accuracy of subsequent comparisons.

[0026] The principle of comparison model optimization is to use fault precursor data, which contains early features of equipment progressing towards failure, as training samples to perform supervised fine-tuning of the Siamese neural network. The core is to enhance the model's ability to extract and identify features of specific fault precursor patterns while preserving its ability to distinguish between healthy and abnormal conditions. Through fine-tuning training, the model can more sensitively capture weak fault signals that are easily overlooked by traditional comparisons, solving the problem of delayed early warning caused by indistinct early fault precursor features, and further improving the sensitivity of early fault identification.

[0027] In Example 3, step 1, the visual monitoring element division is specifically as follows: axial segments are divided at fixed length intervals along the axial direction of the rotary kiln body, and circumferential plates are divided at fixed angle intervals along the circumference of the body. The intersection of the axial segments and the circumferential plates is defined as a single element.

[0028] In Example 4, step 1, constructing a healthy visual fingerprint includes: for each facet element, extracting the mean temperature distribution, standard deviation temperature distribution, and temperature time-series stationarity index of the infrared image sequence; and extracting the mean depth texture features, mean color, and texture cross-period variation tolerance of its visible light image sequence.

[0029] By adopting the above technical solution, the rotary kiln shell has different functional sections along the axial direction (such as the feeding section, calcination section, and cooling section). The operating load and vulnerability of each section are different, but the operating status of the equipment within the same section is relatively consistent. Dividing the axial section into fixed-length intervals ensures that the size of the shell area covered by each axial section is uniform. This avoids both the masking of local anomalies due to excessively long sections and the increase of data redundancy due to excessively short sections. At the same time, the fixed-length axial sections can be adapted to the actual physical structure of the shell (such as refractory brick masonry units and the spacing of support devices), which facilitates the subsequent association of monitoring data with specific equipment components and improves the physical correlation of anomaly location.

[0030] The rotary kiln operates by dividing the circumferential sections at fixed angular intervals. Rotating along a fixed axis, the stress and heat conditions at various positions along the circumference of the kiln change periodically with rotation (e.g., periodic switching of relative position to the flame). However, the fluctuations within the same circumferential angle range exhibit a periodic pattern. Dividing the circumferential sections at fixed angular intervals ensures that each section corresponds to a fixed phase interval during the kiln's rotation, guaranteeing accurate matching of subsequent image sequences with the rotation phase and avoiding discrepancies between health baselines and online data due to phase misalignment. Simultaneously, the fixed-angle circumferential sections ensure uniform distribution of circumferential monitoring units, guaranteeing no blind spots in the kiln's circumference.

[0031] The principle of defining surface elements by the intersection of axial segments and circumferential plates is to form independent surface elements through the intersection of axial segments and circumferential plates. Essentially, it constructs a monitoring unit with unique spatial coordinates: each surface element corresponds to a specific combination of "axial position - circumferential angle" on the cylinder surface, and its spatial position can be accurately identified through a cylindrical coordinate system. This division method discretizes the cylinder surface into regular, uniform, and uniquely positioned monitoring units. This facilitates the visual acquisition unit to plan the shooting path according to the surface element position (such as the pan-tilt control system synchronously tracking each surface element), and also facilitates the subsequent binding of data such as health visual fingerprints and temporal deviation to specific surface elements to achieve accurate spatial positioning of abnormal states. At the same time, it provides a structured unit basis for the spatial correlation analysis (line correlation, surface correlation) in step 3.

[0032] Example 4 clarifies the feature extraction dimensions of health visual fingerprints. The core principle is based on the different physical information carrying capacities of infrared and visible light images, extracting key features that can characterize the "thermal stability state" and "physical morphological stability state" of surface elements respectively. By combining multi-dimensional features, a comprehensive and accurate health status benchmark is constructed, avoiding the limitation that a single feature cannot cover the health status of the device. The specific principle is as follows: Infrared images primarily carry temperature information of surface elements. Temperature stability is a core indicator of the healthy operation of a rotary kiln (e.g., localized overheating often indicates refractory material spalling, cylinder wear, and other faults). Therefore, three types of temperature-related features are extracted to characterize the thermal stability state: The average temperature reflects the average temperature level of the surface element under multiple operating conditions and multiple rotation cycles, and is the basic benchmark for the thermal state of the surface element. The distribution of the average temperature under different operating conditions can cover the temperature range under normal production load fluctuations, avoiding misjudging normal load changes as abnormalities.

[0033] Temperature standard deviation reflects the fluctuation range of surface cell temperature within a period. Temperature fluctuations under healthy conditions should be within a stable range. The standard deviation distribution can quantify the boundary of normal fluctuations. If the standard deviation of subsequent online data exceeds this distribution range, it indicates abnormal temperature fluctuations.

[0034] Temporal stationarity reflects the temperature change trend of the surface element over a continuous period. Under healthy conditions, the temperature time series should not have a significant increasing / decreasing trend. By calculating the stationarity of the temperature series, early signs of thermal anomalies such as slow temperature rise can be captured, avoiding the omission of gradual fault signals due to focusing only on instantaneous temperature.

[0035] Visible light images primarily carry information about the surface physical morphology of surface elements. Physical morphological stability is a core indicator of the structural health of equipment. Therefore, three types of morphological features are extracted to characterize the physical stability state: Depth texture reflects the microscopic uneven structure of the surface of a pixel. In a healthy state, the texture features are consistent. Extracting the mean of the depth texture can establish a baseline morphology for normal texture. If wear or cracks appear on the surface later, the mean of the depth texture will deviate from the baseline, enabling early identification of structural anomalies.

[0036] The surface color of the element is directly related to the material composition and the degree of thermal aging. The color mean can quantify the color benchmark under healthy conditions. Abnormal deviations in the color mean can reflect chemical / physical changes caused by material aging and local overheating, supplementing material state information that cannot be covered by infrared temperature characteristics.

[0037] The principle of texture cross-period variation tolerance extraction: Under healthy conditions, the surface texture of the element will undergo slight periodic changes as the cylinder rotates, but the variation range is within a fixed range; texture cross-period variation tolerance is to quantify the boundary of this normal change. If the texture variation range exceeds the tolerance in subsequent periods, it indicates that the structural integrity is damaged, thus realizing the identification of aberrant morphological anomalies.

[0038] Infrared and visible light features characterize the health status of surface elements from two dimensions: thermal and physical states, respectively, and the two are complementary. By combining multiple features, a comprehensive health benchmark that includes thermal and physical stability can be constructed, avoiding fault types that cannot be identified by single-dimensional features. This ensures that the health visual fingerprint can accurately and comprehensively reflect the stable operating status of surface elements, providing a multi-dimensional reference for subsequent online comparison.

[0039] In Example 5, in step 2, the comparison is achieved through a Siamese neural network. One branch of the Siamese neural network takes the current image patch as input, and the other branch takes the reference image patch generated from the healthy visual fingerprint as input. The output of the Siamese neural network is the similarity score.

[0040] In Example 6, step 2, the temporal deviation is calculated by taking the similarity score sequence of the corresponding surface element within the most recent N consecutive rotation cycles, and calculating the weighted sum of the negative change slope of the sequence and the number of cycles that are continuously below the health threshold.

[0041] In Example 7, step 3, spatial correlation identification includes: identifying surface elements that are continuously arranged along the axial direction and have high deviation, and determining them as line correlation; identifying surface elements that are adjacent circumferentially arranged within the same axial segment and have high deviation, and determining them as surface correlation.

[0042] By adopting the above technical solution and utilizing the dual-branch architecture of the Siamese neural network, the current face image block and the reference image block generated by the health visual fingerprint are respectively input. The features are extracted and the feature similarity is measured in parallel through the dual branches. Finally, a quantified similarity score is output to accurately determine the degree of difference between the current face state and the health state.

[0043] The similarity score sequence of the N most recent consecutive rotation cycles of the surface element is selected. The negative change slope of the sequence and the number of cycles that are continuously below the health threshold are calculated together. The degree of continuous deterioration of the surface element state is quantified by weighting the two.

[0044] Based on the spatial diffusion pattern of rotary kiln faults, surface elements that are continuously arranged along the axial direction and have high deviation are identified as line-related; adjacent circumferentially arranged surface elements with high deviation within the same axial segment are identified as surface-related. This distinguishes isolated anomalies from regional fault precursors and clarifies the scope of fault impact.

[0045] In Example 8, step 4, the dynamic hierarchical early warning specifically refers to: When the temporal deviation of a single face cell exceeds the first threshold for the first time, a level one observation warning is triggered; When a line or area correlation is formed and the abnormal energy in the region exceeds the second threshold, a level two warning is triggered, and the fault type is indicated. When the rate of abnormal energy accumulation in the area exceeds the third threshold, a Level 3 action alert is triggered. The formula for calculating the cumulative value of abnormal energy is as follows: ; Where R represents a potential anomaly region identified through spatial correlation analysis, and region R is composed of spatially adjacent or related visual monitoring elements; i represents the i-th visual monitoring element in region R; It represents the current temporal deviation of cell i, quantifying the degree of continuous deterioration of the state; it represents the rate of change of the temporal deviation of cell i in the most recent unit time, used to characterize the acceleration of the deterioration trend. and These are the state weight coefficient and the trend weight coefficient, respectively. The normalization coefficient represents the surface area of ​​the cylinder represented by element i; This represents the spatial location weighting function, whose value depends on the spatial location s of surface element i relative to the critical component; the closer the distance, the higher the weight.

[0046] In Example 9, in steps 4 and 5, the optimized comparison model uses fault precursor data as training samples to perform supervised fine-tuning training on the twin neural network, maintaining the ability to distinguish healthy states while improving the sensitivity to identify specific precursor patterns.

[0047] By adopting the above technical solution, a three-level logic triggers early warning: the first time the temporal deviation of a single surface element exceeds the first threshold, triggering a level one observation warning; the formation of a line or surface association and the abnormal energy in the region exceeding the second threshold triggers a level two warning and indicates the fault type; the accumulation rate of abnormal energy in the region exceeds the third threshold, triggering a level three action alarm. The abnormal energy in the region is calculated by comprehensively considering the temporal deviation of the surface element, the deterioration acceleration, the surface area normalization coefficient, and the distance weight of key components, quantifying the severity of the regional anomaly.

[0048] By using fault precursor data as training samples, supervised fine-tuning training is carried out on the Siamese neural network. While maintaining its ability to distinguish healthy states, the model's sensitivity to specific fault precursor patterns is improved, and the comparison accuracy is optimized.

[0049] The dynamic tiered early warning system is based on a progressive assessment of anomalies and a quantitative risk evaluation. It incorporates multi-dimensional technical details on top of the basic triggering logic to ensure the accuracy and operability of the early warning system, as detailed below: Details of the three-level early warning threshold setting: The first threshold is determined based on the statistical analysis of the time-series deviation of healthy surface elements. By collecting time-series deviation data of at least 100 healthy rotation cycles, the upper limit of the 95% confidence interval is taken as the threshold to avoid accidental fluctuations triggering false alarms. The second threshold is calibrated in conjunction with historical fault cases, extracting the critical value of abnormal energy in the region before past faults, and adding 10%-15% safety redundancy (e.g., 50). The third threshold is set according to the equipment tolerance limit, referring to the safety boundary of the deterioration rate of key components of the rotary kiln (such as refractory materials and cylinder steel plates), and is set to ensure that the abnormal energy increase per unit time does not exceed 10.

[0050] Details of abnormal energy calculation parameters: α is 0.6-0.8, β is 0.2-0.4. For rapidly deteriorating faults, β can be increased to 0.5, and for slowly accumulating faults, α can be increased to 0.9. The calculation method is based on the maximum surface area of ​​a single element. It equals the current area of ​​the face cell divided by the maximum area of ​​the face cell, ensuring a balanced energy contribution from face cells of different sizes; Values ​​are assigned in segments based on the distance to key components: 1.0 for distances within 1 meter from core components such as rollers and transmission gears, 0.7 for distances between 1 and 3 meters, 0.5 for distances between 3 and 5 meters, and 0.3 for distances above 5 meters, highlighting the weight of anomalies in key areas; the energy accumulation rate is calculated by the difference in abnormal energy between two consecutive monitoring cycles, i.e., v=(E2-E1) / Δt, where Δt is the cycle interval.

[0051] Fault Type Indication Technical Details: The fault type determination for Level 2 early warning is based on matching features of the associated region. When there is a linear correlation and the axial segment is located in the calcination zone, it prioritizes indicating axial spalling of the refractory material; when there is a surface correlation and the circumferential plate corresponds to the flame jet area, it indicates localized flame deviation leading to overheating; simultaneously, it combines abnormal energy... The magnitude of (deterioration acceleration), if >0.2 indicates that the fault is deteriorating rapidly and the review interval needs to be shortened.

[0052] The comparative model optimization revolves around the technical details of sample selection, training control, and performance assurance, ensuring that the fine-tuned model balances health assessment and early warning recognition capabilities, as detailed below: Details of fault precursor sample processing: Samples are selected from historical early warning review data and must meet the conditions of confirmed fault precursors + clear image features, covering 6-8 typical precursors such as slight refractory material peeling (texture abrupt change), slight overheating of the cylinder (abnormal temperature gradient), and micro-cracks on the surface (discontinuous depth texture); the number of samples for each precursor type is not less than 500 sets, each set contains image blocks of the corresponding surface element in 3 consecutive rotation cycles, and is labeled with the precursor type and the degree of deterioration (mild / moderate) to provide clear signals for supervised training.

[0053] Fine-tuning training parameters: During fine-tuning, a small learning rate and batch training strategy are adopted. The initial learning rate is set to 1e-5, and the learning rate decays to 0.8 of the previous round every 5 rounds. The batch size is set to 32, and the training rounds are 15-20 to avoid overfitting. The training set is mixed with 70% healthy samples and 30% precursor samples. Healthy samples are randomly selected from the healthy visual fingerprint benchmark library to ensure that the model does not lose its ability to distinguish healthy states. The cross-entropy loss function is adopted, and a precursor sample weighted loss is introduced. The loss weight of mild precursor samples is increased by 1.2 times to enhance the model's attention to weak precursors.

[0054] Sensitivity verification details: After fine-tuning, the model is verified through two indicators: accuracy of precursor recognition and false positive rate for health. The accuracy of recognition of mild precursor samples should be ≥85%, and the proportion of healthy samples misclassified as abnormal should be ≤3%. At the same time, 10 sets of new precursor data that were not used in training are selected for testing. If the model can identify precursor features 2-3 hours before the failure occurs, the sensitivity optimization is considered to have met the standard. If the false positive rate for health exceeds the standard, the proportion of precursor samples needs to be reduced and 5 rounds of separate training with healthy samples need to be added to balance the model's discrimination tendency.

[0055] Example 10: A cement production equipment fault early warning system, used to implement a cement production equipment fault early warning method. The system includes a visual acquisition unit, a data processing and storage unit, an intelligent analysis unit, an early warning and interaction unit, and a closed-loop learning module. The vision acquisition unit is used to control the shooting angle and rhythm, synchronizing it with the rotation phase of the rotary kiln; The data processing and storage unit is used to store the rotary kiln shell spatial grid model and the health visual fingerprint benchmark library of each surface element, and to perform online image acquisition, tag binding and data preprocessing; The intelligent analysis unit has a built-in Siamese neural network model, which is used to perform temporal comparison and temporal deviation calculation of surface images, and run spatial correlation analysis and anomaly energy assessment algorithms. The early warning and interaction unit is used to generate graded early warning information based on the analysis results, highlight abnormal areas on the 3D visualization model, and provide a historical data query and feedback input interface. The closed-loop learning module is used to receive on-site feedback information, manage incremental updates of the benchmark library, and trigger optimization training tasks for the comparison model.

[0056] The following specific embodiments illustrate the implementation principle of the present invention: Taking a Φ4.8m×72m rotary kiln of a cement plant as the application object, and combining the technical solutions of the aforementioned Examples 1-10, a rotary kiln fault early warning system is implemented. The specific implementation process is as follows: Implementation of Cement Production Equipment Fault Early Warning Methods: Step 1: Visual monitoring surface segmentation and construction of a health visual fingerprint benchmark library; The cylindrical coordinate system was established with the kiln head of the rotary kiln as the origin, the Z-axis along the axis of the cylinder, the θ-axis along the tangent of the cylinder circumference, and the r-axis in the radial direction. Axial segments were divided at fixed intervals of 1.2m along the axial direction, resulting in 60 axial segments for the 72m kiln body. Circumferential segments were divided at fixed intervals of 15°, resulting in 24 circumferential segments. The intersection of the axial segments and the circumferential segments constituted the visual monitoring surface elements, forming a total of 1440 independent surface elements throughout the kiln body. Each surface element was uniquely identified by a sequence of "Z-axis number - θ-axis number".

[0057] The health visual fingerprint benchmark library was constructed by selecting three typical operating conditions of the rotary kiln: full load, 50% load, and start-up / shutdown phases. For each condition, 100 rotation cycles of infrared and visible light image sequences were continuously acquired. Multi-dimensional features were extracted for each surface element to construct the health visual fingerprint. Infrared image features: temperature mean distribution, temperature standard deviation distribution, and temperature time series stationarity index; Visible light image features: mean depth texture features, mean color, and tolerance to cross-period texture variations.

[0058] All feature data is stored in the health visual fingerprint benchmark database, and an index is created according to "operating condition-face number-rotation phase".

[0059] Step 2: Online comparison and time series deviation calculation; Real-time image acquisition and comparison are achieved by acquiring images of each facet in real time through a vision acquisition unit. The acquisition cycle is synchronized with the rotation of the rotary kiln, and the facet number, current operating condition, and rotation phase are bound to the images. A Siamese neural network is used for comparison. The network contains an input layer, a convolutional layer, a pooling layer, a fully connected layer, and a distance metric layer. One branch takes the current facet image block as input, and the other branch takes a reference image block with the same operating condition and phase from the benchmark library as input. The output is a similarity score of 0-1.

[0060] The temporal deviation is calculated by taking the similarity score sequence of the most recent 5 consecutive rotation cycles of each facet element, with a health threshold set at 0.8. The temporal deviation is obtained by calculating the weighted sum of the negative change slope and the number of cycles that are continuously below the health threshold, thus quantifying the degree of continuous deterioration of the facet element's state.

[0061] Step 3: Spatial correlation analysis and anomaly energy calculation; Spatial correlation identification: Real-time analysis of the temporal deviation distribution of 1440 facet cells was performed, with a high deviation threshold set at 1.5. Line association identification: If the temporal deviation of continuous surface elements along the axis is higher than the threshold, it is determined to be a line association; Surface association identification: If the temporal deviation of adjacent surface elements in the same axial segment is higher than the threshold, it is determined to be surface association.

[0062] Calculation of abnormal energy accumulation value and accumulation rate: Using formula ; Calculate the anomalous energy in the associated region, setting α=0.7, β=0.3, and Ai=1. The distance between the surface element and the key component is used to determine the value; the accumulation rate of abnormal energy is calculated by the ratio of the difference in abnormal energy within a continuous period to the time interval, and the trend of abnormal deterioration is quantified.

[0063] Step 4: Dynamic hierarchical early warning; Set warning thresholds: the first threshold is 1.5, the second threshold is 50, and the third threshold is 10 / 10min.

[0064] When the timing deviation of a single surface cell exceeds the first threshold for the first time, a level one observation warning is triggered. When a line or area correlation is formed and the abnormal energy in the region exceeds the second threshold, a level two warning is triggered, indicating the suspected fault type. When the rate of abnormal energy accumulation in the area exceeds the third threshold, a Level 3 action alert is triggered.

[0065] Step 5: Closed-loop optimization: Health visual fingerprint benchmark database updated; After on-site verification, the confirmed normal variant data will be added to the health visual fingerprint benchmark database to expand the coverage of the benchmark database; Fine-tuning of twin neural networks; Historical fault precursor data were collected and labeled. A training set was constructed with 70% healthy samples and 30% precursor samples, and the network parameters were fine-tuned. After fine-tuning, the model's accuracy in recognizing mild fault precursors was improved, and the misclassification rate of healthy samples was reduced.

[0066] Implementation of a fault early warning system for cement production equipment: The system consists of a visual acquisition unit, a data processing and storage unit, an intelligent analysis unit, an early warning and interaction unit, and a closed-loop learning module. The specific configuration and functional implementation are as follows: Visual acquisition unit hardware configuration: infrared camera, visible light camera, and pan-tilt control system; Function implementation: control the camera's shooting rhythm and angle to ensure that the image accurately covers a single surface element and is bound to a coordinate label.

[0067] Data processing and storage unit hardware configuration: server; Function implementation: stores the cylindrical space grid model and health visual fingerprint benchmark library, and performs online image preprocessing and data tag binding.

[0068] Intelligent analysis unit hardware configuration: GPU and processor; Function implementation: Built-in twin neural network, which performs image comparison, temporal deviation calculation, spatial correlation analysis and anomaly energy assessment in real time.

[0069] Early warning and interaction unit software platform: 3D visualization system, web-based early warning interface; Functionality: Highlighting abnormal areas, pushing tiered early warning information, and providing historical data query and feedback entry functions.

[0070] Closed-loop learning module deployment method: edge computing nodes collaborate with cloud servers; Function implementation: receive on-site feedback, manage incremental updates of the benchmark library, trigger model fine-tuning tasks, and verify the effect.

[0071] Fault identification lead time: Identifies early signs of faults 2-3 hours earlier than traditional manual inspections; Early warning accuracy: The accuracy rate of Level II and above early warnings reaches 92%, and the false positive rate for healthy samples is <3%; Operational efficiency: Unplanned equipment downtime is reduced by 40%, and annual operation and maintenance costs are reduced by 15%.

[0072] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for early warning of failure of a cement production plant, characterized in that, The method comprises the following steps: Step 1: A cylindrical coordinate system is established on the surface of the rotary kiln cylinder, and is divided into a plurality of independent visual monitoring surface elements; for each visual monitoring surface element, a plurality of infrared and visible light image sequences under healthy state in multiple working conditions and multiple continuous rotation periods are collected to construct a healthy visual fingerprint reference library representing the stable state characteristics; Step 2: During online operation, the current images of each visual monitoring surface element are collected in real time, compared with the healthy visual fingerprints corresponding to the working conditions and phases in the healthy visual fingerprint reference library, and the similarity between the current state and the healthy state is calculated; Based on the similarity change trend of each surface element in multiple continuous rotation periods, the time sequence deviation degree representing the state deterioration degree is calculated; Step 3: The spatial distribution of the time sequence deviation degrees of all visual monitoring surface elements is analyzed in real time, and the abnormal regions with correlation are identified; the spatial integration of the deviation degrees in the correlated regions is performed to calculate the abnormal energy accumulation value and accumulation speed of the region; Step 4: According to the individual threshold of the time sequence deviation degree, the accumulation value and accumulation speed of the abnormal energy, and the spatial form of the abnormal region, a corresponding different level of early warning is triggered, and the position information and suspected fault type of the correlated visual monitoring surface element are output.

2. The cement production equipment fault early warning method according to claim 1, characterized in that: It further comprises step 5: according to the review results of the on-site early warning, the confirmed normal variant data is supplemented to the healthy visual fingerprint reference library.

3. The method of claim 2, wherein the method further comprises: In step 1, the division of the visual monitoring surface element is specifically: the axial sections are divided along the axial direction of the rotary kiln cylinder at a fixed length interval, the circumferential slices are divided along the circumferential direction of the cylinder at a fixed angle interval, and the intersection of the axial section and the circumferential slice defines a surface element.

4. The cement production equipment fault early warning method according to claim 3, characterized in that, In step 1, the construction of the healthy visual fingerprint comprises: for each surface element, the temperature mean value distribution, temperature standard deviation distribution and temperature time sequence stationarity index of the infrared image sequence are extracted; the depth texture feature mean value, color mean value and texture cross-period change tolerance of the visible light image sequence are extracted.

5. The method of claim 4, wherein: In step 2, the comparison is realized by a twin neural network, one branch of the twin neural network inputs the current image block, the other branch inputs the reference image block generated from the healthy visual fingerprint, and the output of the twin neural network is the similarity score.

6. The method of claim 5, wherein: In step 2, the calculation of the time sequence deviation degree is to take the similarity score sequence of the corresponding surface element in the last N continuous rotation periods, and calculate the negative change slope of the sequence and the weighted sum of the number of periods continuously below the healthy threshold.

7. The method of claim 6, wherein: In step 3, the spatial correlation identification comprises: identifying the surface elements that are continuously arranged along the axial direction and have high deviation degrees, and determining as line correlation; identifying the surface elements that are adjacent in the same axial section and have high deviation degrees, and determining as surface correlation.

8. The method of claim 7, wherein: In step 4, the dynamic grading early warning is specifically: When the time sequence deviation degree of a single surface element exceeds the first threshold for the first time, a first-level observation early warning is triggered; When line correlation or surface correlation is formed, and the regional abnormal energy exceeds the second threshold, a second-level warning early warning is triggered, and the fault type is prompted; When the accumulation speed of the regional abnormal energy exceeds the third threshold, a third-level action alarm is triggered. The calculation formula of the abnormal energy accumulation value is: ； wherein R represents a potential abnormal region identified by spatial correlation analysis, the region R is composed of one or more spatially adjacent or related visual monitoring elements; i represents the i-th visual monitoring element in the region R; represents the current time series deviation degree of the element i, quantifies the degree of continuous deterioration of the state, represents the change rate of the time series deviation degree of the element i in the latest unit time, and is used to represent the acceleration of the deterioration trend; and are state weight coefficient and trend weight coefficient, respectively, denotes the normalized coefficient of the surface area of the cylinder represented by the facet i; denotes the spatial position weight function, the value of which depends on the spatial position s where the facet i is located, and the distance relative to the key component. The closer the distance, the higher the weight.

9. The method of claim 8, wherein: In step 4, in step 5, the optimization comparison model is to take the fault precursor data as a training sample, to carry out supervised fine-tuning training on the twin neural network, to maintain the health state discrimination ability while improving the identification sensitivity to specific precursor patterns.

10. A cement production plant failure early warning system characterized by: The system comprises a visual acquisition unit, a data processing and storage unit, an intelligent analysis unit, an early warning and interaction unit and a closed-loop learning module for realizing the cement production equipment fault early warning method of claim 9. The visual acquisition unit is used for controlling the shooting angle and rhythm, and a cloud control system that is synchronized with the rotary kiln rotation phase; The data processing and storage unit is used for storing the rotary kiln cylinder space grid model, the health visual fingerprint reference library of each surface element, and performing online image acquisition, label binding and data preprocessing; The intelligent analysis unit is built-in with a twin neural network model, which is used for performing time sequence comparison and time sequence deviation calculation of the surface element image, and running spatial correlation analysis and abnormal energy evaluation algorithm; The early warning and interaction unit is used for generating graded early warning information according to the analysis result, and highlighting the abnormal area on the three-dimensional visualization model, while providing a historical data query and feedback input interface; The closed-loop learning module is used for receiving field feedback information, managing incremental updating of the reference library and triggering optimization training tasks of the comparison model.

Images