A deep learning-based charging robot fault diagnosis method and system

Abstract

The application discloses a kind of based on deep learning's charging robot fault diagnosis method and system, specifically relates to robot fault diagnosis technical field, including the following steps: acquisition multiple types of sensing channel data in running process, constructs structured input sequence;Local disturbance amplitude and trend difference are generated weak abnormal confidence value, determine whether trigger subsequent analysis;If trigger, extract modal drift residual index and micro abnormal evolution trend index, splice and form fusion feature vector;The numerical coefficient of each channel is output by trained deep learning model;According to numerical coefficient and feature adjustment sampling frequency;The application realizes the dynamic closed-loop regulation of sampling frequency by unified multi-channel data processing flow, constructs trend-driven anomaly recognition mechanism, and is combined with deep learning model output, improves the data consistency, recognition sensitivity and resource utilization efficiency of charging robot fault diagnosis system, realizes the fine, adaptive monitoring of multiple channel state.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of robot fault diagnosis, and more particularly, to a charging robot fault diagnosis method and system based on deep learning. BACKGROUND

[0002] With the continuous expansion of the application of charging robots in the fields of intelligent transportation, power inspection, warehouse logistics, etc., the continuity, safety and intelligence of their operation have increasingly become the focus of the industry. As a core link in service robot systems, during the long-term autonomous operation of charging robots, internal actuators, perception systems, power systems and other key components are easily affected by environmental disturbances, fatigue loss or control errors, resulting in potential faults or performance degradation, which in turn affects the overall task execution efficiency and system operation safety.

[0003] Currently, for the judgment of the health status of charging robots during operation, the mainstream methods include rule judgment based on threshold, frequency domain signal analysis, expert knowledge base reasoning and shallow learning modeling, etc. These methods mostly rely on manual setting of feature extraction processes and make judgments based on single-channel monitoring results, lacking modeling capabilities for the coupling relationship between channels, trend evolution patterns and complex background noise. Especially in current application scenarios, charging robots often deploy multiple types of sensor channels simultaneously to achieve high-dimensional perception, such as voltage channels and current channels for reflecting energy status, displacement channels and acceleration channels for reflecting structural motion information, torque channels for recording changes in force on the execution unit, and environmental temperature channels for reflecting external operating environment changes, etc.

[0004] Although multi-source data provides a rich information base for fault identification, due to differences in sampling frequency, data distribution and physical meaning, traditional unified rules cannot take into account the characteristics and risk performance of various channels. In addition, in actual systems, potential faults often present as early weak abnormal signals with low intensity, low frequency and slow accumulation, which are manifested as an increase in local disturbance amplitude or unstable changes in trend direction. Traditional diagnosis methods often trigger responses only when the anomaly evolves into a significant deviation, missing the window period for proactive maintenance and early intervention. Therefore, a charging robot fault diagnosis method and system based on deep learning is proposed to address the above problems. SUMMARY

[0005] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0006] A charging robot fault diagnosis method based on deep learning, comprising the following steps:

[0007] Collecting multiple types of sensing channel data of the charging robot during operation, time aligning, normalizing and abnormality rejecting the collected data, constructing a structured data input sequence for subsequent diagnosis process as a unified input basis;

[0008] Based on the structured data input sequence, the local disturbance amplitude of each channel and the trend difference between channels are calculated within a continuous time period to form a weak abnormality confidence value for identifying latent abnormalities, which is used to determine whether to trigger further analysis steps;

[0009] When the analysis step is triggered, two types of feature quantities are extracted, one is a modal drift residual index based on the joint change trend of multiple channel data, and the other is a micro-abnormal evolution trend index based on the historical change direction of a single channel. The two types of feature quantities are spliced to form a fusion feature vector;

[0010] The fusion feature vector is finally output by multiple calls to a deep learning model trained on historical labeled data. Each numerical coefficient corresponds to a channel, representing the severity of potential faults in the current state of the channel.

[0011] Based on the numerical coefficient of each channel, the micro-abnormal evolution trend index, the initial sampling frequency of the channel, and the current modal drift residual index, the sampling frequency of each channel is adjusted, and the frequency after adjustment is used to update the data input sequence of the next period.

[0012] In a preferred embodiment, before time aligning, normalizing and abnormality rejecting the collected data, the channel data is mapped to channel identification, establishing the matching relationship of channel identification of different channel data within the same time interval.

[0013] In a preferred embodiment, when time aligning the collected data, based on a unified timestamp index rule, linear interpolation is used to fill in missing time points in each data channel.

[0014] When normalizing the collected data, the minimum-maximum normalization method is used to convert the numerical value of each channel data within a specified time window into standardized data within a fixed interval range according to the proportional relationship between the minimum value and the maximum value.

[0015] In a preferred embodiment, during the calculation of the local disturbance amplitude of each channel based on the structured data input sequence, the calculation of the local disturbance amplitude uses a window sliding mechanism. The first-order difference of each channel's continuous data segment within a preset length of time window is processed, and the average of the absolute difference between adjacent data points is taken as the local disturbance amplitude of the channel within the corresponding time window.

[0016] The calculation of the inter-channel trend difference comprises performing linear fitting on each channel in a continuous time period, extracting the fitting slope value, and calculating a set of slope difference values between all channels, taking the maximum value of the set of slope difference values to form a quantitative indicator of the trend difference.

[0017] In a preferred embodiment, when forming a weak anomaly confidence value for identifying a latent anomaly, the local disturbance amplitude of each channel in the structured data input sequence in a continuous time period is first standardized to a unified dimension interval respectively with its corresponding inter-channel trend difference, and then the standardized local disturbance amplitude and trend difference are respectively assigned a pre-set static weight factor, and the anomaly sensitivity value of each channel is calculated by merging and calculating, and the maximum value of the anomaly sensitivity values of all channels is extracted as the final weak anomaly confidence value, which is used to depict the most significant latent anomaly performance in the overall multi-channel signal in the current continuous time period; after obtaining the weak anomaly confidence value, it is compared with the set multi-level judgment threshold value, if it is in the critical interval, further analysis steps are triggered, if it is in the safe interval, the subsequent analysis process is terminated.

[0018] In a preferred embodiment, when constructing a modal drift residual index based on the joint change trend of multiple channels, first, a first-order difference sequence is extracted for each channel in the structured data input sequence under the same time window, and each value in the difference sequence is converted into a symbol sequence, with positive values as positive symbols, negative values as negative symbols, and zero values as static symbols, to obtain a symbol trend sequence for each channel.

[0019] The symbol trend sequences of any two channels are position-aligned, and the symbol pairs are compared in sequence from the first position, with the same symbol pair as a match, and the number of matches is recorded; then the average matching density between the two sequences is calculated, defined as the number of matches divided by the length of the symbol sequence, and further subtracting the average symbol jump position difference between the two sequences, which is the average offset distance between the two channels at all symbol change points, and the result obtained is defined as the consistency coefficient of the channel pair.

[0020] A trend cross matrix is constructed, in which each element represents the consistency coefficient of the trend direction of two channels in each coding segment, and the difference between the maximum value and the minimum value of all consistency coefficients is extracted from the matrix as a range span value, which is then combined with the standard deviation and skewness of the matrix to form a weighted sum to constitute a modal drift residual index, which is used to reflect the overall amplitude of the joint trend difference of multiple channels in the current time window.

[0021] In a preferred embodiment, when constructing the micro-anomaly evolution trend index based on single-channel historical change direction, for each channel in the structured data input sequence, equal-interval segments are extracted under the same time window, a linear trend function is fitted for each segment, and the slope direction, slope amplitude, and residual sum of squares of each trend function are recorded. After obtaining the entire trend function sequence within the time window, the trend direction transformation frequency, average residual fluctuation amplitude, and slope amplitude standard deviation are calculated. The three types of statistics are normalized to form a three-dimensional feature vector. The distance between the three-dimensional feature vector and the normal reference trend vector is calculated using the normalized Euclidean distance function. The resulting distance value is the micro-anomaly evolution trend index, reflecting the combination of change direction volatility, trend stability, and residual disturbance within the time window for the current channel.

[0022] In a preferred embodiment, the deep learning model is a convolutional neural network model, and each time a channel corresponding feature combination input is selected. The convolutional neural network model is called multiple times, and each time a numerical coefficient is output. Finally, multiple numerical coefficients are obtained.

[0023] In a preferred embodiment, during the process of adjusting the sampling frequency of each channel, for each channel, the numerical coefficient corresponding to the channel and the micro-anomaly evolution trend index are weighted and averaged, where the weights are the channel pre-set parameters. The resulting weighted value is multiplied by the current modal drift residual index to obtain the risk enhancement factor of the channel.

[0024] The risk enhancement factor is then multiplied by the initial sampling frequency of the channel to obtain the recommended sampling frequency. The recommended sampling frequency is then adjusted with upper and lower boundary limits, where the lower limit is fifty percent of the initial sampling frequency and the upper limit is twice the initial sampling frequency. Values exceeding the boundary are truncated to the boundary value. Finally, the updated sampling frequency of each channel uses the recommended sampling frequency as the collection interval control parameter to schedule the data collection rate of the channel in the next period. This adjustment mechanism is executed once in each data processing period to achieve differentiated dynamic updating of the sampling frequencies of all channels.

[0025] In a preferred embodiment, a deep learning-based charging robot fault diagnosis system specifically includes:

[0026] A data acquisition module for acquiring multiple types of sensor channel data of a charging robot during operation, performing time alignment, normalization, and anomaly rejection processing on the collected data, and constructing a structured data input sequence as a unified input basis for subsequent diagnosis process;

[0027] Anomaly discrimination module, based on the structured data input sequence, calculates the local disturbance amplitude of each channel and the trend difference between channels in a continuous time period, forming a weak anomaly confidence value for identifying potential anomalies, which is used to determine whether to trigger further analysis steps;

[0028] Feature extraction module, for extracting two types of feature quantities when the analysis step is triggered, one is a modal drift residual index constructed based on the joint change trend of multiple channel data, and the other is a micro anomaly evolution trend index constructed based on the historical change direction of a single channel, and the two types of feature quantities are spliced to form a fusion feature vector;

[0029] Fault recognition module, for passing the fusion feature vector through a deep learning model trained with historical labeled data multiple times, and the final output of each numerical coefficient corresponds to a channel, indicating the severity of potential faults in the current state of the channel;

[0030] Frequency regulation module, based on the numerical coefficient of each channel, the micro anomaly evolution trend index, the initial sampling frequency of the channel and the current modal drift residual index, adjusts the sampling frequency of each channel respectively, and the frequency adjustment is used to update the data input sequence of the next period.

[0031] The technical effects and advantages of the present application are:

[0032] The present application collects multiple types of sensing channel data of the charging robot during the running process, and sequentially performs time alignment, normalization and abnormality rejection processing on the collected data, constructs a structured data input sequence, which is used as a unified input basis for the subsequent diagnosis process. Through this processing mechanism, the differences in sampling time, data distribution and noise characteristics between different channels are effectively eliminated, so that the data of all channels can participate in the calculation under a unified time window, and have a unified numerical scale and effectiveness guarantee, avoiding interference with subsequent diagnosis due to sampling delay, amplitude anomaly or drift error of a certain channel. This mechanism ensures the consistency of multi-channel fusion calculation, effectively improves the robustness of subsequent feature extraction and model reasoning, and ensures that the diagnosis system can be applied to various combinations of channel types, with better versatility and engineering deployment flexibility.

[0033] The application is based on structured data input sequence, calculates the local disturbance amplitude of each channel and the trend difference between channels in a continuous time period, and builds a weak anomaly confidence value for identifying potential anomalies from this, which is further used to determine whether to trigger subsequent analysis steps, and builds an anomaly identification mechanism with trend fluctuation as the core indicator. Compared with the traditional single-layer diagnostic strategy based on threshold judgment, this mechanism has higher state recognition sensitivity and process adaptability. When the system is stable, the weak anomaly confidence value is in the safe interval, automatically avoiding resource redundancy investment; and once local disturbance or cross-channel trend appears abnormal, the confidence value will quickly reflect the potential anomaly, automatically activate the deeper feature extraction and depth analysis process, and realize the dynamic scheduling and intelligent focusing of system resources. The trend-driven mechanism not only guarantees the early detection capability of potential anomalies, but also improves the computational resource efficiency of the diagnosis process, so that the entire fault identification system has the running characteristics of high efficiency and adaptability.

[0034] After the analysis step is triggered, the modal drift residual index and the micro anomaly evolution trend index are extracted, a fusion feature vector is built, and a deep learning model trained by historical labeled data is called multiple times to output corresponding numerical coefficients for each channel, which are used to represent the severity of potential faults in the current state of the channel. The numerical coefficients output by the model are not used for static judgment, but further participate in the subsequent sampling frequency regulation process. The system jointly participates in the sampling frequency calculation of each channel, the corresponding micro anomaly evolution trend index, the initial sampling frequency of the channel and the current modal drift residual index, thereby realizing a multi-factor risk fusion index-based sampling frequency dynamic updating mechanism. This mechanism effectively solves the problem of resource waste or fault detection caused by fixed sampling frequency, so that each channel can flexibly adjust the sampling density according to its state change, high-risk channels are sampled to ensure timeliness, and low-risk channels are sampled to save computing power. The whole process is data-driven and has closed-loop regulation ability and periodic execution characteristics, which significantly enhances the adaptability of the system to dynamic changes in the environment and the resource scheduling optimization ability of long-term operation. BRIEF DESCRIPTION OF DRAWINGS

[0035] In order to facilitate the understanding of those skilled in the art, the application will be further described below with reference to the accompanying drawings;

[0036] Figure 1 A schematic diagram of a deep learning-based charging robot fault diagnosis method in the application.

[0037] Figure 2 A schematic diagram of a deep learning-based charging robot fault diagnosis system in the application. DETAILED DESCRIPTION

[0038] The technical solutions in the embodiments of the present application will be clearly and completely described with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative effort belong to the scope of protection of the present application.

[0039] With reference to Figure 1 Figure 2 The following examples are obtained:

[0040] Embodiment 1: A deep learning-based charging robot fault diagnosis method, comprising the following steps:

[0041] Collecting multiple types of sensing channel data of the charging robot during operation, performing time alignment, normalization and abnormality rejection processing on the collected data, constructing a structured data input sequence, and using it as a unified input basis for the subsequent diagnosis process; by performing consistency processing on different types of original sensing data, the data from different channels has comparability and cooperativeness in the same time scale. The time alignment operation ensures that the sampling points of different channels are uniformly analyzed at the same reference time, the normalization process is used to eliminate the numerical scale deviation caused by the dimensional difference of each channel, and the abnormality rejection processing excludes the interference points such as sudden jumps or sampling errors, thereby forming a data sequence format with stable input structure and unified expression ability, providing reliable, clean and continuous basic data support for subsequent algorithm analysis.

[0042] Based on the structured data input sequence, the local disturbance amplitude of each channel and the trend difference between channels are calculated in a continuous time period to form a weak abnormality confidence value for identifying latent abnormalities, which is used to determine whether to trigger further analysis steps; by using the continuity and cross-channel cooperativeness of the structured data, the short-term fluctuation characteristics of each channel and the trend deviation behavior between multiple channels are analyzed in multiple time windows. The local disturbance amplitude reflects whether each channel itself has a micro-amplitude abnormal change in a relatively stable state, and the trend difference between channels reveals whether there is a feature of synchronous decline or inconsistent motion direction between multiple channels. By fusing these two types of analysis quantities to form a weak abnormality confidence value, weak abnormal signs that have not evolved into strong faults can be discovered in advance without relying on obvious fault features, thereby establishing a preliminary discrimination mechanism for early warning, which uses the comparison between the weak abnormality confidence value and the set threshold as the basis for triggering subsequent in-depth analysis.

[0043] ​When the analysis step is triggered, two types of feature quantities are extracted, one is a modal shift residual index constructed based on the joint change trend of multiple channels, and the other is a micro abnormal evolution trend index constructed based on the historical change direction of a single channel. The two types of feature quantities are spliced to form a fusion feature vector. The potential spatial collaborative change pattern and the local historical abnormal evolution trend in the time series data are extracted and fused at the same time to enrich the feature semantics representing the current state. The modal shift residual index emphasizes whether the joint trend behavior between channels has appeared collaborative imbalance or pattern shift, while the micro abnormal evolution trend index is more concerned about the stability of the trend direction and fluctuation of a single channel in its own historical evolution process. By extracting and splicing these two types of complementary feature quantities in time and space dimensions, a fusion feature vector is constructed, which helps to improve the recognition ability of the subsequent model for potential complex fault states, especially for the ambiguous fault feature performance caused by the cross coupling between multi-source heterogeneous data.

[0044] The fusion feature vector is finally output by multiple calls to the deep learning model trained with historical labeled data. Each numerical coefficient corresponds to a channel, representing the severity of potential faults in the current state of the channel. In this step, the deep learning model has the ability to learn complex abnormal patterns through a large number of labeled historical samples. By inputting the fusion feature vector of different channels multiple times, an independent inference process is performed for each channel. Each model call performs nonlinear mapping of the specific channel feature vector trained and induced, thereby outputting a numerical coefficient that reflects the health degree of the channel state. This coefficient can be used to intuitively depict the potential fault risk level of the channel at the current state, facilitating further operations such as risk weighting, strategy response, or information fusion by the subsequent system. At the same time, it also avoids the limitations of traditional fault judgment logic that relies on fixed rules, enabling the system to have stronger adaptive diagnosis and learning capabilities.

[0045] Based on the numerical coefficient of each channel, the micro abnormal evolution trend index, the initial sampling frequency of the channel, and the current modal shift residual index, the sampling frequency of each channel is adjusted. The frequency after adjustment is used to update the data input sequence of the next period. This step establishes a feedback adjustment mechanism based on fault risk assessment results, which dynamically adjusts the current sampling strategy using the current risk expression result and trend evolution index of each channel. The numerical coefficient represents the severity of the fault state of the channel, the micro abnormal evolution trend index represents the persistence of short-term fluctuations, and the modal shift residual index reflects the difference degree of the channel in the global trend. These factors are comprehensively used as weighted factors to input the sampling adjustment strategy, which can adaptively increase the sampling frequency of high-risk channels to obtain more intensive diagnostic data, while reducing the sampling frequency of low-risk channels to save resources. This mechanism realizes the closed-loop feedback between data acquisition and analysis decision, significantly improving the overall response agility and data utilization efficiency of the system.

[0046] Before the time alignment, normalization and outlier rejection processing of the collected data, the channel identification mapping of each channel data is performed to establish the channel identification matching relationship of different channel data in the same time interval. The "collected data" refers to the original sensor measurement data continuously collected based on a preset time period during the actual operation of the charging robot, which contains multiple source channels. The "multi-type sensing channel data" refers to the time series measurement output from different types of sensors, including but not limited to: voltage channel data and current channel data reflecting the state of electrical energy, acceleration channel data and displacement channel data reflecting the structural characteristics of motion, torque channel data reflecting the execution state of the joint, environmental temperature channel data reflecting the external environment, and image channel data for visual identification assistance. Each type of channel may have differences in physical structure, signal dimension, sampling frequency, unit dimension, etc., so a unified mapping logic needs to be performed before analysis.

[0047] The "channel identification" refers to a unique distinguishable label set for each sampling data source channel, including channel category, channel number, physical location index, etc. field information, which is used to accurately identify the data source and corresponding relationship during data preprocessing and feature fusion. The "channel identification mapping of each channel data" refers to that after the original data collection, according to the preset channel configuration table or device sampling strategy, the identification information of the corresponding channel is added to each data point to form a "channel-data-time" three-element relationship structure, which is used for subsequent time alignment and channel correspondence logic construction.

[0048] The "channel identification matching relationship" refers to that in the same time interval, based on the timestamp record of each channel sampling data, the data channels with different sampling start and end time and sampling interval are mapped to each other through time index rules (such as linear interpolation, index interpolation) to establish a corresponding point set, so that multiple channels have a unified reference frame in the structured input sequence. For example, if the voltage channel is sampled at 0.1 second interval and the image channel is sampled at 0.5 second interval, the sampling identification corresponding to the voltage channel needs to be established at each 0.1 second node of the image channel through the interpolation mechanism to realize the channel matching mapping. The establishment of the above channel identification mapping ensures that different types of channel data are processed in a unified time dimension in the subsequent time alignment, normalization and outlier rejection processing steps, avoiding processing errors or feature distortion caused by asynchronous data distribution or channel source confusion. It is one of the prerequisite steps for constructing a structured data input sequence and an important basis for subsequent processing steps such as calculating the local disturbance amplitude, the trend difference between channels and the fusion feature vector.

[0049] When time aligning the collected data, based on a unified timestamp index rule, linear interpolation is used to fill in missing time points in each data channel; when normalizing the collected data, the minimum-maximum normalization method is used to convert the values of each channel data within a specified time window according to the proportional relationship between the minimum value and the maximum value to standardized data within a fixed interval range.

[0050] In the present application, "time alignment" refers to time synchronization operation on the original data obtained by multiple data channels, so that all channels form a consistent data structure under a unified time reference. Due to different sampling frequencies or response delays, there are often missing, asynchronous, and skip sampling problems, so a "unified timestamp index rule" is set, that is, a reference timestamp sequence is constructed according to a global sampling reference period (such as 100 ms) as the target index set for aligning all channel data. For time points with missing samples, data is filled in by performing linear interpolation in each data channel. Linear interpolation is the most commonly used interpolation method in the prior art, which is suitable for the continuity of the multi-channel sensor data of the charging robot, and can effectively avoid the interference of data structure misalignment or inconsistent sampling step of different channels on subsequent diagnosis and analysis. For example, in the scenario where the image channel is sampled every 500 ms and the voltage channel is sampled every 100 ms, linear interpolation can interpolate the image data to the same time step as the voltage channel, ensuring that each time slice in the structured data input sequence has a complete channel correspondence.

[0051] Normalization refers to standardizing the values that are not comparable due to unit, dimension, and measurement scale differences between different channels. The present application uses the minimum-maximum normalization method, which is a classic normalization form in existing data preprocessing technology. The principle is: the data in a channel is linearly mapped according to its minimum value and maximum value within a specified time window, so that it is normalized to the interval [0, 1] or any specified interval [a, b]. This method ensures that the normalized data of different channels are comparable within the same scale interval, and is suitable for subsequent anomaly detection, feature extraction, and deep model input. For example, the original measurement value of the acceleration channel may range from -2g to +2g, while the temperature channel ranges from 20°C to 80°C. Without normalization, direct input to the model may cause gradient imbalance or learning bias. Through the minimum-maximum normalization method, both can be mapped to the same interval, so that the deep learning model has stable training performance.

[0052] Anomaly removal refers to the process of identifying and removing obvious outliers or invalid samples from the input data after time alignment and normalization, preventing interference from noise or erroneous values ​​in subsequent calculations. In this invention, outliers mainly include abrupt changes, static values, and lost sensor packets. Common identification strategies include: sliding window z-score judgment: calculating the z-score of each point using the mean and standard deviation of the data within the sliding window; marking an outlier when its absolute value exceeds a set threshold (e.g., 3); continuous constant value detection: if several consecutive sampling points remain unchanged (exceeding the set static threshold), it is judged as a static anomaly; boundary value strategy: setting reasonable upper and lower physical limits (e.g., voltage should not exceed 300V), exceeding which is judged as an invalid value. Taking the current channel as an example, if the sampled value suddenly jumps to more than five times the normal range within a certain period and differs significantly from surrounding values, it can be considered a spike anomaly, which is removed or replaced with interpolated points. The data after anomaly removal will be used as the final cleaned input sequence for feature construction and model diagnosis. This processing step is a commonly used technique for preprocessing existing industrial data, and it has universal applicability and engineering feasibility.

[0053] In the process of calculating the local perturbation amplitude and inter-channel trend difference based on the structured data input sequence, the calculation of the local perturbation amplitude adopts a window sliding mechanism. Within a preset time window, the continuous data segments of each channel are subjected to first-order difference processing. The absolute difference between adjacent data points is measured and the average value is taken as the local perturbation amplitude of the channel within the corresponding time window. The calculation of inter-channel trend difference includes performing linear fitting on each channel within a continuous time period, extracting the fitting slope value, calculating the set of slope difference values ​​between all channels, and taking the maximum value of the slope difference set to form a quantitative index of trend difference.

[0054] The structured data input sequence refers to the multi-channel data matrix that has undergone the aforementioned processing, including time alignment, normalization, and anomaly removal. Its structural characteristic is a two-dimensional time series array with a unified sampling time sequence and channel identifiers. Each column corresponds to a sensing channel, and each row corresponds to the sampled values ​​of all channels at the same time. "Local disturbance amplitude" refers to a quantitative measurement of the fluctuation degree of a single channel within a continuous time period, used to reflect whether the channel exhibits non-stationary changes within a short time range. To accurately assess the fluctuation degree of a channel within a short period, this invention employs a "window sliding mechanism," that is, a fixed-length time window of L (e.g., 5 seconds) is gradually slid across the channel time series, processing the data segment within each window.

[0055] Within each time window, the data segment of that channel Perform "first-order difference processing", which means calculating the absolute value of the difference between every two adjacent points in the sequence:

[0056] Where i = 1, 2, 3, ..., L-1;

[0057] Then for all differences The arithmetic mean is used to obtain the local disturbance amplitude D of the channel within this time window: This indicator can sensitively capture weak signal anomalies such as sudden changes and small oscillations within a short period of time. For example, in the voltage channel, if the equipment experiences a short-term interruption, the amplitude of this disturbance will be significantly higher than in other windows, thus becoming the basis for subsequent anomaly confidence calculation.

[0058] Inter-channel trend differences are used to measure whether there are significant differences in direction or growth rate in the evolution trends of different channels within a unified time period. This is suitable for capturing potential synchronization misalignments or coordination anomalies within a system. The specific calculation process includes: for each channel within a "continuous time period" (equivalent to the time window mentioned above), performing a "linear fit" using linear regression, that is, fitting a function to the data sequence of each channel: Where k represents the fitted slope (channel trend), and b represents the intercept. Fitting can be performed using existing statistical methods such as least squares. After obtaining the fitted slope values ​​for all channels, a "slope difference set" is constructed. :

[0059] ; i and j are both index numbers. The maximum value among all slope differences in the above set is defined as the "Trend Difference Quantitative Index" within the current time window. Quantitative indicators of this trend difference This can effectively reveal whether a single channel exhibits behavior that deviates from the overall system consistency. For example, in acceleration, displacement, and torque channels, if the slope of one channel is significantly higher than that of the others, it indicates that the channel has exhibited abnormal behavior that differs from the global motion trend, potentially representing localized wear, jamming, or signal drift. The trend difference quantification index, used in conjunction with the aforementioned local disturbance amplitude, provides a comprehensive cross-channel and intra-channel measurement basis for forming subsequent "weak anomaly confidence values."

[0060] The weak anomaly confidence value is a numerical indicator used to represent the degree of the most significant latent abnormal feature in the current time period. This indicator is not a local performance of a specific channel, but extracts the most risk-sensitive point from the overall multi-channel to serve as a basis for further diagnosis of whether it should be triggered. In the specific implementation process, first, the "local disturbance amplitude" of each channel in the current continuous time period and the "inter-channel trend difference" indicator value of the channel are extracted from the structured data input sequence. In order to use two indicators with different sources and scales for unified calculation, the two indicators need to be "standardized". Standardization is a preprocessing method that maps the indicator value to a fixed dimension range. It is usually based on the minimum and maximum values of the indicator in the historical statistical range for linear interval mapping, so that it is converted into a relative value between zero and one.

[0061] After standardization, in order to further reflect the relative importance of the two indicators in different scenarios, the application assigns a "preset static weight factor" to each indicator. This factor is a fixed coefficient, which is set based on experience or training set statistics, representing the influence weight of local disturbance or trend difference on the overall judgment in weak anomaly judgment. For example, for inertial measurement channels, local disturbance amplitude is more sensitive; for voltage and current channels, trend difference is more diagnostic. Then, the standardized local disturbance amplitude is multiplied by its corresponding weight factor, and the trend difference indicator is also multiplied by the corresponding factor, and the sum of the two is obtained. The "anomaly sensitivity value" of each channel in the current time period. This anomaly sensitivity value can be understood as a risk comprehensive score that combines the channel's own abnormal fluctuation characteristics and cross-channel trend difference characteristics.

[0062] After obtaining the anomaly sensitivity values of all channels, the system selects the largest value in the current time period as the final "weak anomaly confidence value". This weak anomaly confidence value is used to represent the most noteworthy latent abnormal channel performance of the entire charging robot system in the time period, which is a mechanism for extracting the strongest signal from the global. Then, the weak anomaly confidence value is compared with the preset "multi-level judgment threshold". The multi-level judgment threshold is a plurality of judgment limits formulated based on a large number of historical normal and abnormal samples, which usually includes a safe interval, a critical interval and a dangerous interval. If the current weak anomaly confidence value is in the "critical interval", the system judges that there is a strong latent abnormal risk, triggers further analysis steps, and enters the subsequent feature extraction and deep reasoning process; if it is in the "safe interval", it is considered that the current system state is stable, and there is no need to start in-depth analysis, terminate the subsequent process, and maintain the normal operation of sampling and monitoring.

[0063] For example, in a typical case, only the displacement channel in the current, temperature, and displacement channels has a sudden high-frequency jitter, causing the local disturbance amplitude of this channel to increase significantly. After standardization and weight fusion, the abnormal sensitivity value of this channel is higher than that of other channels, and its value is selected as the weak anomaly confidence value of the current period. The system detects that the value is in the critical interval of the set threshold, and then starts the further analysis process of the modal drift residual index and the micro anomaly evolution trend index. Through this mechanism, the system does not need to enter the model reasoning stage every time, but only when the latent anomaly sensitivity reaches the preset risk level, the system needs to be processed in detail, improving the diagnosis efficiency and saving computing resources.

[0064] When constructing the modal drift residual index based on the joint change trend of multiple channel data, first, extract the first-order difference sequence of each channel in the structured data input sequence under the same time window, and convert each value in the difference sequence into a symbol sequence, with positive values as positive symbols, negative values as negative symbols, and zero values as static symbols, to obtain the symbol trend sequence of each channel.

[0065] Align the symbol trend sequences of any two channels in position, and compare the symbol pairs from the first position one by one. If they are the same, they are matched, and the number of matches is recorded. Then calculate the average matching density between the two sequences, defined as the number of matches divided by the length of the symbol sequence, and further subtract the average symbol jump position difference between the two sequences. The jump position difference is the average offset distance between the two channels at all symbol change points. The result obtained is defined as the consistency coefficient of the channel pair.

[0066] Construct a trend cross-matrix, where each element in the matrix represents the consistency coefficient of the trend direction of two channels in each encoding segment. Extract the difference between the maximum and minimum values of all consistency coefficients from the matrix as the range span value, and then perform a weighted sum with the standard deviation and skewness of the matrix to form the modal drift residual index, which reflects the overall amplitude of the joint trend difference of multiple channels in the current time window.

[0067] The above process measures the consistency of the change trend of multiple sensor channel output sequences of the charging robot under the current continuous time window, thereby forming the modal drift residual index reflecting the systematic shift phenomenon. Among them, the index is an important criterion for whether the system has a joint trend anomaly, and reflects whether multiple channels show signs of synchronous destruction or increased drift under factors such as function, load, and environmental disturbance.

[0068] In a specific implementation, first, the structured data input sequence is limited to a uniform length time window. The structured sequence has completed time alignment and normalization processing, and the data of each channel forms a one-dimensional continuous numerical sequence in the window. For each channel, a "first-order difference processing" is performed, that is, the subtraction operation is performed on two adjacent data points in the current data sequence to obtain a difference sequence describing the change amplitude. Then, each value in the difference sequence is converted to "symbol encoding", that is: if the value is positive, it indicates an upward trend and is converted to a "positive symbol"; if the value is negative, it indicates a downward trend and is converted to a "negative symbol"; if the value is zero, it indicates stable and no change and is converted to a "still symbol". The sequence generated by this process is the "symbol trend sequence", which reflects the change direction of the channel in the time window. Then, for any two channels, "position alignment of symbol trend sequence" is performed, that is, the symbol sequences of the two channels are compared bit by bit from the first time point to see if the symbols are the same. If the symbols are the same, it is considered that the "trend direction matches", and the position is recorded as a match. Count the number of all matching positions and divide by the total length of the sequence to get the "average matching density" of the channel pair. The higher the value, the more consistent the trend direction of the two channels. Further, in order to consider the time offset problem of trend jump, the average position difference between all symbol change points in the two symbol trend sequences is counted, that is, the "symbol jump position difference". This difference value is used to represent whether the change of the two channels is time delayed or out of sync although the overall trend is consistent. Subtract the jump position difference from the average matching density to get a "modified trend consistency score", which is defined as the "consistency coefficient" of the channel pair.

[0069] After all the pairwise consistency coefficients between channels are calculated, a "trend cross matrix" is constructed, each element of which represents the trend consistency degree between any two channels at the current time window. Next, statistical analysis is performed on the trend cross matrix. First, the maximum consistency coefficient and the minimum consistency coefficient are extracted from the matrix, and the difference between the two is calculated, defined as the "range span value", which represents the dispersion degree of the maximum and minimum trend consistency between channels. Second, the standard deviation and skewness of all consistency coefficients in the matrix are calculated. The standard deviation reflects the fluctuation degree of the consistency distribution, and the skewness reflects the symmetry of the distribution. Finally, the range span value r1, the standard deviation r2, and the skewness r3 are weighted and summed according to the weights set by experience, such as the weights f1, f2, and f3 obtained by expert assignment method, in the form of MTI = f1*r1 + f2*r2 + f3*r3, to form the "modal drift residual index" in the current time window. The modal drift residual index MTI, as a global statistical indicator of multi-channel trend consistency, is used to comprehensively evaluate whether the current system has a trend inconsistency enhancement behavior due to the imbalance of some functional components, mechanical wear or external disturbance. For example, in an actual case, the displacement, acceleration and torque channels originally have a high consistency trend; when a slight mechanism falls off at the end of the charging arm, the torque channel trend suddenly reverses, causing the signs of other channels to be inconsistent, and the corresponding consistency coefficient decreases significantly, eventually increasing the modal drift residual index.

[0070] The present application constructs a microscopic trend change discrimination mechanism for each channel, i.e. through its historical change direction, trend consistency and volatility, etc. multiple dimensions of dynamic behavior, to form a numerical indicator reflecting whether the channel has weak disturbance abnormal evolution characteristics, i.e. "micro abnormal evolution trend index". First, the processing object is each channel in the "structured data input sequence". The structured data input sequence refers to a unified format data table after preprocessing such as time alignment, normalization and abnormality removal, each table representing a channel, and each row corresponding to the multi-channel sampling results at a uniform time point. The sequence of each channel constitutes a one-dimensional time-varying history record. In the "same time window", a continuous fixed-length time interval, such as ten seconds, is selected as the analysis period, and fine-grained trend behavior segmentation extraction is performed on the channel within the period. The so-called "extracting equal-interval segments" refers to uniformly dividing the time sequence of the channel within the window into several segments, such as five segments, each containing an equal number of sampling points.

[0071] For each segment, a "linear trend function" is fitted, i.e. a straight line is used to represent the overall trend direction and slope strength of the data in the time segment. The straight line is obtained based on the least square fitting principle, representing the dominant change direction of the segment in the time period. During the fitting process, three quantities are extracted: one is the "slope direction", indicating whether the trend is rising, falling or stable; the second is the "slope amplitude", indicating the rate of trend rise or fall; the third is the "residual sum of squares", indicating the error degree between the fitted curve and the actual data, used to reflect the fluctuation stability.

[0072] After the trend function fitting of all segments is completed in the entire time window, the following three statistical quantities are calculated: trend direction transformation frequency: the trend direction between adjacent segments is counted to see if it changes, such as from rising to falling or vice versa. The higher the frequency, the more unstable the trend direction, indicating that the system has multiple reverse fluctuation phenomena. Average residual fluctuation amplitude: the residual sum of squares of each segment is averaged to measure the fitting degree of the data to the linear trend in the entire window. The larger the value, the worse the trend fitting effect and the greater the disturbance. Slope amplitude standard deviation: the dispersion of the slope amplitude of all trend functions is calculated. The larger the standard deviation, the more obvious the difference between the trend strengths of different segments in the channel, indicating that there is fluctuation anomaly. The above three statistical quantities represent three different dimensions of the trend direction consistency, trend stability and trend amplitude distribution of the channel in the time period. To facilitate comprehensive processing, the three values are uniformly "normalized", i.e. their values are mapped to the standardized range of zero to one, forming a three-dimensional feature vector with consistent dimensions. Then, the three-dimensional feature vector is compared with a pre-set "normal baseline trend vector", and the "normalized Euclidean distance function" is used to calculate the distance. The distance represents the difference between the current channel change trend characteristics and the normal state trend pattern. The larger the distance, the more the current channel change characteristics deviate from the normal behavior, indicating the possibility of abnormal evolution. The final distance value is the "micro abnormal evolution trend index", which is a fusion index based on the trend direction, fitting residual and trend amplitude statistics of the channel in the current time window, used to represent the micro abnormal fluctuation level of the channel.

[0073] For example, during the operation of the charging robot, if the temperature channel repeatedly switches the trend direction in multiple time segments, and the fitting error of each segment is large and the trend strength fluctuates sharply, the calculated trend direction transformation frequency is high, the average residual is large, and the slope standard deviation is high. The distance value between it and the normal trend vector is much larger than the set safety interval, indicating that there is a potential risk of heat dissipation abnormality, which is significantly reflected by the index.

[0074] In the foregoing process, two types of feature extraction for each channel within the target time window have been completed, namely the modal drift residual index and the micro abnormal evolution trend index, and the two are spliced to form a "fusion feature vector". The fusion feature vector can be understood as a joint representation of the temporal trend discreteness feature and the micro volatility evolution feature embodied by a single channel in the current period. The fusion feature vector is not a set of original data points, but a set of quantitative feature values extracted through structural modeling, with significant data compression and representation capabilities for input processing in deep learning models.

[0075] In the present application, the deep learning model adopts a convolutional neural network model, which is a typical structure widely used in pattern recognition, feature abstraction, and image classification tasks in the prior art. It has multiple layers of convolutional layers and nonlinear activation layers combined to automatically extract deep spatial correlation structures from input features. The convolutional neural network model is not a general model, but is trained with historical labeled data, that is, a large number of charging robots collect multi-channel sample data under known healthy and fault states, label the health state level of the corresponding channel, and perform supervised training on the model. In the specific implementation process, each time the convolutional neural network model is called, a fusion feature combination corresponding to a channel is selected as input, which is the modal drift residual index and the micro abnormal evolution trend index related to the current channel. After inputting into the convolutional neural network, the model undergoes multiple layers of feature convolution, activation transformation, pooling processing, and full connection mapping, and finally obtains a numerical coefficient in the output layer. The numerical coefficient has the following meaning: quantitatively represents the severity of potential failure under the current channel state, and the higher the value, the greater the risk. Since each model call only inputs a feature combination corresponding to a channel, the model needs to be called one by one for each channel in the structured data input sequence. The model is called multiple times in the entire processing flow, and a numerical coefficient is output each time. The multiple numerical coefficients obtained finally correspond to the current state of all channels in the structured data input sequence. Further, a numerical range-risk level table corresponding to the numerical coefficient can be set, and the specific risk level can be obtained by querying the numerical range of the numerical coefficient, for example, the risk level has ten levels, and the higher the level, the higher the risk. Intervention can be performed when the risk level reaches three, and timely risk warning can be issued in the early stage of hidden danger in the subsequent process, which is beneficial to timely intervention by maintenance personnel and has the explicitness of the fault channel.

[0076] For example, in one practical application scenario, the charging robot system accesses six different types of sensing channels, including current, voltage, temperature, displacement, acceleration and image channels. After extracting the fusion feature vectors of the six channels, they are input into the trained convolutional neural network model in turn for judgment, and the model outputs six numerical coefficients in turn. The numerical coefficient of the displacement channel is significantly higher than that of the other channels, indicating that there is a potential fault trend in the displacement signal in the current system, suggesting that there may be problems such as trajectory deviation and limit blockage, thereby providing decision basis for subsequent frequency regulation and response strategy. Compared with the parallel unified determination mode of all channels, this mechanism has stronger channel-level risk identification capability and channel-by-channel modeling advantage. It not only retains the ability of deep learning model to automatically extract high-order features, but also adapts to scenarios with large differences between multiple heterogeneous sensing channels.

[0077] For each channel, the numerical coefficient and the micro-abnormal evolution trend index corresponding to the channel are weighted and averaged, where the weights are channel pre-set parameters. The weighted value is multiplied by the current modal drift residual index to obtain the risk enhancement factor of the channel. Then, the risk enhancement factor is multiplied by the initial sampling frequency of the channel to obtain the recommended sampling frequency. The recommended sampling frequency is then adjusted within the upper and lower boundaries, where the lower boundary is 50% of the initial sampling frequency and the upper boundary is twice the initial sampling frequency. Values exceeding the boundary are truncated to the boundary value. Finally, the updated sampling frequency of each channel is used as the collection interval control parameter to schedule the data collection rate of the channel in the next period. This adjustment mechanism is executed once in each data processing period to achieve differentiated dynamic updating of the sampling frequencies of all channels.

[0078] The purpose is to realize reasonable allocation of system resources and dynamic scheduling of sensing tasks under the premise of ensuring fault diagnosis sensitivity. First, each channel is independently judged, and the input indicators are: the numerical coefficient output by the deep learning model under the current state of the channel, indicating the severity of the potential fault of the channel; the micro-abnormal evolution trend index extracted in the current period of the channel, reflecting the trend direction volatility and trend stability of the channel. Both have been calculated through fusion feature construction and trend modeling in the previous steps, with timeliness and correspondence under the same time window. The above two indicators are combined into a weighted average structure, where the weights are channel pre-set parameters. The set parameters can be developed according to the following factors:

[0079] Sensitivity level of the channel to system safety: if the voltage channel and displacement channel are crucial to the safety of the structure, a higher weight is given; historical statistical anomaly frequency: if the frequency of abnormal detection in the historical sample is higher, the reliability of the micro- abnormal evolution trend index can be improved, and a higher weight is allocated; signal stability characteristics: for channels with poor signal stability and large sampling fluctuations, the weight of the model output value is appropriately adjusted to suppress false positives; empirical engineering weight matrix: a set of channel level parameters is developed based on expert experience, and a constant array is statically configured before actual deployment. Specific examples are as follows: assuming that the importance of the current channel is higher than that of the image channel, a higher weight (for example, seven tenths) can be given to the model output value in the weighted average, while the model output of the image channel accounts for only five tenths, and the rest is supplemented by the trend index. The above weighted average result reflects the local abnormal intensity performance of the channel under the comprehensive state, and then the modal drift residual index calculated by the trend joint between channels at the current time window is multiplied to obtain the risk enhancement factor of the channel. This multiplication structure logically reflects whether the possibility of risk in the channel is amplified in the background of global trend anomalies. Continue processing, multiply the risk enhancement factor by the initial sampling frequency of the channel to obtain the recommended sampling frequency of the channel. On this basis, set the boundary conditions:

[0080] The lower limit is fifty percent of the initial sampling frequency, which is used to prevent slow sampling due to low false positive risk; the upper limit is twice the initial sampling frequency, which avoids excessive system load or over-sampling redundancy. If the calculation result exceeds the boundary, value truncation processing is performed to ensure that the recommended sampling frequency is always within a reasonable and controllable range. Finally, each channel is configured with a data collection interval control parameter according to the updated recommended sampling frequency, which is used to perform sampling scheduling in the next cycle. This sampling frequency adjustment mechanism is executed once every data processing cycle, has a periodic triggering attribute, so that the system maintains diagnostic accuracy while having adaptive sampling resource allocation capability, forming a state-driven differentiated sampling mechanism between multiple channels.

[0081] For example, in a specific implementation scenario, the current model value coefficient of the torque channel is zero point eight, the micro- abnormal evolution trend index is zero point six, the weights are zero point seven and zero point three respectively, the weighted value is zero point seven four, and the modal drift residual index is zero point nine. The risk enhancement factor is about zero point six seven, and the recommended sampling frequency is twenty seven hertz, which is less than the boundary, so it is set as the sampling frequency in the next cycle; while another stable state channel calculates the recommended frequency below its lower limit of twenty hertz, which is truncated to twenty hertz to ensure safe sampling.

[0082] Embodiment 2: A deep learning-based charging robot fault diagnosis system, specifically comprising:

[0083] A data collection module is configured to collect multi-type sensor channel data of the charging robot during operation, perform time alignment, normalization and abnormality elimination on the collected data, construct a structured data input sequence, and use the structured data input sequence as a unified input basis for a subsequent diagnosis process.

[0084] An anomaly discrimination module is configured to calculate local disturbance amplitudes and inter-channel trend differences in a continuous time period based on the structured data input sequence, and form a weak anomaly confidence value for identifying a latent anomaly. The confidence value is used to determine whether to trigger a further analysis step.

[0085] A feature extraction module is configured to extract two types of feature quantities when the analysis step is triggered. One is a modal drift residual index constructed based on joint change trends of multiple channels, and the other is a micro-anomaly evolution trend index constructed based on a historical change direction of a single channel. The two types of feature quantities are spliced to form a fusion feature vector.

[0086] A fault recognition module is configured to input the fusion feature vector into a deep learning model trained based on historical labeled data, and output a numerical coefficient corresponding to each channel. The numerical coefficient represents the severity of a potential fault in the current state of the channel.

[0087] A frequency regulation module is configured to adjust the sampling frequency of each channel based on the numerical coefficient, the micro-anomaly evolution trend index, the initial sampling frequency of the channel, and the current modal drift residual index. The adjusted sampling frequency is used to update the data input sequence of the next period.

[0088] The above formulas are dimensionless numerical calculations. The formulas are obtained by software simulation of a large amount of data to obtain a formula for the latest real situation. The preset parameters in the formulas are set by a person skilled in the art according to the actual situation.

[0089] It should be understood that the sequence numbers of the above processes do not mean the order of execution in various embodiments of the present application. The execution order of the processes should be determined according to their functions and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0090] Those skilled in the art can realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. A person skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0091] Those skilled in the art can clearly understand the specific working process of the system, device and unit described above for the convenience and brevity of description, which can refer to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0092] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A deep learning-based charging robot fault diagnosis method, characterized in that, The method comprises the following steps: Collecting multiple types of sensor channel data of the charging robot during operation, performing time alignment, normalization and abnormality elimination on the collected data, and constructing a structured data input sequence as a unified input basis for the subsequent diagnosis process; Based on the structured data input sequence, the local disturbance amplitude of each channel and the trend difference between channels are calculated within a continuous time period to form a weak anomaly confidence value for identifying latent anomalies, which is used to determine whether to trigger further analysis steps; When the analysis step is triggered, two types of feature quantities are extracted, one is a modal shift residual index constructed based on the joint change trend of multiple channel data, and the other is a micro anomaly evolution trend index constructed based on the historical change direction of a single channel. The two types of feature quantities are spliced to form a fusion feature vector; The fusion feature vector is input into a deep learning model trained based on historical labeled data, and the final output of each numerical coefficient corresponds to a channel, representing the severity of potential faults under the current state of the channel; Based on the numerical coefficient of each channel, the micro anomaly evolution trend index, the initial sampling frequency of the channel, and the current modal shift residual index, the sampling frequency of each channel is adjusted, and the frequency after adjustment is used to update the data input sequence of the next period; When constructing the modal shift residual index based on the joint change trend of multiple channel data, first extract the first-order difference sequence of each channel in the structured data input sequence under the same time window, and convert each value in the difference sequence into a symbol sequence, with positive values as positive symbols, negative values as negative symbols, and zero values as static symbols, to obtain the symbol trend sequence of each channel. Align the position of the symbol trend sequences of any two channels, and compare the symbol pairs from the first position one by one. If they are the same, they are matched, and the number of matches is recorded. Then calculate the average matching density between the two sequences, defined as the number of matches divided by the length of the symbol sequence, and further subtract the average symbol jump position difference between the two sequences. The jump position difference is the average offset distance between the two channels at all symbol change points. The result obtained is defined as the consistency coefficient of the channel pair. Construct a trend cross matrix, where each element in the matrix represents the consistency coefficient of the trend direction of the two channels in each coding segment. From the matrix, extract the difference between the maximum and minimum values of all consistency coefficients as the range span value, then jointly with the standard deviation and skewness of the matrix, perform weighted summation to constitute the modal shift residual index. When constructing the micro-anomaly evolution trend index based on the single-channel historical change direction, for each channel in the structured data input sequence, equal-interval segments are extracted under the same time window, a linear trend function is fitted for each segment, and the slope direction, slope amplitude and residual sum of squares of each trend function are recorded; after obtaining the entire trend function sequence within the time window, the trend direction transformation frequency, average residual fluctuation amplitude and slope amplitude standard deviation are calculated, and the three types of statistics are normalized to form a three-dimensional feature vector; the distance between the three-dimensional feature vector and the normal reference trend vector is calculated using the normalized Euclidean distance function, and the obtained distance value is the micro-anomaly evolution trend index.

2. The deep learning-based charging robot fault diagnosis method according to claim 1, characterized in that, Before the time alignment, normalization and abnormality rejection processing of the collected data, channel identification mapping is performed on the channel data, and the channel identification matching relationship of different channel data in the same time interval is established.

3. The deep learning-based charging robot fault diagnosis method according to claim 2, characterized in that, When performing time alignment on the collected data, based on the unified timestamp index rule, linear interpolation is used to fill in the missing time points in each data channel; When performing normalization processing on the collected data, the minimum-maximum normalization method is used to convert the values of each channel data within a specified time window into standardized data within a fixed interval range according to the proportional relationship between the minimum value and the maximum value.

4. The deep learning-based charging robot fault diagnosis method according to claim 3, characterized in that, In the process of calculating the local disturbance amplitude and the inter-channel trend difference based on the structured data input sequence, the window sliding mechanism is used to perform first-order difference processing on the continuous data segments of each channel within a preset length of time window, the absolute difference between adjacent data points is measured and averaged to obtain the local disturbance amplitude of the channel within the corresponding time window; The calculation of the inter-channel trend difference includes performing linear fitting on each channel in the continuous time period, extracting the fitting slope value, and calculating the slope difference set between all channels, and taking the maximum value of the slope difference set as the quantitative indicator of the trend difference.

5. The deep learning-based charging robot fault diagnosis method according to claim 4, characterized in that, When forming the weak anomaly confidence value for identifying latent anomalies, the local disturbance amplitude and the corresponding inter-channel trend difference of each channel in the structured data input sequence within the continuous time period are standardized to a unified dimension interval, and the standardized local disturbance amplitude and trend difference are assigned with a preset static weight factor, and the anomaly sensitivity value of each channel is calculated by combining and calculating, and the maximum value of the anomaly sensitivity values of all channels is extracted as the final weak anomaly confidence value; after obtaining the weak anomaly confidence value, it is compared with the set multi-level judgment threshold value, if it is in the critical interval, further analysis steps are triggered, if it is in the safe interval, the subsequent analysis process is terminated.

6. The deep learning-based charging robot fault diagnosis method according to claim 5, characterized in that, The deep learning model is a convolutional neural network model, which inputs a feature combination corresponding to one channel each time, and the convolutional neural network model is called multiple times, outputting a numerical coefficient each time, and finally obtaining multiple numerical coefficients.

7. The deep learning-based charging robot fault diagnosis method according to claim 6, characterized in that, In the process of adjusting the sampling frequency of each channel respectively, for each channel, the numerical coefficient corresponding to the channel and the micro-anomaly evolution trend index are weighted and averaged, wherein the weights are respectively the pre-set parameters of the channel, the obtained weighted value is multiplied by the current modal drift residual index to obtain the risk enhancement factor of the channel; Then the risk enhancement factor is multiplied by the initial sampling frequency of the channel to obtain the recommended sampling frequency, and the recommended sampling frequency is adjusted by upper and lower boundary limits, wherein the lower limit is fifty percent of the initial sampling frequency and the upper limit is twice the initial sampling frequency, and the values exceeding the boundary are truncated to the boundary value; finally, the updated sampling frequency of each channel takes the recommended sampling frequency as the collection interval control parameter, which is used to schedule the data collection rate of the channel in the next period.

8. A deep learning-based charging robot fault diagnosis system based on the deep learning-based charging robot fault diagnosis method of any one of claims 1-7, characterized in that, Specifically includes: The data acquisition module is used to collect multiple types of sensor channel data of the charging robot during operation, to perform time alignment, normalization and abnormality rejection processing on the collected data, to construct a structured data input sequence, and to use it as a unified input basis for the subsequent diagnosis process; The anomaly discrimination module calculates the local disturbance amplitude of each channel and the trend difference between channels within a continuous time period based on the structured data input sequence, forms a weak anomaly confidence value for identifying latent anomalies, and uses the confidence value to determine whether to trigger a further analysis step; The feature extraction module is used to extract two types of feature quantities when the analysis step is triggered, one is a modal drift residual index constructed based on the joint change trend of multiple channel data, and the other is a micro-anomaly evolution trend index constructed based on the historical change direction of a single channel. The two types of feature quantities are spliced to form a fusion feature vector; The fault recognition module uses the fusion feature vector to call a deep learning model trained based on historical labeled data multiple times, and finally outputs a numerical coefficient corresponding to each channel, indicating the severity of potential faults in the current state of the channel; The frequency control module adjusts the sampling frequency of each channel based on the numerical coefficient of each channel, the micro-anomaly evolution trend index, the initial sampling frequency of the channel, and the current modal drift residual index. The frequency after adjustment is used to update the data input sequence in the next period.

Images