A method and system for detecting abnormal equipment operation data in extremely cold environments

Abstract

This invention discloses a method and system for detecting abnormal equipment operation data in extremely cold environments, belonging to the field of industrial equipment condition monitoring. To solve the problem of false alarms and missed alarms caused by the difficulty in distinguishing between real equipment faults and abnormalities and pseudo-anomalies caused by low temperatures in sensors and acquisition links under extremely cold conditions, this invention generates residual sequences by predicting physical constraints based on temperature and operating conditions, performs online change point detection on the residuals, and adjusts the noise covariance and observation channel enable based on the change point results in a closed loop in a Bayesian state-space model that includes equipment state variables, sensor bias variables, and sensor frozen state variables, performs filtering updates, and assigns alarms, thus achieving the technical effects of pseudo-anomaly suppression, anomaly attribution, and reliable early warning.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment condition monitoring, and in particular to a method and system for detecting abnormal equipment operation data in extremely cold environments. Background Technology

[0002] As wind farms, oil and gas field stations, mining areas, and polar research stations expand to higher latitudes and altitudes, the long-term operation of equipment in extremely cold environments, ranging from -30°C to -50°C, is becoming increasingly common. To ensure the continuous and reliable operation of critical equipment, the industry widely adopts multi-sensor-based condition monitoring systems to collect operational data such as current, vibration, temperature, speed, and power, and uses this data for anomaly detection and fault early warning. Existing anomaly detection technologies have evolved from fixed threshold and rule-based methods, statistical process control and residual analysis methods, to model-based methods based on machine learning and deep learning, as well as gray-box modeling and state-space estimation methods that combine mechanistic models and data-driven models. Simultaneously, the demand for online detection has driven the application of sliding window detection and some online filtering estimation methods in engineering.

[0003] However, the aforementioned existing technologies still have shortcomings under extremely cold working conditions:

[0004] 1. Most methods assume that the sensor and acquisition link are reliable, or only perform simple data cleaning. They are unable to cope with zero drift caused by low temperature, sensitivity changes, output freezing caused by probe frost and ice, intermittent interruptions caused by hardened cables and poor contact, sampling loss caused by power supply attenuation, and timestamp misalignment caused by communication jitter. They are prone to misjudging false anomalies as equipment failures.

[0005] 2. Some data-driven models rely on training under normal temperature or stable operating conditions. When faced with distribution drift caused by strong coupling between temperature and operating conditions, their generalization ability is insufficient. Fixed thresholds are also difficult to adjust adaptively at low temperatures, leading to increased false alarm rates or missed detection of critical faults.

[0006] 3. Existing methods often separate anomaly detection from health diagnosis of the acquisition chain, lacking online identification of the time and type of anomalies and dynamic adjustment mechanisms for model parameters and noise assumptions, making it difficult to achieve reliable attribution of anomaly sources and graded alarms.

[0007] Therefore, there is a need for a method and system for detecting abnormal equipment operation data in extremely cold environments that can overcome the shortcomings of the existing technologies. Summary of the Invention

[0008] One objective of this invention is to propose a method and system for detecting anomalies in equipment operating data in extremely cold environments. Addressing the problems of existing technologies in distinguishing between genuine equipment malfunctions and false anomalies caused by low temperatures in sensors and data acquisition links under extremely cold conditions, leading to false alarms and missed alarms, and the lack of online diagnostic and anomaly attribution mechanisms, the following technical solution is proposed: Multi-channel observation data from multiple sensors are acquired and time-aligned to form an observation sequence; simultaneously, ambient temperature and operating parameters are acquired to form a conditional sequence; the observation sequence and conditional sequence are input into a physical constraint prediction model, and the model parameters are adaptively updated based on the conditional sequence to generate a predicted observation sequence. The process involves calculating residual sequences, performing online change point detection on each observation channel based on these sequences to obtain the change point occurrence time, involved channels, change point type, and statistics. A Bayesian state-space model is established, incorporating equipment state variables, sensor bias variables, and sensor frozen state variables. The process noise covariance matrix and measurement noise covariance matrix are determined based on the conditional sequence. At each sampling time, a closed-loop adjustment is performed on the noise covariance matrix and observation channel enable based on the change point detection results, followed by Bayesian filtering updates. Anomaly attribution is performed by combining posterior estimation results with change point detection results, outputting equipment anomalies or acquisition link anomalies and their anomaly confidence levels. This invention possesses the technical effects of suppressing false anomalies, reducing false alarms and missed alarms, and achieving anomaly source attribution and reliable early warning in extremely cold environments.

[0009] This invention provides a method for detecting abnormal equipment operation data in extremely cold environments, including:

[0010] S1. Acquire multi-channel observation data from multiple sensors, perform time alignment processing to form an observation sequence arranged by sampling time, along with corresponding environmental temperature and operating parameters, to form a conditional sequence; S2. Input the observation sequence and conditional sequence into the physical constraint prediction model, adaptively update the model parameters based on the conditional sequence, and predict the observation sequence based on the updated physical constraint prediction model to obtain the predicted observation sequence. Generate a residual sequence based on the observation sequence and the predicted observation sequence; S3. Perform online change point detection on each observation channel of the multi-channel observation data based on the residual sequence, and output the change point detection results, including the time of change point occurrence, the observation channel involved, the change point type, and the change point statistics; S4. Establish a Bayesian state-space model, including equipment state variables, transmission... The sensor bias variable and sensor frozen state variable are determined according to the condition sequence, and the process noise covariance matrix and measurement noise covariance matrix corresponding to each sampling time are determined. At each sampling time, based on the change point detection result, at least one of the process noise covariance matrix, measurement noise covariance matrix and observation channel enable state is adjusted in a closed loop and then Bayesian filtering is performed to update it. The posterior estimates of the equipment state variable, the posterior estimates of the sensor bias variable and the posterior probability of the sensor frozen state variable are output to obtain the posterior estimation result. S5. Based on the posterior estimation result and the change point detection result, anomaly attribution processing is performed. Based on the preset judgment rule, the anomaly category is determined and alarm information is output. The anomaly categories include equipment anomaly and acquisition link anomaly, and the anomaly confidence level is output for the anomaly category.

[0011] Optionally, S1 includes:

[0012] Collect multi-channel observation data output from the multiple sensors, as well as the corresponding ambient temperature and operating parameters; add timestamps to the multi-channel observation data.

[0013] Based on the preset sampling period, multiple sampling times are determined in chronological order, and the multi-channel observation data is mapped to the multiple sampling times according to the timestamp, so as to form corresponding multi-channel observation data at each sampling time.

[0014] When any observation channel is missing an observation value at any sampling time, a missing value marker is generated for that observation channel;

[0015] The observation sequence is obtained by arranging the multi-channel observation data and missing markers corresponding to each sampling time in chronological order.

[0016] The environmental temperature and operating parameters are arranged in a time sequence consistent with the observation sequence to obtain the condition sequence.

[0017] Optionally, S2 includes:

[0018] The model parameters of the physical constraint prediction model are updated based on the ambient temperature and operating parameters in the condition sequence to match the model parameters with the ambient temperature and operating parameters corresponding to each sampling time. The physical constraint prediction model is a gray box mechanism model, which includes at least one of the following constraints: constraints between power, torque, and speed; energy conservation constraints; thermal balance constraints; and correlation constraints between vibration response, load, and speed. The model parameters of the physical constraint prediction model include at least one of friction coefficient, transmission efficiency, heat capacity, and heat transfer coefficient, and are updated according to the ambient temperature. Based on the updated physical constraint prediction model, predictions are made for each sampling time in the observation sequence to output a predicted observation sequence that corresponds one-to-one with the observation sequence. For each sampling time, the difference between the multi-channel observation data of that sampling time in the observation sequence and the predicted observation data of that sampling time in the predicted observation sequence is calculated to obtain the residual vector of that sampling time. The residual vectors of each sampling time are arranged in chronological order to obtain the residual sequence.

[0019] Optionally, S3 includes:

[0020] For each observation channel corresponding to the observation sequence, the change point statistics of the observation channel are calculated within a preset detection window based on the residual sequence, and a preset judgment threshold is determined based on the ambient temperature in the condition sequence.

[0021] When the ambient temperature is lower than a preset low temperature threshold, a low temperature judgment threshold lower than the normal temperature threshold is used; the change point statistics are compared with the preset judgment threshold to determine the candidate change point occurrence time of the observation channel; after determining the candidate change point occurrence time, the change point type corresponding to the candidate change point occurrence time is determined based on the residual sequence change characteristics and observation sequence change characteristics before and after the candidate change point occurrence time, the change point type includes mean abrupt change, variance abrupt change, output pinning, and intermittent discontinuity; the candidate change point occurrence time, the involved observation channel, the change point type, and the change point statistics are combined to generate a change point detection result;

[0022] Furthermore, the change point detection result also includes a multi-channel consistency identifier, which is used to characterize the number of observation channels where a change point occurs within a preset time tolerance range; when the number of observation channels is greater than a preset consistency threshold, the change point type is marked as a condition-related change point for subsequent anomaly attribution processing.

[0023] Optionally, S4 includes:

[0024] Receive observation sequences and conditional sequences, predicted observation sequences, and change point detection results;

[0025] A Bayesian state-space model is established, and device state variables, sensor bias variables, and sensor frozen state variables are set in the Bayesian state-space model. The sensor bias variables are used to characterize the bias change of the observation channel, and the sensor frozen state variables are used to characterize whether the observation channel is in a normal output state or a frozen state.

[0026] Based on the ambient temperature and operating parameters in the condition sequence, determine the process noise covariance matrix and measurement noise covariance matrix corresponding to each sampling time. When the ambient temperature is lower than a preset low temperature threshold, increase the prior probability that the sensor's frozen state variable is in a frozen state, and amplify the process noise covariance matrix and the measurement noise covariance matrix. At each sampling time, read the change point occurrence time, the observation channel involved, the change point type, and the change point statistics corresponding to that sampling time from the change point detection results, and adjust the measurement noise covariance matrix, the process noise covariance matrix, and the enable status of the observation channels participating in the observation update in a closed loop. Specifically, when the change point type is output pinning or intermittent discontinuity, set the enable status of the corresponding observation channel to not participate in the observation update, and adjust the enable status of the measurement noise covariance matrix corresponding to that sampling time. The noise variance corresponding to the observation channel is set to a value greater than a preset upper limit. When the change point type is a variance mutation, the noise variance corresponding to the observation channel in the measurement noise covariance matrix is ​​adaptively amplified according to the amplification factor determined by the change point statistic. When the change point type is a mean mutation, the submatrix corresponding to the sensor bias variable in the process noise covariance matrix is ​​adaptively amplified according to the amplification factor determined by the change point statistic. After completing the closed-loop adjustment, Bayesian filtering is performed based on the observation sequence, the predicted observation sequence, the process noise covariance matrix, the measurement noise covariance matrix, and the observation channel enable state. The posterior estimate of the device state variable, the posterior estimate of the sensor bias variable, and the posterior probability of the sensor frozen state variable are output to form the posterior estimation result and output the posterior estimation result.

[0027] Furthermore, the Bayesian state-space model is a switching state-space model, which includes at least two modes: a normal mode and an abnormal mode, and the transition probability between the two modes is determined by the ambient temperature and operating parameters in the condition sequence.

[0028] Optionally, S5 includes:

[0029] Receive posterior estimation results and change point detection results;

[0030] Based on the posterior estimation results, the deviation index of the equipment state variable, the deviation index of the sensor bias variable, and the freezing probability index of the sensor frozen state variable are calculated respectively.

[0031] Based on the change point detection results, extract the change point type indication information corresponding to each sampling time;

[0032] When the freeze probability index is greater than the preset freeze probability threshold, or the deviation index of the sensor bias variable is greater than the preset bias threshold, or the change point type indication information indicates that the change point type is output pinning or intermittent discontinuity, the abnormal category associated with the corresponding sampling time is determined to be the acquisition link abnormality, and the abnormal confidence level of the acquisition link abnormality is calculated based on the freeze probability index, the deviation index of the sensor bias variable and the change point type indication information.

[0033] When the deviation index of the device state variable is greater than the preset device deviation threshold, and the change point type indication information indicates that the change point type is a mean change or a variance change, the anomaly category associated with the corresponding sampling time is determined to be a device anomaly, and the anomaly confidence of the device anomaly is calculated based on the deviation index of the device state variable and the change point type indication information.

[0034] The anomaly category and its corresponding anomaly confidence level are combined to generate alarm information and output.

[0035] Optionally, the condition sequence further includes acquisition link health parameters, which include at least one of power supply voltage, communication signal strength, and sampling packet loss rate; and in S4, the measurement noise covariance matrix is ​​adaptively amplified based on the acquisition link health parameters.

[0036] On the other hand, the present invention also provides a system for detecting abnormal equipment operation data in extremely cold environments, comprising:

[0037] The data acquisition and alignment module is used to acquire multi-channel observation data corresponding to multiple sensors and perform time alignment to form an observation sequence, and to acquire the ambient temperature and operating parameters corresponding to the observation sequence to form a condition sequence.

[0038] The residual generation module is used to input the observation sequence and the condition sequence into the physical constraint prediction model, adaptively update the model parameters according to the condition sequence, output the predicted observation sequence, and generate a residual sequence from the observation sequence and the predicted observation sequence.

[0039] The online change point detection module is used to perform online change point detection on each observation channel based on the residual sequence, and outputs the change point detection results including the time of change point occurrence, the observation channel involved, the type of change point, and the change point statistics.

[0040] The Bayesian closed-loop filtering module is used to establish a Bayesian state-space model that includes equipment state variables, sensor bias variables, and sensor frozen state variables. It determines the process noise covariance matrix and the measurement noise covariance matrix based on the condition sequence. At each sampling time, it performs closed-loop adjustment on at least one of the process noise covariance matrix, the measurement noise covariance matrix, and the observation channel enable state based on the change point detection results, and then performs Bayesian filtering update to output the posterior estimation result.

[0041] The attribution alarm module is used to determine the anomaly category and output alarm information based on the posterior estimation result and the change point detection result. The anomaly category includes equipment anomaly and acquisition link anomaly, and outputs the anomaly confidence level.

[0042] The beneficial effects of this invention are:

[0043] 1. By generating residual sequences through a physical constraint prediction model based on temperature and operating conditions, and adaptively updating model parameters and noise, anomaly detection can adapt to operating condition drift and data distribution changes in extremely cold environments, reducing false alarms and missed alarms caused by fixed thresholds or models trained at normal temperatures.

[0044] 2. By using residual-based online change point detection, the timing of change points, the channels involved, the type of change point, and the statistics are obtained. The change point results are then applied in a closed loop to the process noise covariance matrix, measurement noise covariance matrix, and observation channel enable of the Bayesian state-space model. This achieves closed-loop coupling between change point detection and Bayesian filtering, improving the online identification capability of low-temperature pseudo-anomalies and real anomalies such as output pinning, intermittent discontinuity, abrupt changes in mean, and abrupt changes in variance.

[0045] 3. By introducing sensor bias variables and sensor frozen state variables into the Bayesian state-space model, and combining posterior estimation and change point type to perform anomaly attribution, anomalies can be distinguished into equipment anomalies and data acquisition link anomalies, and anomaly confidence scores can be output. This reduces unnecessary downtime and maintenance assignments, and improves the reliable early warning capability for critical faults. Attached Figure Description

[0046] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0047] Figure 1 This is a flowchart of a method for detecting abnormal equipment operation data in extremely cold environments, as proposed in this invention.

[0048] Figure 2 This is a schematic diagram of the process of updating the Bayesian state-space closed-loop filter in step S4 of the present invention. Detailed Implementation

[0049] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0050] refer to Figures 1-2 A method for detecting abnormal equipment operation data in extremely cold environments, comprising:

[0051] S1. Acquire multi-channel observation data from multiple sensors, perform time alignment processing to form an observation sequence arranged by sampling time, along with corresponding environmental temperature and operating parameters, to form a conditional sequence; S2. Input the observation sequence and conditional sequence into the physical constraint prediction model, adaptively update the model parameters based on the conditional sequence, and predict the observation sequence based on the updated physical constraint prediction model to obtain the predicted observation sequence. Generate a residual sequence based on the observation sequence and the predicted observation sequence; S3. Perform online change point detection on each observation channel of the multi-channel observation data based on the residual sequence, and output the change point detection results, including the time of change point occurrence, the observation channel involved, the change point type, and the change point statistics; S4. Establish a Bayesian state-space model, including equipment state variables, transmission... The sensor bias variable and sensor frozen state variable are determined according to the condition sequence, and the process noise covariance matrix and measurement noise covariance matrix corresponding to each sampling time are determined. At each sampling time, based on the change point detection result, at least one of the process noise covariance matrix, measurement noise covariance matrix and observation channel enable state is adjusted in a closed loop and then Bayesian filtering is performed to update it. The posterior estimates of the equipment state variable, the posterior estimates of the sensor bias variable and the posterior probability of the sensor frozen state variable are output to obtain the posterior estimation result. S5. Based on the posterior estimation result and the change point detection result, anomaly attribution processing is performed. Based on the preset judgment rule, the anomaly category is determined and alarm information is output. The anomaly categories include equipment anomaly and acquisition link anomaly, and the anomaly confidence level is output for the anomaly category.

[0052] In this specific embodiment, S1 includes:

[0053] The data acquisition and alignment module is deployed on the device-side acquisition gateway and synchronizes with a unified clock source. The gateway writes a timestamp to each reported data entry at the moment of reception and records the channel identifier, value, and validity flag. Multi-channel observation data is output from multiple sensors and numbered by observation channel. The total number of observation channels is recorded as Simultaneously, ambient temperature and operating parameters are collected to form a condition sequence, and link health parameters are further collected to supplement the condition sequence. The collected link health parameters are measured by the gateway side and include power supply voltage, communication signal strength and sampling packet loss rate.

[0054] The system pre-configures the sampling period and generates a time-ordered sequence of sampling times accordingly, where the sampling times satisfy... ,in Indicates the first Each sampling time, Indicates the alignment start time. Indicates the preset sampling period. Indicates the sampling time sequence number. Indicates the length of the observation sequence;

[0055] Time alignment processing uses a fixed time tolerance window And mapping is performed with the sampling time as the center, for each sampling time. With each observation channel Searching for timestamps falling within the range in the original reported data. The validity flag is a valid observation value; if multiple observations are found, the timestamp and the value of the observation are selected. The one with the smallest deviation is taken as the observation value of the observation channel at that sampling time. If the deviations are the same, the one with the later timestamp is selected to resist out-of-order arrival caused by communication jitter. If no observation value that meets the conditions is found, a missing mark is generated for the observation channel and a missing placeholder value is written into the observation sequence.

[0056] The missing marker uses binary encoding. A value of 0 indicates that there is a valid observation value in the observation channel at the sampling time and it participates in subsequent processing. A value of 1 indicates that there is a lack of a valid observation value in the observation channel at the sampling time and observation update suppression is performed based on the missing marker in subsequent processing.

[0057] The ambient temperature and operating parameters are mapped to each sampling time using the same time alignment rule as the multi-channel observation data to ensure a one-to-one correspondence with the observation sequence. In this embodiment, the operating parameters are organized in vector form and include three items: speed, load and power. They are reported by the equipment controller under the same time synchronization system and then aligned after being written with timestamps.

[0058] The health parameters of the acquisition link are also mapped to each sampling time. The power supply voltage is sampled by the gateway power monitoring circuit, written to a timestamp, and then aligned. The communication signal strength is read by the gateway communication module when receiving data packets and selected at each sampling time. The record with the smallest deviation is taken as the communication signal strength at that sampling moment. The sampling packet loss rate is calculated in intervals with the sampling moment as the boundary. The data is statistically obtained and written into the conditional sequence, using a preset sampling period. The expected number of data packets to arrive within this interval is estimated and compared with the actual number of arrived and valid data packets to obtain the packet loss ratio;

[0059] Finally, the multi-channel observations and missing markers corresponding to each sampling time are arranged in chronological order to form an observation sequence, and the ambient temperature, operating parameters and acquisition link health parameters that are consistent with its chronological order are arranged in chronological order to form a condition sequence.

[0060] In this specific embodiment, S2 includes:

[0061] The residual generation module receives the observation sequence and the condition sequence, and at each sampling time... After updating the model parameters of the physical constraint prediction model with temperature and operating condition conditions, a prediction observation sequence is generated and a residual sequence is calculated.

[0062] The physical constraint prediction model is a gray-box mechanism model. The model structure consists of constraints between power, torque, and speed; thermal balance constraints; and the correlation constraints between vibration response, load, and speed. The condition sequence at the sampling time... Including ambient temperature Rotation speed ,load With power And thus drive the prediction calculation;

[0063] Model parameters at sampling time The composition is as follows:

[0064] ;

[0065] in Indicates the coefficient of friction. Indicates transmission efficiency. Indicates equivalent heat capacity, Indicates the equivalent heat transfer coefficient. Represents the motor torque constant. , Represents the coefficients of the vibration correlation model;

[0066] In this embodiment, the temperature adaptive update of the model parameters uses a reference temperature. The calibration value and linear temperature compensation factor are implemented under the reference temperature, where the calibration value at the reference temperature is set to... The temperature compensation coefficient is set to and will Set as Multiply by a factor ,Will Set as Multiply by a factor ,Will Set as Multiply by a factor ,Will Set as Multiply by a factor and will Limited to the range To avoid numerical divergence and and The value is limited to a positive value to satisfy the physical realizability of the thermal equilibrium model;

[0067] In the power, torque, and speed constraint section, the module is based on... and Calculate the angular velocity and obtain the equivalent torque on the transmission side, then base it on the motor torque constant. Convert the torque into predicted current And it serves as the current channel output for predictive observation;

[0068] In the vibration correlation constraint section, the module uses a deterministic linear correlation model to map rotational speed and load to predicted effective vibration values. ,in and use The square term and The linear terms are generated together To characterize the correlation constraints between vibration response and load and rotational speed;

[0069] In the thermal balance constraint section, the module uses the bearing temperature as the thermal state and the difference between frictional loss power and convective heat transfer power as the net heat input, where the frictional loss power is determined by... Determined together with the equivalent torque and angular velocity and denoted as The forward Euler discretization method is used to realize thermal state recursion and temperature channel prediction, and the predicted bearing temperature is also used. As the temperature channel output for the predicted observation, the recursion and residual calculation satisfy the following equation:

[0070] ;

[0071] in Indicates the sampling time The bearing temperature thermal state, Indicates the sampling time The bearing temperature thermal state, This indicates the preset sampling period in step S1. Indicates the sampling time The equivalent heat capacity, Indicates the sampling time Frictional power loss Indicates the sampling time The equivalent heat transfer coefficient, Indicates the sampling time Ambient temperature, Indicates the sampling time The multi-channel observation vector, and defined in this embodiment as ,in Represents the observed current value. This represents the observed effective value of vibration. This represents the observed bearing temperature. Indicates and One-to-one corresponding prediction observation vector and defined as ,in Indicates the predicted current. Indicates the predicted effective value of vibration. This indicates the predicted bearing temperature. Indicates the sampling time The residual vector;

[0072] The bearing temperature thermal state The initialization time is determined by taking the first valid observation value of the bearing temperature channel in step S1. If the channel is missing at the first sampling time, the initialization value is set to [value to be filled in]. And when the channel first produces a valid observation at a subsequent time, the valid observation is used to overwrite the initial value to eliminate the initial value bias;

[0073] When the missing marker in step S1 indicates that a certain observation channel lacks a valid observation value at a certain sampling time, the residual generation module still calculates the predicted observation value corresponding to that channel according to the above physical constraint prediction model and adds it to the residual vector. The missing placeholder value is written to the channel, and the missing marker is output along with the residual sequence so that the subsequent online change point detection only performs statistical calculation and threshold comparison on the valid residual components, thereby obtaining the predicted observation sequence and residual sequence that correspond one-to-one with the observation sequence according to the sampling time.

[0074] In this specific embodiment, S3 includes:

[0075] The online change point detection module receives the residual sequence and performs online detection channel by channel for each observation channel. At sampling time Maintenance length is A circular cache to store the most recently used items individual residual scalars With corresponding missing markers ,in Indicates the sampling time residual vector The Middle One portion, Indicates the sampling time Is the channel missing and Indicates validity Indicates missing, window length Set to 30 and only if the interval Internal satisfaction Only then did they address the issue of the channel. Calculate the change point statistics;

[0076] This statistic is constructed using the difference between the mean and variance of two adjacent windows, where the current window is... The previous window is The mean and variance of the two windows are maintained using the Welford recursive method, which is consistent with one-time calculation and online updating, to avoid repeated traversal. The mean of the residuals of the current window is denoted as . The mean of the residuals in the previous window is The variance of the current window residual is The variance of the residuals in the previous window is And calculate the change point statistics:

[0077] ;

[0078] in Indicates observation channel At sampling time The change point statistics, Indicates the weight of the mean term and sets it to... This represents the mean of the residuals in the current window. This represents the mean of the residuals from the previous window. This represents the variance of the residuals in the current window. This represents the variance of the residuals from the previous window. This indicates that positive numbers with a denominator of zero are prevented and are set to... ,symbol Absolute value operation, symbol Represents the square root operation;

[0079] The module is based on ambient temperature in the conditional sequence. Select a threshold and output the time when candidate change points occur. Low temperature determination threshold is used. ,when The threshold for determining temperature is based on normal temperature. ,in and in satisfying And the interval between it and the previously confirmed change point is not less than Under the condition of sampling time, Register the candidate change point occurrence time for this channel and record the statistics. ;

[0080] After candidate change points are registered, the module determines the change point type, first calculating the standardized mean difference. Ratio of variance ,when and When the change point type is determined to be a mean mutation, when and The change point type is then determined to be a variance mutation;

[0081] When neither of the above two conditions is met, the module further combines the observation sequence. The system distinguishes between "output pinning" and "intermittent interruptions," with output pinning determined by the nearest... The range of valid observations and sensor resolution are used to determine the settings. And for the observation channel Pre-configured sensor resolution When recently Each sampling time satisfies And the difference between the maximum and minimum observed values ​​is no greater than The variable point type will be determined as output pinned;

[0082] Intermittent discontinuity was determined by the density of missing markers, in the most recent Within each sampling time when The variable point type is determined to be intermittent and discontinuous. The occurrence time of the candidate change point is determined as the sampling time when the missing marker first appears within the window;

[0083] After obtaining the occurrence time and type of change points for each channel, the module marks the operating condition-related change points based on the multi-channel consistency identifier and sets the time tolerance to [value missing]. Each sampling time and for each sampling time Statistics in the interval Number of observation channels for endogenous mutation points ,when The system generates a multi-channel consistency flag that is true and adds a marker to the change point type of these change points as operating condition-related change points, where the consistency threshold is... Set to 3;

[0084] Finally, the change point detection results are generated by combining the time of change occurrence, the observation channels involved, the change point type, the change point statistics, and the multi-channel consistency identifier, and then output according to the sampling time.

[0085] In this specific embodiment, S4 includes:

[0086] Bayesian closed-loop filtering module receives observation sequences Predicted observation sequence The system uses conditional sequences and change point detection results to establish a switching state-space model that includes equipment state variables, sensor bias variables, and sensor frozen state variables. The equipment state variables are represented by "equipment deviation states that deviate from the predicted physical constraints" and denoted as... The sensor bias variable is used to characterize the zero-point drift of each observation channel and is denoted as... The sensor frozen state variable is used to characterize whether each observation channel is in normal output or frozen output and is denoted as . Pattern variables Used to characterize normal and abnormal patterns and Indicates normal mode, Indicates an abnormal mode;

[0087] In this embodiment, the number of observation channels is: And defined in the same way as step S2 as well as And construct residual observations within the filter. At each sampling time, the augmented continuous state Perform a linear Gaussian Bayes filter and use the following equation as the state transition and observation equation:

[0088] ;

[0089] in Indicates the sampling time The augmented continuous state vector, Indicates the sampling time The device deviation state vector, Indicates the sampling time The sensor bias vector, Indicates the sampling time The residual observation vector, Represents the state transition matrix and is taken in this embodiment. The identity matrix is ​​used to represent and Random walk evolution, Represents the observation matrix and is taken in this embodiment. The residual is characterized by being obtained by superimposing the equipment bias and the sensor bias, where express identity matrix Indicates the sampling time In mode The process noise is zero-mean and has a process noise covariance matrix. Indicates the sampling time In mode The measurement noise is zero and has a measurement noise covariance matrix. Indicates the sampling time Pattern variables;

[0090] Construct a diagonal matrix based on "equipment deviation sub-blocks and offset sub-blocks" and in normal mode. The diagonal element of the lower fixed equipment deviation sub-block is And the diagonal element of the biased sub-block is In abnormal mode The diagonal elements of the equipment deviation sub-block will be magnified 25 times in the same channel while keeping the diagonal elements of the offset sub-block unchanged;

[0091] Under the baseline condition, the diagonal element is fixed as... And based on the ambient temperature in the conditional sequence Power supply voltage Communication signal strength With sampling packet loss rate Perform adaptive scaling, where when When Multiply the whole by 4 and ,when When Multiply the whole by 4, when When Multiply the whole by 4, when When Multiply the whole by 9, and limit the upper limit of the measurement noise variance of each channel to 100 times its baseline variance to ensure numerical stability;

[0092] The mode transition probability of switching state-space models is determined by both ambient temperature and operating parameters, where the operating parameters are load. And when and Time setting and Set when the condition is not met. and This is used to perform an interactive multi-model filtering process between the two modes at each sampling time, including mixing based on the mode probabilities of the previous time step and applying filters to the two modes respectively. and Perform a Kalman prediction and update, update the mode probabilities based on the likelihood of residual observations, and fuse them to obtain... and The posterior estimate;

[0093] Sensor frozen state variables A channel-by-channel two-state hidden Markov update is used to output the frozen posterior probability, where for each channel... Maintain freeze probability In the prior prediction stage Use transition probability and ,when Use transition probability and During the observation update phase, based on the channel resolution of the observation sequence Construct a frozen evidence indicator and perform a Bayesian update, where the missing marker in step S1 characterizes the channel in If the evidence is missing, the frozen evidence indicator will be set to frozen; if the channel is not missing, then... The frozen evidence indication is set to frozen; otherwise, it is set to unfrozen, and a fixed emission probability is used. (Evidence is the freeze) and (Evidence is the freeze) Updated ;

[0094] At each sampling time, the change point detection result is read and closed-loop adjustment is performed. For observation channels with change point types of output pinning or intermittent discontinuity, the observation channel enable state is recorded as 0 and maintained for 60 sampling times. The corresponding diagonal element is directly set to To achieve consistent implementation where "it does not participate in observation updates and the noise variance is greater than a preset upper limit", for observation channels with variance mutation type, the change point statistics are used. Segmented amplification corresponds to the measurement of noise variance and... Multiply by 4, in Multiply by 9, in Multiply by 16, where The threshold value used in step S3 under the current temperature conditions will be used for observation channels where the change point type is a sudden change in mean. The diagonal elements of the corresponding bias sub-block are multiplied by 100 in the subsequent 30 sampling times to accelerate bias absorption and suppress device misjudgment caused by bias abrupt changes, while the posterior probability is frozen and satisfies The channel enable state is then recorded as 0, and the corresponding measurement noise variance is set to 0. To achieve observation update suppression in the frozen channel;

[0095] After completing the above closed-loop adjustment, perform Bayesian filtering update at this sampling time and output the posterior estimates of the equipment state variables, the posterior estimates of the sensor bias variables, and the posterior probabilities of the sensor frozen state variables. The posterior estimates of the equipment state variables are obtained by superimposing the predicted observation sequence with the posterior estimates of the equipment bias states and maintaining consistency with the observation channel order. The posterior estimates of the sensor bias variables are directly obtained from... The posterior probability is given by each channel. It is composed of and output along with the posterior probability of the pattern.

[0096] In this specific embodiment, S5 includes:

[0097] The attribution alarm module receives the posterior estimation result and the change point detection result output in step S3, and at the sampling time... Attribution and determination of causes for "device malfunction" and "acquisition link malfunction" and generation of alarm information;

[0098] The posterior estimation results include posterior estimates of the equipment state variables. and its posterior covariance matrix Posterior estimation of sensor bias variables and its posterior covariance matrix Frozen posterior probability vector of sensor frozen state variables ,in Estimation of equipment deviation status across three channels: current, effective vibration value, and bearing temperature. The bias estimation corresponds to the three channels. and Indicates the first Each observation channel at the sampling time The posterior probability of being in a frozen state;

[0099] The attribution alarm module first calculates the deviation index and the freeze probability index. These indices are obtained using posterior estimation and posterior uncertainty normalization, satisfying the following:

[0100] ,

[0101] ,

[0102] ;

[0103] in Indicates the sampling time The deviation index of equipment state variables, This represents the posterior estimate vector of the equipment state variables. express The corresponding posterior covariance matrix, sign The symbol represents the matrix inversion operation. This indicates the transpose operation. Indicates the sampling time No. The deviation index of sensor bias variables for each observation channel. Representing vectors The One portion, Representation matrix The One diagonal element, This indicates preventing positive numbers with a denominator of zero from being counted. This represents the absolute value operation. • Represents the square root operation. Indicates the sampling time No. The probability index of freezing in each observation channel;

[0104] The attribution alarm module extracts the data from the change point detection results and the sampling time. The associated variable point type indication information is generated, and a discrete indication value is generated for each observation channel. ,in And respectively represent no change point, abrupt change in mean, abrupt change in variance, output pinning, and intermittent discontinuity, and are marked with "in the interval The memory will change the channel at the moment the change point occurs. of The method of setting the value to the corresponding type and otherwise setting it to 0 is used to achieve real-time association.

[0105] After completing indicator calculation and type extraction, anomaly category determination is performed, and the freeze probability threshold is set to [value missing]. The bias threshold is set to The device deviation threshold is set to When satisfied or Or there may be observation channels that meet the requirements. At that time, it will be related to the sampling time. The associated anomaly category was determined to be an acquisition link anomaly;

[0106] When satisfied or Or there may be observation channels that meet the requirements. At that time, it will be related to the sampling time. The associated anomaly category was determined to be an acquisition link anomaly;

[0107] When satisfied And there are observation channels that satisfy... At that time, it will be related to the sampling time. The associated anomaly category was determined to be equipment anomaly;

[0108] Anomaly confidence is obtained by mapping from indicators according to fixed rules and then truncated to... The confidence level of the data collection link anomaly is formed by a weighted sum of three factors: frozen evidence, biased evidence, and change-point type evidence. The frozen evidence is taken as... Biased evidence collection Variable point type evidence exists The value is 1 if the condition is met, and 0 otherwise. The three weights are set as follows: With 0.2;

[0109] The confidence level of equipment anomalies is determined jointly by evidence of equipment deviation and evidence of change point type, wherein the evidence of equipment deviation is taken from... Variable point type evidence exists If the value is 1, then the value is 0; multiply the two values ​​to obtain the confidence level of the device malfunction.

[0110] When sampling time When both the acquisition link abnormality and device abnormality conditions are met, two types of alarm information are output, each carrying its corresponding abnormality confidence level. When neither of the two conditions is met, no alarm information is output.

[0111] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0112] This invention directly addresses the core contradiction in extremely cold conditions by combining "physical constraint residuals, online change point detection, Bayesian state-space estimation, and attribution alarms," ​​namely, the difficulty in distinguishing between real faults and low-temperature pseudo-anomalies due to their similar appearances in the original observations. Specifically, it first uses a gray-box mechanism model with temperature and operating conditions to generate predictive observations and form a residual sequence, thus pre-absorbing the interpretable operating condition effects such as changes in ambient temperature and load, thereby highlighting abnormal features that do not conform to physical constraints in the residual domain. Then, online change point detection is performed on the residuals to obtain the time and type of anomaly occurrence, making patterns such as abrupt changes in mean, variance, output pinning, and intermittent interruptions explicit. Subsequently, filtering and updating are performed in the Bayesian state-space model, which includes equipment state variables, sensor bias variables, and sensor frozen state variables. Finally, the attribution of equipment anomalies and acquisition link anomalies is completed by combining posterior estimation and change point types, and the anomaly confidence is output, thereby reducing false alarms and false negatives and improving the reliability of early warning in extremely cold environments.

[0113] In terms of algorithm structure, this invention makes targeted improvements for extreme cold scenarios to address the technical problems. Firstly, it constructs a "variable point and Bayesian filter closed-loop coupling" mechanism. At each sampling moment, the variable point statistics and variable point type are inversely applied to the process noise covariance matrix, measurement noise covariance matrix, and observation channel enable state. For output pinning and intermittent interruptions, channel elimination and measurement noise enhancement are performed; for variance mutations, measurement noise is adaptively amplified; and for mean mutations, process noise in the bias subspace is adaptively amplified. This allows the filter to quickly absorb bias drift and suppress the propagation of false anomalies caused by freezing and intermittent interruptions. Secondly, it introduces latent variables for sensor bias and freezing state, and adaptively adjusts their priors and noise levels according to ambient temperature and link health parameters. Simultaneously, multi-channel consistency identifiers can be used to mark condition-related variable points to reduce the misjudgment of linkage changes caused by operating conditions as faults. These structural improvements enable anomaly detection not only to discover anomalies but also to perform interpretable differentiation and robust online processing of anomaly sources, thereby better achieving false anomaly suppression and reliable early warning of real faults in extreme cold environments.

Claims

1. A method for detecting abnormal equipment operation data in extremely cold environments, characterized in that, include: S1. Acquire multi-channel observation data from multiple sensors, perform time alignment processing to form an observation sequence arranged according to sampling time, as well as the corresponding ambient temperature and operating parameters, to form a conditional sequence; S2. Input the observation sequence and condition sequence into the physical constraint prediction model, adaptively update the model parameters according to the condition sequence, and predict the observation sequence based on the updated physical constraint prediction model to obtain the predicted observation sequence. Generate the residual sequence based on the observation sequence and the predicted observation sequence. S3. Based on the residual sequence, perform online change point detection on each observation channel of the multi-channel observation data, and output the change point detection results, including the time of change point occurrence, the observation channel involved, the type of change point, and the change point statistics; S4. Establish a Bayesian state-space model, including equipment state variables, sensor bias variables, and sensor frozen state variables. Determine the process noise covariance matrix and measurement noise covariance matrix corresponding to each sampling time according to the condition sequence. At each sampling time, perform closed-loop adjustment on at least one of the process noise covariance matrix, measurement noise covariance matrix, and observation channel enable state based on the change point detection results, and then perform Bayesian filtering update. Output the posterior estimate of the equipment state variables, the posterior estimate of the sensor bias variables, and the posterior probability of the sensor frozen state variables to obtain the posterior estimation results. S5. Based on the posterior estimation results and change point detection results, perform anomaly attribution processing, determine the anomaly category based on the preset judgment rules and output alarm information. The anomaly categories include equipment anomalies and acquisition link anomalies, and output the anomaly confidence level for the anomaly category.

2. The method for detecting abnormal equipment operation data in extremely cold environments according to claim 1, characterized in that, S1 includes: Collect multi-channel observation data output from the multiple sensors, as well as the corresponding ambient temperature and operating parameters; Add timestamps to the multi-channel observation data; Multiple sampling times are determined in chronological order according to a preset sampling period, and the multi-channel observation data is mapped to the multiple sampling times according to timestamps, so as to form corresponding multi-channel observation data at each sampling time. When any observation channel is missing an observation value at any sampling time, a missing value marker is generated for that observation channel; The observation sequence is obtained by arranging the multi-channel observation data and missing markers corresponding to each sampling time in chronological order. The environmental temperature and operating parameters are arranged in a time sequence consistent with the observation sequence to obtain the condition sequence.

3. The method for detecting abnormal equipment operation data in extremely cold environments according to claim 1, characterized in that, S2 include: The model parameters of the physical constraint prediction model are updated based on the ambient temperature and operating condition parameters in the condition sequence, so that the model parameters of the physical constraint prediction model match the ambient temperature and operating condition parameters corresponding to each sampling time. The physical constraint prediction model is a gray box mechanism model, which includes at least one of the following constraints: constraints between power, torque and speed, energy conservation constraints, thermal balance constraints, and correlation constraints between vibration response and load and speed. The model parameters of the physical constraint prediction model include at least one of the following: friction coefficient, transmission efficiency, heat capacity, and heat transfer coefficient, and are updated according to the ambient temperature. Based on the updated physical constraint prediction model, predictions are made for each sampling time in the observation sequence to output a predicted observation sequence that corresponds one-to-one with the observation sequence. For each sampling time, the difference between the multi-channel observation data of that sampling time in the observation sequence and the predicted observation data of that sampling time in the predicted observation sequence is calculated to obtain the residual vector of that sampling time. The residual vectors at each sampling time are arranged in chronological order to obtain the residual sequence.

4. The method for detecting abnormal equipment operation data in extremely cold environments according to claim 1, characterized in that, S3 includes: For each observation channel corresponding to the observation sequence, the change point statistics of the observation channel are calculated within a preset detection window based on the residual sequence, and a preset judgment threshold is determined based on the ambient temperature in the condition sequence. When the ambient temperature is lower than the preset low temperature threshold, a low temperature judgment threshold lower than the normal temperature threshold is used; the change point statistics are compared with the preset judgment threshold to determine the candidate change point occurrence time of the observation channel; After determining the time of occurrence of the candidate change point, the change point type corresponding to the time of occurrence of the candidate change point is determined based on the change characteristics of the residual sequence before and after the time of occurrence of the candidate change point and the change characteristics of the observed sequence. The change point type includes abrupt change in mean, abrupt change in variance, output pinning, and intermittent discontinuity. The change point detection result is generated by combining the occurrence time of the candidate change point, the observation channel involved, the change point type, and the change point statistics.

5. The method for detecting abnormal equipment operation data in extremely cold environments according to claim 1, characterized in that, S4 include: Receive observation sequences and conditional sequences, predicted observation sequences, and change point detection results; A Bayesian state-space model is established, and device state variables, sensor bias variables, and sensor frozen state variables are set in the Bayesian state-space model. The sensor bias variables are used to characterize the bias change of the observation channel, and the sensor frozen state variables are used to characterize whether the observation channel is in a normal output state or a frozen state. Based on the ambient temperature and operating parameters in the condition sequence, determine the process noise covariance matrix and the measurement noise covariance matrix corresponding to each sampling time. When the ambient temperature is lower than a preset low temperature threshold, the prior probability that the sensor's frozen state variable is in a frozen state is increased, and the process noise covariance matrix and the measurement noise covariance matrix are amplified. At each sampling time, the change point occurrence time, the involved observation channels, the change point type, and the change point statistics corresponding to that sampling time are read from the change point detection results. Based on this, the measurement noise covariance matrix, the process noise covariance matrix, and the enable status of the observation channels involved in the observation update are adjusted in a closed loop, wherein: When the variable point type is output pinning or intermittent discontinuity, the observation channel enable state of the corresponding observation channel is set to not participate in the observation update, and the noise variance in the measurement noise covariance matrix corresponding to the observation channel is set to a value greater than the preset upper limit. When the change point type is a variance mutation, the noise variance corresponding to the observation channel in the measurement noise covariance matrix is ​​adaptively amplified according to the amplification factor determined by the change point statistic; When the change point type is a mean abrupt change, the submatrix in the process noise covariance matrix corresponding to the sensor bias variable is adaptively amplified according to the amplification factor determined by the change point statistic; After completing the closed-loop adjustment, Bayesian filtering is performed based on the observation sequence, the predicted observation sequence, the process noise covariance matrix, the measurement noise covariance matrix, and the observation channel enable state. The posterior estimates of the equipment state variables, the posterior estimates of the sensor bias variables, and the posterior probabilities of the sensor frozen state variables are output, forming the posterior estimation results, which are then output.

6. The method for detecting abnormal equipment operation data in extremely cold environments according to claim 1, characterized in that, S5 include: Receive posterior estimation results and change point detection results; Based on the posterior estimation results, the deviation index of the equipment state variable, the deviation index of the sensor bias variable, and the freezing probability index of the sensor frozen state variable are calculated respectively. Based on the change point detection results, extract the change point type indication information corresponding to each sampling time; When the freeze probability index is greater than the preset freeze probability threshold, or the deviation index of the sensor bias variable is greater than the preset bias threshold, or the change point type indication information indicates that the change point type is output pinning or intermittent discontinuity, the abnormal category associated with the corresponding sampling time is determined to be the acquisition link abnormality, and the abnormal confidence of the acquisition link abnormality is calculated based on the freeze probability index, the deviation index of the sensor bias variable and the change point type indication information. When the deviation index of the device state variable is greater than the preset device deviation threshold, and the change point type indication information indicates that the change point type is a mean change or a variance change, the anomaly category associated with the corresponding sampling time is determined to be a device anomaly, and the anomaly confidence of the device anomaly is calculated based on the deviation index of the device state variable and the change point type indication information. The anomaly category and its corresponding anomaly confidence level are combined to generate alarm information and output.

7. The method for detecting abnormal equipment operation data in extremely cold environments according to claim 4, characterized in that, The change point detection result also includes a multi-channel consistency identifier, which is used to characterize the number of observation channels where a change point occurs within a preset time tolerance range. When the number of observation channels is greater than a preset consistency threshold, the change point type is marked as a condition-related change point for subsequent anomaly attribution processing.

8. The method for detecting abnormal equipment operation data in extremely cold environments according to claim 5, characterized in that, The Bayesian state-space model is a switching state-space model, which includes at least two modes: normal mode and abnormal mode. The transition probability between the two modes is determined by the ambient temperature and operating parameters in the condition sequence.

9. The method for detecting abnormal equipment operation data in extremely cold environments according to claim 1, characterized in that, The condition sequence also includes acquisition link health parameters, which include at least one of the following: power supply voltage, communication signal strength, and sampling packet loss rate; and in S4, the measurement noise covariance matrix is ​​adaptively amplified based on the acquisition link health parameters.

10. A system for detecting abnormal equipment operation data in extremely cold environments, used to execute the method for detecting abnormal equipment operation data in extremely cold environments as described in any one of claims 1 to 9, characterized in that, include: The data acquisition and alignment module is used to acquire multi-channel observation data corresponding to multiple sensors and perform time alignment to form an observation sequence, and to acquire the ambient temperature and operating parameters corresponding to the observation sequence to form a condition sequence. The residual generation module is used to input the observation sequence and the condition sequence into the physical constraint prediction model, adaptively update the model parameters according to the condition sequence, output the predicted observation sequence, and generate a residual sequence from the observation sequence and the predicted observation sequence. The online change point detection module is used to perform online change point detection on each observation channel based on the residual sequence, and outputs the change point detection results including the time of change point occurrence, the observation channel involved, the type of change point, and the change point statistics. The Bayesian closed-loop filtering module is used to establish a Bayesian state-space model that includes equipment state variables, sensor bias variables, and sensor frozen state variables. It determines the process noise covariance matrix and the measurement noise covariance matrix based on the condition sequence. At each sampling time, it performs closed-loop adjustment on at least one of the process noise covariance matrix, the measurement noise covariance matrix, and the observation channel enable state based on the change point detection results, and then performs Bayesian filtering update to output the posterior estimation result. The attribution alarm module is used to determine the anomaly category and output alarm information based on the posterior estimation result and the change point detection result. The anomaly category includes equipment anomaly and acquisition link anomaly, and outputs the anomaly confidence level.

Images