Deep learning-based equipment phm failure mode recognition prediction method

Abstract

This invention discloses a deep learning-based method for equipment health management and fault prediction, specifically relating to the field of equipment health management and fault prediction technology. It addresses the problem that existing methods struggle to continuously obtain stable and comparable health status representations when observable parameter combinations change dynamically under abnormal sensor operating conditions, thus affecting the accuracy and continuity of fault mode identification and remaining lifetime prediction. By collecting vibration, temperature, and load time series data, a three-dimensional availability code is generated. A common health status vector is constructed by calling the corresponding subset state projector, and a fault mode category, monotonic degradation index, and remaining lifetime interval are generated by combining the state prototype set. The prototype matching boundary is updated when the code changes, achieving continuous robust inference.

Description

Technical Field

[0001] This invention relates to the field of equipment health management and fault prediction technology, and more specifically, to a deep learning-based equipment PHM fault mode recognition and prediction method. Background Technology

[0002] Equipment health management (PHM) technology is typically applied to equipment requiring long-term continuous operation, such as aerospace power plants, rail transit traction systems, industrial compressors, CNC machining equipment, electric drive assemblies, and continuous production facilities. It monitors the equipment's health status by collecting operating parameters such as vibration, temperature, and load, and uses this data to perform failure mode identification, degradation analysis, and remaining life prediction. Existing data-driven PHM methods generally establish a state characterization model based on a pre-defined set of parameters. Their modeling premise is that each monitoring parameter can be continuously, synchronously, and stably collected during operation, and that the input dimensions between adjacent monitoring windows remain consistent.

[0003] However, in practical applications, factors such as high temperature, impact vibration, electromagnetic interference, loose interfaces, sensor aging, maintenance and disassembly, and limited installation space often cause vibration, temperature, or load parameters to be missing, exceed limits, change abruptly, drift, or experience localized failures. This leads to changes in the combination of observable parameters for the same equipment at different monitoring periods. These changes are not a direct reflection of equipment degradation, but rather a reflection of changes in monitoring conditions at the data level. When existing methods still perform uniform feature extraction and status determination on data from different monitoring windows according to fixed input conditions, the distribution shift, correlation changes, and representation boundary drift caused by changes in parameter availability are easily superimposed on the actual health changes caused by component wear, fatigue, loosening, or performance degradation, thus causing confusion between perceived changes in conditions and actual equipment degradation. In this case, the fault categories, degradation trends, and life results output by the model may be discontinuous, jump, or distorted, thereby affecting the effectiveness of PHM results in operation and maintenance decisions, maintenance arrangements, and operational support.

[0004] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a deep learning-based equipment PHM (Problem Mode Recognition and Prediction) method. This method addresses the issue that existing methods struggle to consistently obtain stable and comparable health status representations when observable parameter combinations dynamically change under abnormal sensor operating conditions, thus affecting the accuracy and continuity of fault mode recognition and remaining lifetime prediction. By collecting vibration, temperature, and load time series data, a three-dimensional availability code is generated. A common health status vector is constructed by calling the corresponding subset state projector. Furthermore, a fault mode category, monotonic degradation index, and remaining lifetime interval are generated by combining the state prototype set. The prototype matching boundary is updated when the code changes, achieving continuous robust inference and thus resolving the problems mentioned in the background section.

[0006] To achieve the above objectives, the present invention provides the following technical solution: S1. Collect vibration time series, temperature time series and load time series within the preset monitoring window, perform unified time reference alignment, fixed-length window processing and amplitude standardization, and generate three-dimensional parameter availability codes arranged by vibration, temperature and load based on missing, out-of-bounds and abrupt change states; S2. When at least one bit of the three-dimensional parameter availability code is one, the corresponding subset state projector is called according to the three-dimensional parameter availability code to map the available parameters in the vibration time series, temperature time series and load time series to a common health state vector, and to make the complete parameter sample and the corresponding subset sample converge to the same state prototype neighborhood in the state prototype set. S3. Based on the public health state vector, three-dimensional availability code, and state prototype set, generate failure mode category, monotonic degradation index, and remaining life interval; S4. When the availability codes of the three-dimensional parameters of consecutive adjacent monitoring windows change and the switching confirmation conditions are met, switch the subset state projector, update the prototype matching boundary according to the change in distance between the common health state vector and the state prototype before and after the switch, and output the verified fault mode category, monotonic degradation index and remaining lifetime range.

[0007] Furthermore, the unified time reference alignment includes generating a unified time index according to the preset monitoring window length and unified sampling interval, and completing the alignment and writing of vibration time series, temperature time series and load time series based on the time deviation between each original sampling time and the corresponding unified time index.

[0008] Furthermore, the out-of-bounds state of the temperature time series is determined based on the temperature safety boundary corresponding to the load, and the abrupt change state of the load time series is determined jointly based on the differential amplitude of adjacent sampling points and the continuous duration, and a three-dimensional availability code is generated accordingly.

[0009] Furthermore, the available parameters are arranged in a uniform time index order and a fixed order of vibration, temperature, and load to form a subset observation sequence. The subset state projector corresponds one-to-one with a three-bit parameter availability code that has at least one bit set to one, and receives the corresponding subset observation sequence to perform state projection.

[0010] Furthermore, the state prototype set is established in the order of degradation stages, and the complete parameter samples form a complete observation sequence according to the corresponding degradation stage. After deleting parameter observations that are inconsistent with the training encoding, the complete observation sequence forms a subset sample, which is constrained to the same state prototype neighborhood together with the complete parameter samples.

[0011] Furthermore, the subset state projector sequentially performs observation coding, temporal state coding, and geometric median convergence on the subset observation sequence to obtain a pre-public health state vector, and performs scale normalization on the pre-public health state vector to form a public health state vector.

[0012] Furthermore, the failure mode category is determined based on the first distance between the reference center of each failure mode within the degradation stage corresponding to the public health state vector and the current most recent state prototype number, and the monotonic degradation index is generated based on the continuous position of the public health state vector in the direction of the line connecting adjacent state prototypes.

[0013] Furthermore, the remaining lifetime interval is generated based on the lifetime table of the failure mode stage corresponding to the current failure mode category. The center of the interval is determined by the current degradation stage position, and the half-width of the interval is determined by the base half-width of the failure mode stage, the excess prototype distance, and the number of unavailable parameters in the three-dimensional parameter availability code.

[0014] Furthermore, the switching confirmation conditions include that the current three-dimensional availability code has a continuous window number in the continuous monitoring window that is not less than the continuous occurrence window number threshold, and that the first distance between the current public health status vector and the previously verified public health status vector is not higher than the status continuity threshold.

[0015] Furthermore, after the switching confirmation conditions are met, coordinate median alignment is performed based on the previously verified public health state vector, the current public health state vector, and the most recent state prototype of the previous monitoring window to generate the verified public health state vector, and the prototype matching boundary is updated between the neighborhood radius of the basic prototype and the upper bound radius of the corresponding encoded boundary.

[0016] The technical effects and advantages of the deep learning-based equipment PHM fault mode recognition and prediction method of this invention are as follows: This invention does not follow the existing approach of first supplementing missing or abnormal parameters and then uniformly feeding them into the model for processing. Instead, it first determines which parameters within the current monitoring window are truly usable, and then directly enters the corresponding state projection path based on the available parameter combinations. Furthermore, it uniformly constrains the states obtained under different parameter combinations into the same health semantic space. The key to this approach is distinguishing between changes in observation conditions caused by sensor anomalies and state changes caused by equipment degradation. This ensures that fault mode identification, degradation assessment, and lifespan prediction are based on a consistent state foundation, rather than on speculatively recovered data. Therefore, in scenarios where parameters are missing, out of bounds, or abruptly changing, the source of the results is clearer, and the judgment criteria are more unified.

[0017] During continuous monitoring, this invention incorporates continuous verification, state checking, and boundary update mechanisms to address projection path switching caused by changes in parameter availability coding. This ensures that inference results between different combinations of observable parameters can be smoothly connected along the same state prototype link, preventing fault category jumps, degradation location distortions, or abnormal fluctuations in the lifespan due to short-term disturbances or instantaneous switching. Therefore, it maintains the continuity and consistency of output results under abnormal sensor operating conditions and directly reflects the impact of current observable conditions on the reliability range of results within the remaining lifespan. This makes the results more consistent with field operating characteristics and easier to use directly in practical health management scenarios. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the equipment PHM fault mode recognition and prediction method based on deep learning according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 This invention provides a deep learning-based method for equipment PHM (Problem Mode Recognition and Prediction) fault mode identification, including: S1. Collect vibration time series, temperature time series and load time series within the preset monitoring window, perform unified time reference alignment, fixed-length window processing and amplitude standardization, and generate three-dimensional parameter availability codes arranged by vibration, temperature and load based on missing, out-of-bounds and abrupt change states; S2. When at least one bit of the three-dimensional parameter availability code is one, the corresponding subset state projector is called according to the three-dimensional parameter availability code to map the available parameters in the vibration time series, temperature time series and load time series to a common health state vector, and to make the complete parameter sample and the corresponding subset sample converge to the same state prototype neighborhood in the state prototype set. S3. Based on the public health state vector, three-dimensional availability code, and state prototype set, generate failure mode category, monotonic degradation index, and remaining life interval; S4. When the availability codes of the three-dimensional parameters of consecutive adjacent monitoring windows change and the switching confirmation conditions are met, switch the subset state projector, update the prototype matching boundary according to the change in distance between the common health state vector and the state prototype before and after the switch, and output the verified fault mode category, monotonic degradation index and remaining lifetime range.

[0021] This invention addresses the core issue of distorted health management results under abnormal sensor operating conditions. It first identifies truly usable parameters within the current monitoring window from vibration and temperature loads, and solidifies these observable conditions into parameter availability codes. This eliminates the need for subsequent processing to supplement abnormal data, allowing direct state analysis based on truly observable combinations. Then, for different observable combinations, corresponding subset state projection paths are invoked to map available parameters to a unified health semantic space. State prototypes arranged according to degradation stages constrain the convergence of complete and subset observations to a consistent health state, thus distinguishing representational drift caused by missing parameters from actual equipment degradation. Furthermore, the unified health state simultaneously generates a fault mode category monotonic degradation index and a remaining lifetime interval, ensuring that diagnostic and predictive results are based on the same state, and reflecting the impact of current observable conditions on the reliability range of the results within the lifetime interval. Finally, when the parameter availability code changes, the output continuity of the monitoring window before and after is maintained by continuously confirming the state verification and boundary update. Therefore, this invention can not only maintain the consistency of fault identification and lifetime prediction under the condition of sensor missing and out-of-bounds sudden change, but also reduce the false switching, false diagnosis and lifetime jump caused by instantaneous sensing fluctuations, and improve the stability, interpretability and engineering usability of the results.

[0022] Step S1 is responsible for establishing the input boundary. During continuous operation, the vibration time series, temperature time series and load time series of the equipment are not always synchronous and stable. Local missing, out-of-bounds and sudden changes will directly change the observable conditions within the current monitoring window. If the real usable parameters are not identified at the source stage, the subsequent state characterization will confuse the interpretation of sensor anomalies with equipment degradation. Therefore, step S1 needs to organize the original samples into standardized inputs under a unified time reference and form a three-dimensional parameter availability code that can accurately characterize the current combination of observable parameters.

[0023] S101. Establishment of a unified monitoring window and time alignment processing.

[0024] Step S1 first sets the preset monitoring window length and the uniform sampling interval. The preset monitoring window length is determined by the shortest duration required for the target equipment to complete one stable load change process under the calibrated working conditions. The uniform sampling interval is jointly determined by the nominal refresh cycles of the vibration sampling link, temperature sampling link, and load sampling link.

[0025] The specific value selection logic is as follows: From the three types of nominal refresh cycles, a time interval is selected that can form a fixed number of sampling points within a preset monitoring window and is not later than the slowest refresh link update cycle. Then, a unified time index sequence is generated within the preset monitoring window according to the unified sampling interval. The original vibration sampling sequence, original temperature sampling sequence, and original load sampling sequence are matched for the nearest time according to the unified time index sequence. For each unified time index position, the nearest vibration sampling time, temperature sampling time, and load sampling time are searched. When the time deviation between the searched corresponding sampling time and the unified time index position does not exceed the corresponding channel time tolerance, the corresponding original sampled value is written into the vibration time sequence and temperature time sequence. For timing and load timing sequences with the same name, when the time deviation exceeds the corresponding channel's time tolerance, a blank marker is written in the corresponding position. The time tolerance of the vibration channel, temperature channel, and load channel are all determined by the time deviation samples under fault-free calibration conditions. The determination process is as follows: the absolute deviation between the actual sampling time and the theoretical sampling time of the three types of sampling links under fault-free calibration conditions is statistically analyzed. After sorting them from smallest to largest, the absolute deviation at the last preset percentage position, such as 5%, is taken as the basic tolerance. Then, it is compared with half of the corresponding nominal refresh cycle, and the larger value of the two is taken as the corresponding channel's time tolerance. This forms a vibration timing sequence, temperature timing sequence, and load timing sequence that share a unified time reference.

[0026] S102. Point-level valid determination and dimensionless sequence formation.

[0027] After the vibration time series, temperature time series, and load time series are formed, step S1 continues to determine the validity of the three types of time series point by point. Among them, the point-level validity determination of the vibration time series is completed only based on whether the corresponding position is a missing marker. Non-missing positions are recorded as valid vibration positions, and missing positions are recorded as invalid vibration positions. The point-level validity determination of the temperature time series is based on the missing marker determination, and the temperature safety boundary determination under the load constraint is further superimposed. The temperature safety boundary is generated by the load boundary node table, which is derived from the target equipment's fault-free stable operation sample.

[0028] The specific generation process is as follows: First, divide the load into several load boundary node intervals according to the load value from low to high. Then, collect the corresponding temperature sample at each load boundary node. After sorting the temperature samples from low to high, take the temperature at the first preset percentage position (e.g., 1%) as the lower lower boundary and the temperature at the second preset percentage position (e.g., 1%) as the upper temperature boundary. When the current load value is between two adjacent load boundary nodes, both the lower and upper temperature boundaries are linearly interpolated according to the load position ratio. When the current load value is outside the first node or outside the last node, the first node boundary value or the last node boundary value is used respectively. When the temperature time sequence position is neither a missing identifier nor falls between the lower and upper temperature boundaries corresponding to the current load value, it is recorded as a valid temperature position; other positions are recorded as invalid temperature positions. The point-level validity determination of the load time sequence is completed based on the load change rate and duration. The load change rate is calculated by subtracting the dimensionless load value at the current unified time index position from the load change rate. The dimensionless load value at the previous unified time index position is divided by the unified sampling interval to obtain the load change rate at the current position. When the absolute value of the load change rate at several consecutive positions is not lower than the load change rate judgment threshold and the change direction is consistent, the corresponding continuous segment is regarded as the load change mutation segment. The load change rate judgment threshold is jointly determined by the converted value of the target equipment's rated load ramp rate and the load change rate at the position located at the next preset percentage, such as 5%, in the fault-free switching condition sample. The larger of the two values ​​is taken as the load change rate judgment threshold. After the point-level effective judgment is completed, all non-empty and effective positions are normalized. The vibration time series is normalized by the vibration calibration lower bound and vibration calibration upper bound, the temperature time series is normalized by the temperature calibration lower bound and temperature calibration upper bound, and the load time series is normalized by the load calibration lower bound and load calibration upper bound. Invalid positions continue to retain the empty mark, thus forming a dimensionless vibration sequence, a dimensionless temperature sequence, and a dimensionless load sequence.

[0029] S103. Determine the availability of a window.

[0030] Based on the clearly defined effective locations at the point level, step S1 makes window-level availability determinations for vibration time series, temperature time series, and load time series. The window-level availability of the vibration time series is jointly determined by the longest continuous invalid vibration duration and the effective vibration coverage duration. The longest continuous invalid vibration duration represents the longest duration of consecutive occurrences of invalid vibration locations within the preset monitoring window, and the effective vibration coverage duration represents the total time corresponding to all effective vibration locations within the preset monitoring window. When the longest continuous invalid vibration duration does not exceed the continuous invalid vibration determination duration, and the effective vibration coverage duration is not less than the minimum effective vibration coverage duration, the vibration window-level availability is recorded as available; otherwise, it is recorded as unavailable. The continuous invalid vibration determination duration is determined by the continuous gap duration sample in the fault-free communication sample. The determination process is as follows: sort the continuous gap duration samples from smallest to largest, and take the duration at the last preset percentage position, such as 5%, as the continuous invalid vibration determination duration. The minimum effective coverage time for vibration is obtained by subtracting the continuous invalidation judgment time for vibration from the preset monitoring window length. The window-level availability status of temperature time series and load time series is determined using the same structure, that is, the longest continuous invalidation time of temperature and the effective coverage time of temperature, and the longest continuous invalidation time of load and the effective coverage time of load are statistically analyzed, and then compared with the continuous invalidation judgment time of temperature, the minimum effective coverage time of temperature, the continuous invalidation judgment time of load, and the minimum effective coverage time of load, respectively. The continuous invalidation judgment time of temperature is determined by the sample of the duration of fault-free temperature out-of-bounds, and the continuous invalidation judgment time of load is determined by the sample of the duration of fault-free load sudden change. Both are determined using the same sorting and digitization process. The minimum effective coverage time of temperature and the minimum effective coverage time of load are obtained by subtracting the corresponding continuous invalidation judgment time from the preset monitoring window length, respectively. Thus, the window-level availability status of vibration, the window-level availability status of temperature, and the window-level availability status of load are obtained.

[0031] S104. The three-dimensional parameter availability encoding is generated and the output of step S1 is fixed.

[0032] After the available states at the vibration window level, temperature window level, and load window level are all determined, step S1 generates a three-dimensional availability code according to a fixed order of vibration, temperature, and load. The first digit corresponds to the available state at the vibration window level, the second to the available state at the temperature window level, and the third to the available state at the load window level. A value of one in the three-dimensional availability code indicates that the corresponding parameter is available for subset state projection in the second step within the current preset monitoring window, while a value of zero indicates that the corresponding parameter does not participate in subset state projection within the current preset monitoring window. Simultaneously, the dimensionless vibration sequence and dimensionless temperature... The degree sequence and dimensionless load sequence continue to maintain the same time index order, and invalid positions retain blank markers without any numerical completion. This allows the second step to first identify the current observable parameter combination based on the three-dimensional parameter availability code, and then read the corresponding available parameter time series from the dimensionless vibration sequence, dimensionless temperature sequence, and dimensionless load sequence. Thus, the output objects of step S1 are fixed as the dimensionless vibration sequence, dimensionless temperature sequence, dimensionless load sequence, and three-dimensional parameter availability code. The subsequent second, third, and fourth steps all use the aforementioned names, and the terminology mapping remains unique throughout the entire implementation process.

[0033] After step S1 is completed, the original vibration samples, original temperature samples, and original load samples are converted into vibration time series, temperature time series, and load time series under a unified time reference. Furthermore, dimensionless vibration sequences, dimensionless temperature sequences, dimensionless load sequences, and three-dimensional parameter availability codes are formed, which clearly defines which parameters within the current monitoring window can be used for subsequent analysis. The input objects thus possess consistency, verifiability, and traceability, and the impact of parameter anomalies on subsequent processing is limited to the level of interpretable monitoring conditions.

[0034] Step S2 does not deal with a complete and stable multi-parameter input, but with the actual observable parameter combination filtered by step S1. Under this condition, if different monitoring window data are still processed with a fixed input structure, the common health state vector will drift with the change of parameter combination and cannot be used as a unified judgment basis. Therefore, step S2 needs to select the corresponding subset state projection path based on the three-dimensional parameter availability code and compress the states formed under different parameter combinations into the health semantic space constrained by the same state prototype set.

[0035] S201. Subset observation sequence generation.

[0036] Step S2 first reads the three-dimensional parameter availability code output in step S1 and interprets it sequentially as a vibration availability flag, a temperature availability flag, and a load availability flag. The vibration availability flag corresponds to the window-level availability status of the dimensionless vibration sequence, the temperature availability flag corresponds to the window-level availability status of the dimensionless temperature sequence, and the load availability flag corresponds to the window-level availability status of the dimensionless load sequence. Then, according to the unified time index order fixed in step S1, the dimensionless vibration sequence, dimensionless temperature sequence, and dimensionless load sequence within the current monitoring window are scanned time-by-time. At each unified time index position, the window-level availability status of the corresponding parameter is checked first, and then it is checked whether the corresponding position is a valid observation position. Only when the window-level availability status of the corresponding parameter is available and the corresponding position is a valid observation position, is the observation value at that position written into the subset observation sequence.

[0037] Each observation unit in the subset observation sequence consists of three parts: the first part is the current valid observation value, the second part is the parameter category to which the current observation belongs, and the third part is the relative time position of the current observation in the preset monitoring window. The relative time position is obtained by dividing the time difference between the current unified time index and the start time of the preset monitoring window by the preset monitoring window length, thereby mapping all valid observations uniformly to a time position range between zero and one.

[0038] The arrangement rule of the subset observation sequence is fixed as follows: first sort it from front to back according to a unified time index, and then sort it in a fixed order of vibration, temperature and load within the same time position. This allows the subset observation sequence to retain both the temporal sequence relationship and the parameter source relationship. For parameters whose three-dimensional availability code is unavailable, none of their observation positions are written into the subset observation sequence, and no numerical completion, interpolation substitution or missing estimation is performed.

[0039] To ensure consistent observation organization between the training and inference phases, step S2 constructs a complete observation sequence corresponding to a scenario where vibration, temperature, and load are all in an available state. During construction, all valid observations are written in a uniform time index order and a fixed order for vibration, temperature, and load, for subsequent state prototype set establishment and training with consistent constraints. Through this processing, the dimensionless vibration sequence, dimensionless temperature sequence, dimensionless load sequence, and three-dimensional parameter availability code output in step S1 are converted into a subset observation sequence that can directly enter the subset state projector. Therefore, the formation of the subsequent common health state vector is based on the actual combination of observable parameters, rather than on the supplemented speculative observations.

[0040] S202. Subset state projection execution and public health state vector formation.

[0041] After the subset observation sequence has been formed, step S2 first determines whether at least one of the three-dimensional parameter availability codes is 1. If so, a unique subset state projector is selected based on the three-dimensional parameter availability codes. The non-zero combinations of vibration, temperature, and load parameters are pre-set to correspond to each subset state projector, and each subset state projector processes only one fixed combination of observable parameters. If all three-dimensional parameter availability codes are zero, the current monitoring window is recorded as a completely unavailable monitoring window, and no subset state projection result is generated for the current monitoring window. After verification, the common health state vector, nearest state prototype number, prototype distance, and prototype neighborhood radius use the corresponding values ​​written back to the cache from the previous monitoring window, and the process directly proceeds to step S4 for continuity processing. For the case where at least one parameter is 1, the subset state projector internally includes an observation coding layer, a bidirectional gated cyclic unit layer, and a geometric median convergence layer.

[0042] The observation coding layer performs structured coding on each observation unit. Specifically, the observation value, parameter category, and relative time position are written into the observation coding vector in a fixed order. The parameter category is represented by a three-bit category indicator set in a fixed order of vibration, temperature, and load. When it belongs to vibration, it is written as 1, 0, 0; when it belongs to temperature, it is written as 0, 1, 0; and when it belongs to load, it is written as 0, 0, 1. In this way, each observation coding vector carries amplitude information, time information, and parameter identity information at the same time.

[0043] The bidirectional gated cyclic unit layer processes all observation coding vectors according to the order of the subset observation sequence, and outputs a hidden state sequence corresponding to each observation unit. Each hidden state vector in the hidden state sequence represents the degradation characterization of the corresponding observation unit after the interaction of the current monitoring window and the preceding and following temporal contexts. For example, in the current monitoring window, the three-dimensional parameter availability code indicates that the vibration parameter and load parameter are available, while the temperature parameter is unavailable. Therefore, according to the unified time index order, the effective vibration observations and effective load observations at each time position are first written into the subset observation sequence in sequence, and corresponding observation coding vectors are formed respectively. The bidirectional gated cyclic unit layer then reads in the observation coding vectors one by one according to the order of the subset observation sequence. In the forward path, the influence of the preceding observation on the current position is accumulated in the order of time. In the reverse path, the influence of the subsequent observation on the current position is accumulated in the reverse order. Then, the forward state and the reverse state corresponding to the same observation coding vector are concatenated, and the hidden state vector at that position is obtained through linear transformation. Thus, vibration observations at a certain time location not only retain their own amplitude and parameter category information, but also introduce the contextual constraints of previous load changes and subsequent vibration evolution on that location, so that all hidden state vectors together form a temporal state sequence that can characterize the degradation process of the current monitoring window.

[0044] After the hidden state sequence is generated, a geometric median convergence method is used to form a pre-common health state vector. Specifically, a candidate vector is searched in the state space such that the sum of its Euclidean distances to all hidden state vectors is minimized. The candidate vector obtained at this point is the pre-common health state vector. Subsequently, the pre-common health state vector is scaled and normalized according to its own Euclidean norm to form the common health state vector. If the Euclidean norm of the pre-common health state vector is lower than the normalization lower bound, the normalization lower bound is used instead of the Euclidean norm for normalization. The process of determining the normalization lower bound is as follows: zero-value samples are removed from the Euclidean norm samples of all pre-common health state vectors during the training phase, and then the samples are sorted in ascending order. The Euclidean norm of the sample in the top 5% (e.g., 5%) is taken as the normalization lower bound.

[0045] The reason for using geometric median convergence is that local observation units may exhibit discrete shifts under abnormal sensor conditions. Geometric median convergence is less sensitive to small shifts in observations than mean convergence, making it more suitable for compressing latent state sequences with different combinations of observable parameters into a unified and stable common health state vector. After this sub-step, step S2 directly obtains a common health state vector of a unified dimension from the subset observation sequence and fully preserves the observable conditions given in step S1 in the mapping result, laying the foundation for subsequent alignment of the state prototype set.

[0046] S203. Establish the set of state prototypes and train consistent constraints with the prototypes.

[0047] To ensure that the public health state vectors formed under different three-dimensional parameter availability encoding conditions have a unified degradation semantics, step S2 pre-establishes a set of state prototypes during the training phase. The set of state prototypes is arranged in the order of degradation stages, with each degradation stage corresponding to one state prototype. The process of determining the total number of degradation stages is as follows: first, based on the remaining lifespan labels in the lifespan test samples, the entire lifespan process is divided into several candidate stages; then, the maximum number of stages that can simultaneously satisfy two conditions is selected from the candidate stages. The first condition is that the number of training samples contained in each degradation stage is not less than the preset minimum stage sample size; the second condition is that the remaining lifespan intervals of adjacent degradation stages do not overlap.

[0048] For any training sample, first read the complete observation sequence, then read the total lifetime and current remaining lifetime label of the corresponding lifetime test unit. Based on the proportion of the current remaining lifetime to the total lifetime, map the training sample to a unique target degradation stage number. The mapping logic is: the closer the remaining lifetime proportion is to the initial state, the earlier the target degradation stage number; the closer the remaining lifetime proportion is to the failure state, the later the target degradation stage number. The initialization process of the state prototype set is as follows: input all complete observation sequences belonging to the same target degradation stage number into the subset state projector under complete parameter conditions to obtain the corresponding common health state vector. Then, within the same degradation stage, calculate the geometric median vector of all common health state vectors and define this geometric median vector as the state prototype corresponding to that degradation stage.

[0049] After the state prototype set is initialized, prototype consistency constraint training needs to be implemented. Specifically, starting from the complete observation sequence, all observation units that do not participate in the current observable parameter combination are deleted according to the training encoding to form a masked observation sequence. Then, the common health state vector corresponding to the complete observation sequence and the common health state vector corresponding to the masked observation sequence are calculated separately. The training loss consists of three parts: the first part is the first distance from the common health state vector corresponding to the complete observation sequence to the target state prototype; the second part is the first distance from the common health state vector corresponding to the masked observation sequence to the same target state prototype; and the third part is the first distance between the common health state vector corresponding to the complete observation sequence and the common health state vector corresponding to the masked observation sequence. The first distance is calculated by summing the absolute differences of each dimension.

[0050] The first two parts simultaneously bring together the complete observation conditions and subset observation conditions into the neighborhood of the state prototypes at the same degradation stage. The third part further compresses the representational differences between the complete and subset observation conditions. Meanwhile, step S2 introduces a prototype order constraint, imposing a minimum interval requirement on state prototypes at adjacent degradation stages. The prototype order interval is determined by calculating the first distance between the geometric midpoint vectors of complete observation samples at adjacent degradation stages, sorting them from smallest to largest, and taking the distance value in the first quarter as the prototype order interval. Through prototype consistency constraint training and prototype order constraint training, the set of state prototypes no longer merely corresponds to the state clustering centers under complete parameter conditions, but becomes a shared degradation semantic reference for various combinations of observable parameters.

[0051] S204. Generation of the nearest state prototype number, prototype distance, and prototype neighborhood radius.

[0052] After entering the inference phase, step S2, after obtaining the public health state vector of the current monitoring window, further calculates the nearest state prototype number, prototype distance, and prototype neighborhood radius, so that the third step can perform fault mode category discrimination, monotonic degradation index generation, and remaining lifetime interval generation, and also facilitates the fourth step to update the prototype matching boundary when the three-dimensional parameter availability encoding changes.

[0053] The most recent state prototype index is generated by calculating the first distance from the common health state vector to each state prototype in the state prototype set, and taking the index of the state prototype with the smallest first distance as the most recent state prototype index. This smallest first distance is also recorded as the prototype distance. The prototype neighborhood radius consists of two parts: the first part is the basic prototype neighborhood radius corresponding to the degradation stage under complete parameter conditions, and the second part is the neighborhood expansion introduced by the reduction of observable parameters.

[0054] The process for determining the basic prototype neighborhood radius is as follows: During the training phase, collect the common health state vectors corresponding to all complete observation sequences belonging to the same degradation stage. Then, calculate the first distance from these common health state vectors to the state prototype of that degradation stage. Sort all distance samples in ascending order, and take the distance value located at the last preset percentage, such as 90%, as the basic prototype neighborhood radius of that degradation stage. The process for determining the neighborhood expansion is as follows: During the training phase, construct complete observation sequences and masked observation sequences for the same training sample, calculate the first distance from both to the target state prototype, and then calculate the distance difference between them. Sort all distance difference samples in ascending order, and take the distance difference located at the last preset percentage, such as 5%, as the prototype distance increment caused by a single unavailable parameter. The prototype neighborhood radius of the current monitoring window is jointly determined by the basic prototype neighborhood radius and the prototype distance increment. The more unavailable parameters in the three-dimensional availability code, the larger the prototype neighborhood radius. The specific generation logic is as follows: first, count the number of unavailable code bits in the three-dimensional availability code, then multiply this number by the prototype distance increment, and then add the product to the basic prototype neighborhood radius. The resulting prototype neighborhood radius reflects both the normal range of characterization fluctuations in the current degradation stage under complete parameter conditions and the degree of characterization expansion of the current observable parameter combination relative to complete parameter conditions.

[0055] After completing the above calculations, step S2 outputs the common health state vector to step S3, and simultaneously retains the nearest state prototype number, prototype distance, and prototype neighborhood radius to step S4, so that subsequent fault mode classification, monotonic degradation index generation, remaining lifetime interval generation, and prototype matching boundary update are all based on the same set of state semantic benchmarks.

[0056] After step S2 is completed, the available parameter time series corresponding to the current monitoring window is mapped to a common health state vector of a unified dimension. At the same time, the nearest state prototype number, prototype distance and prototype neighborhood radius are generated, so that the state representation under different parameter combinations can fall into the same degradation semantic framework for comparison. This process distinguishes the representation drift caused by parameter missing from the actual degradation of the equipment, so that the health state representation has a consistent interpretation basis across parameter combinations.

[0057] Step S3 builds upon the unified health semantic space results obtained in step S2. At this point, it is necessary to convert the public health state vector into a discrimination result for health management decision-making. However, under abnormal sensor conditions, simply providing a category label or a single lifetime value cannot fully reflect the combined influence of the current degree of degradation and observable conditions. Therefore, step S3 must synchronously generate fault mode categories, monotonic degradation index, and remaining lifetime intervals around the public health state vector, state prototype set, and three-dimensional parameter availability code.

[0058] After step S3 is completed, the public health state vector is further transformed into failure mode category, monotonic degradation index and remaining lifetime interval, forming a joint discrimination result corresponding to the current monitoring window. This allows the diagnosis result, degradation location and lifetime range to be based on the same state basis. The remaining lifetime interval not only reflects the degree of degradation, but also reflects the impact of the current observable conditions on the credibility range of the result. Therefore, the output result has a more complete state interpretation capability.

[0059] S301. Establishment of fault mode category discrimination criteria and generation of current fault mode category.

[0060] Step S3 first reads the common health state vector and the most recent state prototype number output in step S2, and pre-establishes a set of fault mode reference centers during the training phase. Each fault mode reference center in the set is determined by the fault mode category and the degradation stage. That is, for each fault mode category, a dedicated fault mode reference center is established within each degradation stage. The process of establishing the fault mode reference center is as follows: first, the training samples are divided into different fault mode categories according to the real fault labels; then, the training samples within the same fault mode category are divided into different degradation stage subsets according to the target degradation stage number; then, in the intersection subset of each fault mode category and degradation stage, the geometric median vector of all common health state vectors is obtained, and the obtained geometric median vector is defined as the corresponding fault mode reference center. When the current monitoring window enters the fault mode category discrimination stage, the degradation stage corresponding to the most recent state prototype index is first taken as the current degradation stage. Then, the common health state vector is compared one by one with all fault mode reference centers within the current degradation stage. The comparison method is to sum the absolute differences between the common health state vector and each fault mode reference center in each dimension to obtain the first distance from the common health state vector to each fault mode reference center. Then, the fault mode category with the smallest first distance is selected as the fault mode category of the current monitoring window. If multiple fault mode categories have the same first distance, the fault mode category with the highest ranking is selected as the current fault mode category according to the pre-fixed fault mode category order. Through the above processing, the fault mode category discrimination is not directly based on the nearest neighbor result of a single sample, but is based on the robust centers of samples of the same type in the unified health semantic space. Furthermore, the most recent state prototype index here limits the range of degradation stages where the discrimination occurs, so that the fault mode category discrimination is always carried out within the corresponding degradation stage, without introducing cross-stage misjudgment.

[0061] S302. Generation of the monotonic degradation index.

[0062] After the fault mode categories are obtained, step S3 continues to generate a monotonic degradation index. The monotonic degradation index is used to characterize the relative position of the current monitoring window in the complete degradation process and maintains a fixed direction of evolution throughout the lifetime. To this end, the set of state prototypes that has been fixed in the training phase of step S2 is read first, and the prototype segment length between adjacent state prototypes is calculated in the order of degradation stages. The prototype segment length is calculated by summing the absolute differences of two adjacent state prototypes in each dimension. Considering that adjacent state prototypes may experience distance compression in local regions, step S3 also sets a lower bound for the prototype segment length. The process of determining the lower bound for the prototype segment length is as follows: all non-zero adjacent prototype segment lengths are sorted in ascending order, and the prototype segment length in the top 5% is taken as the lower bound for the prototype segment length. When a prototype segment length is lower than the lower bound for the prototype segment length, the lower bound for the prototype segment length is used instead of the corresponding prototype segment length. Subsequently, the lengths of all prototype segments from the first state prototype to the nearest state prototype number are accumulated to obtain the prefix segment length, and the lengths of all prototype segments are accumulated to obtain the total segment length. If the nearest state prototype number is not the last degradation stage, the local projection ratio of the common health state vector in the direction of the prototype connection of the current degradation stage is calculated. The local projection ratio is calculated by first calculating the vector of the common health state vector relative to the current state prototype, then calculating the vector of the next state prototype relative to the current state prototype, then taking the inner product of the two, and dividing by the sum of the square length of the vector of the next state prototype relative to the current state prototype and the lower bound of the projection denominator. The process of determining the lower bound of the projection denominator is as follows: sort all adjacent state prototypes by their L2 distance squares in ascending order, and take the square value in the top 5% as the lower bound of the projection denominator. If the local projection ratio is less than zero, it is zero; if it is greater than one, it is one. If the nearest state prototype number is already in the last degradation stage, the local projection ratio is directly taken as one. Finally, the prefix segment length and the result of multiplying the local projection ratio by the current prototype segment length are added together and then divided by the total segment length to obtain the pre-monotonic degradation index. During the online inference phase, if the degradation direction of the target equipment is predefined as monotonically increasing, the larger of the monotonic degradation index of the previous monitoring window and the pre-monotonic degradation index is taken as the monotonic degradation index of the current monitoring window. If the degradation direction of the target equipment is predefined as monotonically decreasing, the smaller of the two is taken as the monotonic degradation index of the current monitoring window. Through this processing, the monotonic degradation index has both the meaning of continuous position on the state prototype link and the constraint characteristic of unidirectional change along the monitoring time sequence, so it can be directly used as the continuous degradation coordinates for the subsequent remaining lifetime interval generation.

[0063] S303. Remaining lifetime interval center generation.

[0064] After obtaining the fault mode category and monotonic degradation index, step S3 continues to generate the center of the remaining lifetime interval. To ensure that the center of the remaining lifetime interval can reflect the lifetime differences under different fault mode categories, step S3 establishes a fault mode stage lifetime table for each fault mode category and each degradation stage during the training phase. The fault mode stage lifetime table includes at least the median remaining lifetime of the stage, the lower reference lifetime of the stage, and the upper reference lifetime of the stage. Specifically, the training samples are first divided into different fault mode categories according to the real fault labels, and then divided into different degradation stage subsets according to the target degradation stage number. Then, all remaining lifetime labels within the same fault mode category and the same degradation stage are sorted from smallest to largest. The remaining lifetime value in the middle position is taken as the median remaining lifetime of the stage, the remaining lifetime value in the first 10% position is taken as the lower reference lifetime of the stage, and the remaining lifetime value in the last 90% position is taken as the upper reference lifetime of the stage. The center of the remaining lifetime interval of the current monitoring window is jointly determined by the current fault mode category, the most recent state prototype number, and the local projection ratio. If the most recent state prototype number is not the last degradation stage, interpolation is performed between the median remaining lifetimes of two adjacent degradation stages corresponding to the current fault mode category, according to the local projection ratio. The closer the local projection ratio is to zero, the closer the center of the remaining lifetime interval is to the median remaining lifetime of the current degradation stage; the closer the local projection ratio is to one, the closer the center of the remaining lifetime interval is to the median remaining lifetime of the next degradation stage. If the most recent state prototype number is already in the last degradation stage, the median remaining lifetime corresponding to the last degradation stage is directly used as the center of the remaining lifetime interval. Thus, the center of the remaining lifetime interval no longer depends solely on the discrete numbering of the most recent state prototype number, but is obtained through continuous position correction within the degradation stage, thereby maintaining smooth changes between adjacent monitoring windows.

[0065] S304. Remaining lifetime range half-width generation.

[0066] After the center of the remaining lifetime interval has been determined, step S3 continues to generate the half-width of the remaining lifetime interval. The half-width of the remaining lifetime interval is used to characterize the uncertainty range of the current monitoring window lifetime estimate. Its size is simultaneously affected by the degree of lifetime dispersion within the degradation stage, the deviation of the public health state vector from the state prototype, and the changes in observable conditions reflected by the three-dimensional parameter availability encoding. Step S3 first reads the lower-stage reference lifetime and upper-stage reference lifetime corresponding to the current failure mode category and the most recent state prototype number, and defines half of the difference between the two as the basic half-width of the failure mode stage. The basic half-width of the failure mode stage reflects the natural dispersion range of lifetime labels within the same failure mode category and the same degradation stage under complete parameter conditions. Subsequently, step S3 also calculates the failure mode stage mapping coefficient. The failure mode stage mapping coefficient is used to convert the excess prototype distance in the state space into an expansion amount on the lifetime unit. It is determined by taking the absolute value of the difference in the median remaining lifetime between adjacent degradation stages for the current failure mode category, and then dividing it by the prototype segment length between the aforementioned adjacent state prototypes; if the most recent state prototype number is located in the last degradation stage, the failure mode stage mapping coefficient corresponding to the previous degradation stage is used. The prototype distance excess is calculated by subtracting the prototype neighborhood radius from the prototype distance output in step S2. If the difference is negative, it is treated as zero; if the difference is positive, the original value is retained. Next, step S3 uses the three-dimensional parameter availability code to count the number of unavailable parameters in the current monitoring window. The more unavailable parameters, the weaker the current observable conditions. Therefore, a fault mode stage expansion step size needs to be introduced. The process for determining the fault mode stage expansion step size is as follows: during the training phase, complete observation samples and masked observation samples belonging to the same fault mode category and the same degradation stage are selected. The center of their remaining lifetime intervals is calculated separately. The absolute values ​​of their differences form a sample sequence, which is then sorted from smallest to largest. The difference at the last 75% position is taken as the fault mode stage expansion step size. The half-width of the remaining lifetime interval of the current monitoring window is obtained by adding three parts: the first part is the base half-width of the fault mode stage; the second part is the boundary deviation expansion amount obtained by multiplying the fault mode stage mapping coefficient by the prototype distance excess; and the third part is the perception condition expansion amount obtained by multiplying the number of unavailable parameters by the fault mode stage expansion step size. With this structure, the half-width of the remaining lifetime interval is not determined by a fixed empirical constant, but is generated by the verifiable stage discrepancy patterns, prototype deviation patterns, and parameter missing patterns in the training data.

[0067] S305. Generation of remaining lifetime intervals and processing of joint discrimination results.

[0068] After the center of the remaining lifetime interval and the half-width of the remaining lifetime interval have been obtained, step S3 continues to generate the lower bound and the upper bound of the remaining lifetime interval. The lower bound of the remaining lifetime interval is obtained by subtracting the half-width of the remaining lifetime interval from the center of the remaining lifetime interval. When the calculation result is less than zero, the lower bound of the remaining lifetime interval is zero. The upper bound of the remaining lifetime interval is obtained by adding the half-width of the remaining lifetime interval to the center of the remaining lifetime interval. After calculating the lower and upper bounds of the remaining lifetime interval, step S3 combines the fault mode category, monotonic degradation index, lower and upper bounds of the remaining lifetime interval, nearest state prototype number, prototype distance, and prototype neighborhood radius into a joint discrimination result for the current monitoring window. The fault mode category indicates the closest fault mode, the monotonic degradation index indicates the current degradation position, and the lower and upper bounds of the remaining lifetime interval jointly characterize the current lifetime prediction range. Simultaneously, step S3 generates boundary deviation markers based on the relationship between the prototype distance and the prototype neighborhood radius. If the prototype distance does not exceed the prototype neighborhood radius, the boundary deviation marker is marked as normal; if the prototype distance exceeds the prototype neighborhood radius, the boundary deviation marker is marked as deviation. The boundary deviation markers in step S3 do not change the obtained fault mode category and monotonic degradation index but serve as an important basis for updating the prototype matching boundary in step S4. After this process, step S3 transforms the unified health semantic space result provided in step S2 into a monitoring result sequence that can be directly output and continuously invoked by step S4.

[0069] Step S4 addresses the aspect most prone to result fluctuations under continuous monitoring conditions: how to maintain the continuity of the common health state vector and joint discrimination results when the three-dimensional availability code changes between adjacent monitoring windows. Without a dedicated switching confirmation, state verification, and boundary update process, the active subset state projector will switch frequently due to short-term fluctuations, and the fault mode category, monotonic degradation index, and remaining lifetime range will also experience unnecessarily sudden changes. Therefore, step S4 requires conditional and controlled processing of the code changes.

[0070] S401. Encoding change confirmation and active subset state projector update determination.

[0071] Step S4 first reads the three-dimensional availability code, verified common health state vector, prototype distance, and active subset state projector identifier already written to the cache in the previous monitoring window. Simultaneously, it reads the three-dimensional availability code, common health state vector, nearest state prototype number, prototype distance, prototype neighborhood radius, and boundary deviation marker generated in steps S1, S2, and S3 for the current monitoring window. If the three-dimensional availability code of the current monitoring window is completely consistent with that of the previous monitoring window, the active subset state projector remains unchanged, and step S4 proceeds directly to subsequent result verification. If they are inconsistent, it first determines whether the current three-dimensional availability code is continuously changing.

[0072] The process for determining continuous changes is as follows: Step S4 accumulates the repeated occurrence length of the current three-dimensional availability code in the continuous monitoring window to form the code change duration window number, and compares the code change duration window number with the continuous occurrence window number threshold. The logic for determining the continuous occurrence window number threshold is as follows: extract all short-term code fluctuation segments in the fault-free operation sample, count the duration window number of each fluctuation segment, sort them from smallest to largest, and take the duration window number located at the last preset percentage, such as 95%, as the continuous occurrence window number threshold. At the same time, Step S4 also calculates the first distance between the current public health state vector and the previous verified public health state vector to form a state offset, and compares the state offset with the state continuity threshold. The logic for determining the state continuity threshold is as follows: during the training phase, select adjacent monitoring windows in the same life test unit where the three-dimensional availability code does not change, calculate the first distance between the corresponding public health state vectors, sort them from smallest to largest, and take the first distance located at the last preset percentage, such as 90%, as the state continuity threshold. Step S4 only recognizes the current three-parameter availability code as a stable change code and determines the subset state projector uniquely corresponding to the current three-parameter availability code as the target subset state projector when the number of sustained window periods of the code change is not less than the threshold for consecutive window periods and the state offset is not higher than the state continuity threshold. If either condition is not met, the active subset state projector remains the subset state projector corresponding to the previous monitoring window, and the active state remains unchanged. Through this process, the switching of the three-parameter availability code is no longer determined by the instantaneous fluctuations of a single monitoring window, but by the combined constraints of persistence and continuity, thereby distinguishing between short-term sampling disturbances and changes in stable observable conditions.

[0073] S402. Generation of public health status vector after verification.

[0074] After the target subset state projector has been confirmed, step S4 requires a post-switch verification of the current public health state vector to ensure that it remains within the continuous semantic range of the same state prototype set before and after the switch. Step S4 first reads the nearest state prototype number of the previous monitoring window and the nearest state prototype number of the current monitoring window. If they match, the current public health state vector is directly used as the verified public health state vector and written into subsequent processing. If they do not match, the distance difference relative to the nearest state prototype of the previous monitoring window is calculated. The distance difference is calculated by first finding the first distance from the current public health state vector to the nearest state prototype of the previous monitoring window, then finding the first distance from the verified public health state vector to the same state prototype, and finally taking the absolute value of the difference as the distance difference. Step S4 then compares the distance difference with a distance difference threshold. The logic for determining the distance difference threshold is as follows: during the training phase, select the complete observation sequence and the masked observation sequence corresponding to the same lifetime test sample, calculate the first distance from both to the same target state prototype, sort the absolute values ​​of the differences between the two in ascending order, and take the distance difference at the last preset percentage, such as 90%, as the distance difference threshold. If the distance difference is not higher than the distance difference threshold, the current public health state vector is directly used as the verified public health state vector; if the distance difference is higher than the distance difference threshold, step S4 performs coordinate median alignment.

[0075] The meaning of coordinate median alignment is to take the median of three scalars for each dimension of the public health state vector. These three scalars are, in order, the component of the previously verified public health state vector in that dimension, the component of the current public health state vector in that dimension, and the component of the most recent state prototype from the previous monitoring window in that dimension. This process is used to construct a new verified public health state vector dimension by dimension. The reason for using coordinate median alignment is that, for the objective function of summing absolute deviations by dimension, taking the median of each dimension directly yields the vector with the minimum sum of absolute deviations. Therefore, it has both clear mathematical meaning and a directly implementable solution method. Through this process, the verified public health state vector maintains consistency with the current monitoring window's public health state vector while being jointly constrained by the previously verified public health state vector and the most recent state prototype from the previous monitoring window, thus keeping state transitions during the switching process within an interpretable range.

[0076] S403. Prototype matching bounded update.

[0077] After the public health status vector is obtained after verification, step S4 continues to update the prototype matching boundary corresponding to the current monitoring window. The update object is the prototype neighborhood radius corresponding to the current most recent state prototype number and the current three-dimensional parameter availability code.

[0078] Step S4 first calculates the first distance between the verified public health state vector and the previous verified public health state vector, forming the verified state offset; then, it calculates the first distance from the verified public health state vector to the current nearest state prototype, forming the verified prototype distance; finally, it compares the verified prototype distance with the prototype distance recorded in the previous monitoring window to obtain the prototype distance change. The boundary correction increment is the larger of the verified state offset and the prototype distance change, and its physical meaning is the maximum offset caused by the switching event in the unified health semantic space and the prototype space.

[0079] To limit the cumulative amplification of the boundary correction increment, step S4 pre-establishes a boundary correction upper limit. The logic for determining the boundary correction upper limit is as follows: during the training phase, the first distance difference between the complete observation sequence and the masked observation sequence relative to the same state prototype is collected. All distance difference samples are sorted from smallest to largest, and the distance difference located at the last preset percentage, such as 95%, is taken as the boundary correction upper limit. Simultaneously, step S4 also establishes a boundary upper bound radius for each degradation stage and each three-dimensional parameter availability code. The logic for determining the boundary upper bound radius is as follows: during the training phase, the first distance from all common health state vectors belonging to the same degradation stage and the same three-dimensional parameter availability code to the corresponding state prototype is collected, sorted from smallest to largest, and the first distance located at the last preset percentage, such as 95%, is taken as the boundary upper bound radius for that degradation stage and that three-dimensional parameter availability code. The updated prototype neighborhood radius of the current monitoring window is constrained by three parts: first, the prototype neighborhood radius output in step S2 is used as the base radius; second, the smaller value between the boundary correction increment and the boundary correction upper limit is superimposed on the base radius; finally, the superposition result is compared with the boundary upper bound radius, and the smaller value is taken as the updated prototype neighborhood radius. If the current boundary deviation is marked as normal, and the verified prototype distance has fallen back into the range of the basic prototype neighborhood radius, then step S4 directly reverts the updated prototype neighborhood radius to the basic radius, no longer retaining the previous expansion. Through this process, the prototype matching boundary can absorb the necessary offset brought about by stable switching, without continuously amplifying during continuous switching and losing its discriminative ability.

[0080] S404. Continuity verification of joint results after switching.

[0081] After the prototype matching boundary update is completed, step S4 performs a post-switch continuity check on the fault mode category, monotonic degradation index, and remaining lifetime interval given in step S3. Step S4 first compares the checked prototype distance with the updated prototype neighborhood radius. If the checked prototype distance does not exceed the updated prototype neighborhood radius, it indicates that the checked common health state vector is within the current acceptable prototype neighborhood, and the fault mode category, monotonic degradation index, and remaining lifetime interval generated in step S3 are directly used as the final output of the current monitoring window. If the checked prototype distance exceeds the updated prototype neighborhood radius, it indicates that the state after the switch is still outside the boundary region. In this case, step S4 does not re-determine the fault mode category but retains the fault mode category already generated in step S3, and performs continuity correction on the monotonic degradation index and remaining lifetime interval.

[0082] The monotonic degradation index is corrected by interval truncation. The upper and lower limits of the interval are obtained by expanding the monotonic degradation index change threshold to both sides of the final output of the previous monitoring window. The logic for determining the monotonic degradation index change threshold is as follows: during the training phase, the absolute values ​​of the differences in monotonic degradation indices between adjacent monitoring windows of the same life test unit are collected and sorted from smallest to largest. The difference at the last preset percentage, such as 90%, is taken as the monotonic degradation index change threshold. If the monotonic degradation index given in step S3 of the current monitoring window falls within the above interval, it is directly retained. If it exceeds the interval, it is truncated to the nearest interval boundary.

[0083] For the remaining lifetime interval, step S4 further calculates the boundary excess expansion. The boundary excess expansion is generated by first calculating the difference between the verified prototype distance and the updated prototype neighborhood radius, then treating negative values ​​as zero, and finally using the fault mode stage mapping coefficient established in step S3 to convert the above difference into an expansion amount in lifetime units. After obtaining the expansion amount, the lower bound of the remaining lifetime interval given in step S3 is expanded downward by the same amount. If the expansion result is less than zero, it is treated as zero. The upper bound of the remaining lifetime interval given in step S3 is expanded upward by the same amount. Through this continuous verification, the monitoring window outside the boundary will not cause non-physical jumps in the fault mode category, monotonic degradation index, and remaining lifetime interval at the switching point, but will instead smoothly transition according to interpretable correction rules.

[0084] S405. Activity status write-back and next monitoring window initialization.

[0085] After the final failure mode category, final monotonic degradation index, and final remaining lifetime range are all determined, step S4 writes the activity status corresponding to the current monitoring window back to the continuous monitoring cache for initialization of the next monitoring window.

[0086] The write-back content includes at least the current three-dimensional parameter availability code, the verified public health state vector, the current most recent state prototype number, the verified prototype distance, the updated prototype neighborhood radius, the final failure mode category, the final monotonic degradation index, the final remaining lifetime interval, and the current active subset state projector identifier.

[0087] If the current monitoring window already meets the stable change switching conditions, the target subset state projector is written back as the new active subset state projector; if the current monitoring window does not meet the stable change switching conditions, the active subset state projector of the previous monitoring window continues to be used. Through this write-back process, step S4 solidifies all the key state variables formed by code change confirmation, state verification, boundary update, and result verification into the initial conditions of the next monitoring window, enabling the next monitoring window to directly inherit the continuity constraints already completed by the current monitoring window without having to revisit earlier monitoring history. Thus, step S1 defines the current observable conditions, step S2 constructs a unified health semantic space, step S3 generates fault mode categories, monotonic degradation indices, and remaining lifetime intervals, and step S4 maintains the continuity of the above results when the three-dimensional parameter availability code changes. The data objects, processing boundaries, and output semantics between the preceding and following steps remain consistent.

[0088] After step S4 is completed, the three-dimensional parameter availability coding changes are distinguished into stable changes that can trigger switching and short-term disturbances that do not require switching. After verification, the public health state vector, the updated prototype matching boundary, the final failure mode category, the monotonic degradation index, and the remaining lifetime interval are uniformly written back to the continuous monitoring link, so that the result switching between different combinations of observable parameters has continuous rules, and the monitoring output shows a smooth transition rather than a sudden jump in the coding change scenario.

[0089] Specifically, the above are merely preferred embodiments of this application and are not intended to limit this application.

[0090] The percentile parameters, threshold values, and number of stages mentioned in the instruction manual can be determined based on offline training samples, calibration samples, or on-site operating procedures.

[0091] In the description of this specification, references to terms such as "an embodiment," "example," and "specific example" indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0092] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A deep learning-based equipment PHM fault mode recognition and prediction method, characterized in that, Including the following steps: S1. Collect vibration time series, temperature time series and load time series within the preset monitoring window, perform unified time reference alignment, fixed-length window processing and amplitude standardization, and generate three-dimensional parameter availability codes arranged by vibration, temperature and load based on missing, out-of-bounds and abrupt change states; S2. When at least one bit of the three-dimensional parameter availability code is one, the corresponding subset state projector is called according to the three-dimensional parameter availability code to map the available parameters in the vibration time series, temperature time series and load time series to a common health state vector, and to make the complete parameter sample and the corresponding subset sample converge to the same state prototype neighborhood in the state prototype set. S3. Based on the public health state vector, three-dimensional availability code, and state prototype set, generate failure mode category, monotonic degradation index, and remaining life interval; S4. When the availability codes of the three-dimensional parameters of consecutive adjacent monitoring windows change and the switching confirmation conditions are met, switch the subset state projector, update the prototype matching boundary according to the change in distance between the common health state vector and the state prototype before and after the switch, and output the verified fault mode category, monotonic degradation index and remaining lifetime range.

2. The equipment PHM fault mode recognition and prediction method based on deep learning according to claim 1, characterized in that, Step S1 includes: Unified time reference alignment includes generating a unified time index according to the preset monitoring window length and unified sampling interval, and completing the alignment and writing of vibration time series, temperature time series and load time series based on the time deviation between each original sampling time and the corresponding unified time index.

3. The equipment PHM fault mode recognition and prediction method based on deep learning according to claim 2, characterized in that, Step S1 also includes: The out-of-bounds state of the temperature time series is determined based on the temperature safety boundary corresponding to the load, and the sudden change state of the load time series is determined by the differential amplitude of adjacent sampling points and the continuous duration, and a three-dimensional availability code is generated accordingly.

4. The equipment PHM fault mode recognition and prediction method based on deep learning according to claim 1, characterized in that, Step S2 includes: Available parameters are arranged in a uniform time index order and a fixed order of vibration, temperature and load to form a subset observation sequence. The subset state projector corresponds one-to-one with a three-bit parameter availability code that has at least one bit set to one, and receives the corresponding subset observation sequence to perform state projection.

5. The deep learning-based equipment PHM fault mode recognition and prediction method according to claim 4, characterized in that, Step S2 also includes: The set of state prototypes is established in the order of degradation stages. The complete parameter samples form a complete observation sequence according to the corresponding degradation stage. After deleting parameter observations that are inconsistent with the training encoding, the complete observation sequence forms a subset sample, which is constrained to the same state prototype neighborhood together with the complete parameter samples.

6. The deep learning-based equipment PHM fault mode recognition and prediction method according to claim 5, characterized in that, Step S2 also includes: The subset state projector sequentially performs observation coding, temporal state coding, and geometric median convergence on the subset observation sequence to obtain a pre-common health state vector, and then performs scale normalization on the pre-common health state vector to form a common health state vector.

7. The equipment PHM fault mode recognition and prediction method based on deep learning according to claim 1, characterized in that, Step S3 includes: The failure mode category is determined based on the first distance between the reference centers of each failure mode within the degradation stage corresponding to the public health state vector and the current most recent state prototype number. The monotonic degradation index is generated based on the continuous positions of the public health state vector in the direction of the line connecting adjacent state prototypes.

8. The deep learning-based equipment PHM fault mode recognition and prediction method according to claim 7, characterized in that, Step S3 also includes: The remaining lifetime interval is generated based on the lifetime table of the failure mode stage corresponding to the current failure mode category. The center of the interval is determined by the current degradation stage position, and the half-width of the interval is determined by the base half-width of the failure mode stage, the excess prototype distance, and the number of unavailable parameters in the three-dimensional parameter availability code.

9. The equipment PHM fault mode recognition and prediction method based on deep learning according to claim 1, characterized in that, Step S4 includes: The switching confirmation conditions include that the number of continuous windows in the current three-dimensional availability code in the continuous monitoring window is not less than the threshold for the number of consecutive windows, and that the first distance between the current public health state vector and the previously verified public health state vector is not higher than the state continuity threshold.

10. The deep learning-based equipment PHM fault mode recognition and prediction method according to claim 9, characterized in that, Step S4 also includes: After the switching confirmation conditions are met, coordinate median alignment is performed based on the previously verified public health state vector, the current public health state vector, and the most recent state prototype of the previous monitoring window to generate the verified public health state vector, and the prototype matching boundary is updated between the neighborhood radius of the basic prototype and the upper bound radius of the corresponding encoded boundary.

Images