Energy storage system test method and system based on AI camera and AI analysis

Abstract

The invention relates to the technical field of energy storage system testing, and discloses an AI camera and AI analysis-based energy storage system testing method and system, and the method comprises the steps: injecting a group of mutually orthogonal exploration excitation signals into an energy storage system, and collecting the global space-time response data of the system; comparing the global space-time response data with a preset health baseline model to identify a suspicious area with abnormal response; for the suspicious region, adaptively generating a resonance excitation signal capable of amplifying the fault characteristics to the maximum extent through an optimization algorithm, and injecting the resonance excitation signal; and collecting amplification response data of the suspicious region, and inputting the amplification response data into the deep learning model to diagnose a specific fault type. The system comprises a distributed excitation network, a global space-time sensing array and an AI analysis processing module. Targeted amplification is carried out on latent fault features by actively generating resonance excitation, the problem that early weak faults are difficult to find in a traditional method is solved, and the fault detection accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of energy storage system testing technology, specifically to an energy storage system testing method and system based on AI cameras and AI analysis. Background Technology

[0002] Energy storage systems, as a crucial component of the modern energy system, play a vital role in balancing energy supply and demand, improving energy efficiency, and ensuring the stability of energy supply. From the grid integration and consumption of renewable energy generation to peak-valley regulation of the power system, energy storage systems are indispensable. Throughout the entire lifecycle of an energy storage system, from performance verification during the R&D phase to quality control during production, and then to condition monitoring and fault diagnosis during actual operation, accurate and efficient testing is fundamental to ensuring its reliable operation and performance optimization.

[0003] However, traditional energy storage system testing methods have significant limitations in comprehensively assessing the health status of the system. Current technologies typically rely on manual visual inspections to check the physical condition of energy storage systems. This method is not only inefficient, but its results are also highly susceptible to the subjective judgment and fatigue of the inspectors. Therefore, manual inspections have a high rate of missed detection for early physical defects that are difficult to detect, such as tiny cracks in the battery casing or slight loosening of connecting components, thus creating potential safety hazards for the system.

[0004] Meanwhile, while traditional testing methods can collect the operating electrical parameters of energy storage systems through various sensors, they often struggle to perform in-depth data mining and effective correlation at the data analysis level. Faced with the massive, high-dimensional, and complex data generated during system operation, traditional data processing and analysis methods are particularly cumbersome, making it difficult to quickly and accurately correlate abstract electrical parameter changes with the specific physical equipment status. This results in an inability to provide timely warnings of potential system failure risks, hindering accurate assessment of system operating status and predictive maintenance.

[0005] The rapid development of cutting-edge technologies such as artificial intelligence and computer vision has brought opportunities for innovation in the field of energy storage system testing. Introducing AI cameras and AI analysis technology into energy storage system testing is expected to break through the bottlenecks of traditional testing methods and achieve intelligent, automated, and precise testing processes.

[0006] Therefore, this invention proposes a testing method and system for energy storage systems based on AI cameras and AI analysis to address the shortcomings of existing technologies. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a testing method and system for energy storage systems based on AI cameras and AI analysis, which solves the deficiencies of energy storage system testing technology in early and subtle fault detection, and improves the sensitivity of fault detection and the accuracy of diagnosis.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a testing method for energy storage systems based on AI cameras and AI analysis, the method comprising the following steps:

[0009] S1. Inject a set of mutually orthogonal exploration excitation signals into multiple preset locations of the energy storage system;

[0010] S2. A high-speed thermal imaging camera is used to collect global spatiotemporal response data characterizing the dynamic changes in the surface temperature of the energy storage system under a set of mutually orthogonal probing excitation signals.

[0011] S3. Compare the global spatiotemporal response data with the preset health baseline model in the spatiotemporal dimension to identify suspicious areas where the response pattern deviates.

[0012] S4. Define the amplified fault response characteristics of the suspected area as the optimization target, and generate a resonant excitation signal for the suspected area by iteratively adjusting the excitation signal parameters to solve the optimization target.

[0013] S5. Inject a resonant excitation signal into the suspicious area, and use the high-speed thermal imaging camera to collect magnified response data of the suspicious area. Then, use the matching relationship between the magnified response data and the preset fault feature library to diagnose the fault type of the suspicious area.

[0014] Preferably, in step S1, the step of injecting a set of mutually orthogonal detection excitation signals into multiple preset locations of the energy storage system includes:

[0015] A pseudo-random sequence modulation method is used to modulate multiple different sinusoidal carrier signals to generate a set of mutually orthogonal detection excitation signals;

[0016] The pseudo-random sequence modulation method uses m-sequences or Gold codes;

[0017] The generated probe excitation signal has broadband spectral characteristics.

[0018] Preferably, in step S2, the step of using a high-speed thermal imaging camera to acquire global spatiotemporal response data characterizing the dynamic changes in the surface temperature of the energy storage system under a set of mutually orthogonal probing excitation signals includes:

[0019] The AI ​​analysis and processing module serves as the master clock source, sending synchronization trigger commands to the distributed excitation network and the high-speed thermal imaging camera to align the start time of excitation injection with the start time of image acquisition.

[0020] Each frame of two-dimensional thermal image acquired by the high-speed thermal imaging camera is appended with a timestamp relative to the master clock source;

[0021] A series of two-dimensional thermal images with the timestamps are combined in chronological order to form global spatiotemporal response data, which is a three-dimensional data structure.

[0022] Preferably, in step S3, the step of comparing the global spatiotemporal response data with a preset health baseline model in the spatiotemporal dimension to identify suspicious areas where the response pattern deviates includes:

[0023] For any spatial coordinate point (x, y) in the global spatiotemporal response data, extract the response time series vector v(x, y);

[0024] Based on the response time series vector v(x,y), calculate the baseline time series vector v(x,y) of the corresponding coordinate point in the health baseline model. base The deviation between (x, y) is used to obtain the outlier score A(x, y), and the formula for calculating the outlier score A(x, y) is as follows:

[0025] A(x, y) = ||v(x, y) - v base (x, y)||2;

[0026] In the formula, A(x, y) is the anomaly score at the spatial coordinate point (x, y); v(x, y) is the response time series vector collected at the spatial coordinate point (x, y); v base (x,y) is the baseline time series vector corresponding to the spatial coordinate point (x,y) in the health baseline model; ||·||2 is the Euclidean norm operation;

[0027] Based on the anomaly scores of all spatial coordinate points, a two-dimensional anomaly score map is generated.

[0028] The set of spatial coordinate points where the abnormal score A(x,y) is greater than the preset abnormal threshold is determined as the suspicious area.

[0029] Preferably, in step S4, the step of defining the amplified fault response characteristics of the suspected region as the optimization objective, and generating a resonant excitation signal for the suspected region by iteratively adjusting the excitation signal parameters to solve for the optimization objective includes:

[0030] The optimization objective is to solve for the optimal excitation signal S.opt The optimization problem is defined by the following formula:

[0031]

[0032] In the formula, S opt Ω is the optimal excitation signal; S is the excitation signal to be optimized; S R represents the physically permissible excitation signal space. ROI (S) represents the amplified response data of the suspicious region under the action of excitation signal S; J(R) ROI (S) is the objective function used to quantify and amplify the significance of fault modes in response data;

[0033] The optimization problem is solved iteratively using a Bayesian optimization algorithm to obtain the resonant excitation signal.

[0034] Preferably, in step S5, the steps of injecting a resonant excitation signal into the suspected area, acquiring magnified response data of the suspected area using the high-speed thermal imaging camera, and then diagnosing the fault type of the suspected area by matching the magnified response data with a preset fault feature database include:

[0035] The field of view of the high-speed thermal imaging camera is focused on the suspicious area to acquire the magnified response data as a three-dimensional data tensor;

[0036] The amplified response data is input into a deep convolutional neural network model trained offline;

[0037] The deep convolutional neural network model outputs a confidence score vector based on the spatiotemporal features of the extracted amplified response data. The confidence score vector contains confidence scores corresponding to multiple preset fault types.

[0038] The fault type corresponding to the highest confidence score in the confidence score vector is determined as the diagnosed fault type.

[0039] Preferably, the method further includes the step of generating a system-level causal analysis report, the steps of which include:

[0040] The global spatiotemporal response data at any spatial coordinate point is modeled as the result of the linear superposition of all exploration excitation signals after they have passed through their respective propagation paths;

[0041] Based on the mutual orthogonality of the exploration excitation signals, a blind source separation algorithm is applied to the global spatiotemporal response data to separate the response components independently triggered by each exploration excitation signal.

[0042] The set of spatiotemporal impulse response functions is obtained by reverse engineering. Each function in the set defines the heat propagation characteristics from the injection location of the probe excitation signal to any spatial coordinate point on the surface of the energy storage system.

[0043] Preferably, the calculated set of spatiotemporal impulse response functions is used as the basis for calculation to quantify the causal influence between any two components in the energy storage system, and a system-level causal graph characterizing the causal influence path between components within the energy storage system is constructed based on the causal influence measurement and a preset causal association threshold.

[0044] Preferably, the fault type of the diagnosed suspicious area is used as a node attribute and integrated with the constructed system-level causal graph to generate an integrated diagnostic report that includes both local fault information and fault propagation path.

[0045] This invention also provides a testing system for energy storage systems based on AI cameras and AI analysis, the system comprising:

[0046] A distributed excitation network is used to inject a set of mutually orthogonal probing excitation signals into multiple preset locations of the energy storage system, and to inject resonant excitation signals into identified suspicious areas according to instructions.

[0047] The global spatiotemporal sensing array includes a high-speed thermal imaging camera. The global spatiotemporal sensing array is used to acquire global spatiotemporal response data under the action of the set of mutually orthogonal probing excitation signals, as well as amplified response data of the suspicious area under the action of the resonant excitation signal.

[0048] The AI ​​analysis and processing module is used to compare the global spatiotemporal response data with a preset health baseline model to identify the suspicious areas; to define the amplified fault response characteristics of the suspicious areas as optimization targets, to generate the resonant excitation signal through iterative solution, and to instruct the distributed excitation network to inject it; and to diagnose the fault type of the suspicious areas by using the matching relationship between the amplified response data and the preset fault feature library, and to generate an integrated diagnostic report.

[0049] The data storage module is used to store the global spatiotemporal response data, amplified response data, and integrated diagnostic reports;

[0050] The user interaction module is used to present the integrated diagnostic report in a visual manner.

[0051] This invention provides a testing method and system for energy storage systems based on AI cameras and AI analysis. It offers the following advantages:

[0052] 1. This invention actively injects a set of probing excitation signals into the energy storage system and preliminarily screens suspicious areas based on the comparison of global response data and a health baseline model, changing the traditional passive waiting test mode. In particular, by generating and injecting resonant excitation signals through iterative optimization, it can effectively amplify the characteristics of early and weak latent faults, making them stand out from the system background noise, thus solving the problem that traditional methods are difficult to detect early faults.

[0053] 2. This invention employs a closed-loop optimization method to generate a resonant excitation signal. This signal is specifically designed for the response characteristics of a particular suspicious region, ensuring the targeted amplification of fault features. Subsequent diagnosis is based on this amplified response data with improved signal-to-noise ratio, combined with a pre-trained deep convolutional neural network for classification. Compared to directly analyzing the original weak signal, this significantly reduces the probability of false positives and false negatives, improving the accuracy of the diagnostic results.

[0054] 3. This invention is not limited to diagnosing faults in individual components. By performing blind source separation calculations on global response data, it reverse-engineers the set of spatiotemporal impulse response functions from each excitation source to various points in the system, and constructs a system-level causal graph based on this. This method can clearly depict the transmission path and mutual influence relationships of faults among different components, helping users understand the root causes and development trends of faults at the system level, and providing data support for preventive maintenance and system optimization.

[0055] 4. From injecting the initial detection signal and identifying suspicious areas to generating optimized stimuli, performing amplified diagnosis, and finally generating an integrated report, the entire testing process of this invention is automatically executed by the AI ​​analysis and processing module based on preset models and algorithms, without the need for manual intervention. This approach not only greatly shortens the testing cycle, but more importantly, it eliminates subjective factors in manual detection and analysis, ensuring the repeatability of the testing process and the objective consistency of the results. Attached Figure Description

[0056] Figure 1 This is a block diagram of the energy storage system test system of the present invention;

[0057] Figure 2 This is a flowchart of the energy storage system testing method of the present invention;

[0058] Figure 3 This is a schematic diagram of the resonance excitation signal generation method of the present invention;

[0059] Figure 4 This is a visual illustration of the integrated diagnostic report of the present invention.

[0060] Among them, 10 is a distributed excitation network; 20 is a global spatiotemporal sensing array; 21 is a high-speed thermal imaging camera; 30 is an AI analysis and processing module; 31 is a data preprocessing unit; 32 is a suspicious area identification unit; 33 is a resonance excitation generation unit; 34 is a fault diagnosis unit; 35 is a causal graph construction unit; 40 is a data storage module; and 50 is a user interaction module. Detailed Implementation

[0061] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] Reference Figure 1 The present invention provides a test system for energy storage system based on AI camera and AI analysis. The system includes a distributed excitation network 10, a global spatiotemporal sensing array 20, an AI analysis and processing module 30, a data storage module 40, and a user interaction module 50.

[0063] The distributed excitation network 10 includes multiple signal generators and power amplifiers deployed at multiple preset electrical nodes in the energy storage system. The distributed excitation network 10 is communicatively connected to the AI ​​analysis and processing module 30 to receive instructions from the AI ​​analysis and processing module 30, and injects a set of mutually orthogonal probing excitation signals into multiple preset locations in the energy storage system according to the instructions, as well as injecting resonant excitation signals into identified suspicious areas.

[0064] The global spatiotemporal sensing array 20 has a high-speed thermal imaging camera 21 as its core component. The high-speed thermal imaging camera 21 is mounted on a programmable multi-axis gimbal to ensure its field of view covers all critical surfaces of the energy storage system under test. The global spatiotemporal sensing array 20 is connected to the AI ​​analysis and processing module 30 via a high-speed data interface. It is used to acquire global spatiotemporal response data characterizing the dynamic changes in the surface temperature of the energy storage system under a set of mutually orthogonal probing excitation signals, as well as amplified response data of suspicious areas under resonant excitation signals. The acquired data is then transmitted to the AI ​​analysis and processing module 30.

[0065] The AI ​​analysis and processing module 30 is physically housed in a high-performance computing server equipped with a graphics processing unit (GPU). The AI ​​analysis and processing module 30 establishes communication connections with the distributed excitation network 10, the global spatiotemporal sensing array 20, the data storage module 40, and the user interaction module 50. The AI ​​analysis and processing module 30 integrates multiple functional units, including: a data preprocessing unit 31, a suspicious region identification unit 32, a resonance excitation generation unit 33, a fault diagnosis unit 34, and a causal graph construction unit 35.

[0066] The data preprocessing unit 31 is used to perform precise time synchronization between the response data received from the global spatiotemporal sensing array 20 and the excitation signal injection record from the distributed excitation network 10, so as to ensure that any point in time in the response data has a deterministic correspondence with the instantaneous state of the excitation signal.

[0067] The suspicious region identification unit 32 receives global spatiotemporal response data and compares it with a preset health baseline model stored in the data storage module 40. Specifically, for any spatial coordinate point in the global spatiotemporal response data, this unit extracts its response time series vector and calculates the deviation between this vector and the baseline time series vector of the corresponding coordinate point in the health baseline model to obtain an anomaly score A(x,y). The suspicious region identification unit 32 identifies the set of spatial coordinate points with an anomaly score greater than a preset anomaly threshold as suspicious regions and outputs the coordinate information of the suspicious region to the resonance excitation generation unit 33.

[0068] The resonance excitation generation unit 33 is used to generate a resonance excitation signal by iteratively solving for the fault response characteristics of the suspected region, based on the coordinate information of the region. opt This unit uses the Bayesian optimization algorithm to solve the optimization problem and obtains the optimal excitation signal S. opt The parameter instructions are sent to the distributed excitation network 10.

[0069] The fault diagnosis unit 34 receives amplified response data and inputs it into a pre-trained deep convolutional neural network. The deep convolutional neural network extracts and classifies the input amplified response data, outputs confidence scores for different preset fault types, and takes the fault type with the highest confidence score as the diagnosis result.

[0070] The causal graph construction unit 35 performs blind source separation calculations on the collected global spatiotemporal response data and a set of injected mutually orthogonal probing excitation signals, solves inversely for the set of spatiotemporal impulse response functions, and constructs a system-level causal graph characterizing the causal influence paths between components within the energy storage system based on this set. Finally, this unit integrates the fault type output by the fault diagnosis unit 34 as a node attribute with the constructed system-level causal graph to generate an integrated diagnostic report.

[0071] The data storage module 40 is physically hosted in a distributed file system. The data storage module 40 is connected to the global spatiotemporal sensing array 20 and the AI ​​analysis and processing module 30, and is used to persistently store global spatiotemporal response data, amplified response data, a health baseline model, a pre-trained deep convolutional neural network model, and an integrated diagnostic report generated by the AI ​​analysis and processing module 30.

[0072] The user interaction module 50 is physically represented by an operating terminal with a graphical user interface. The user interaction module 50 is connected to the AI ​​analysis and processing module 30, and is used to receive and present integrated diagnostic reports in a visual manner (such as heatmaps or topology maps), allowing operators to set test parameters and initiate test tasks.

[0073] Reference Figure 2 The figure is a flowchart of a testing method for an energy storage system according to an embodiment of the present invention. Before executing the testing method, a preparation phase needs to be completed in advance. The core of this phase is to establish an objective and accurate reference standard for subsequent anomaly comparison.

[0074] Preparation phase: Establishment of a health baseline model

[0075] This phase is performed on an energy storage system that has been confirmed to be in a healthy state. A healthy state means that the energy storage system is a new system that has passed factory acceptance testing, or an in-service system that has just undergone comprehensive maintenance and whose performance indicators have been confirmed to be normal. The goal of this phase is to collect and process the thermal response characteristics of the healthy system under a set of standardized stimuli to construct a health baseline model.

[0076] In this phase, the system is first set to baseline acquisition mode via the user interaction module 50. Subsequently, the AI ​​analysis and processing module 30 instructs the distributed excitation network 10 to inject a set of standardized, mutually orthogonal probe excitation signals into multiple preset locations on the energy storage system. The parameters of this set of probe excitation signals (such as waveform, frequency composition, and duration) are consistent with the parameters of the probe excitation signals used in subsequent actual tests to ensure the consistency of the comparison benchmark.

[0077] Simultaneously with the injection of the probing excitation signal, the high-speed thermal imaging camera 21 in the global spatiotemporal sensing array 20 continuously images the surface of the energy storage system, acquiring global spatiotemporal response data. This data is a three-dimensional dataset recording the temperature change T(x,y,t) of each spatial coordinate point (x,y) on the system surface within time t. To reduce the impact of environmental random noise on the baseline model, this data acquisition process is repeated multiple times, yielding multiple sets of independent global spatiotemporal response data.

[0078] The data preprocessing unit 31 in the AI ​​analysis and processing module 30 receives multiple sets of global spatiotemporal response data. For each spatial coordinate point (x, y) on the system surface, the data preprocessing unit 31 extracts the corresponding response time series vector from each set of data. Subsequently, it performs point-by-point averaging on the multiple response time series vectors extracted from multiple sets of data corresponding to the same spatial coordinate point to generate a single, statistically stable baseline time series vector v. base (x,y).

[0079] Finally, the baseline time series vector v of all spatial coordinate points is... base The set of (x, y) together constitutes the health baseline model. This model is structurally stored in the data storage module 40, serving as an objective basis for spatiotemporal dimension comparison by the suspicious region identification unit 32 in subsequent tests. The establishment of this model provides a quantitative reference standard for accurately identifying minor deviations in response patterns.

[0080] Investigating the generation and injection of excitation signals:

[0081] This step is executed by the distributed excitation network 10 under the control of the AI ​​analysis and processing module 30. The AI ​​analysis and processing module 30 sends instructions to the distributed excitation network 10, which include a set of digital waveform definitions for the probe excitation signals. The probe excitation signals are a set of designed electrical signals with mutually orthogonal characteristics. Specifically, if this set of probe excitation signals is represented as a set {s1(t), s2(t), ..., s...} N Let N be the total number of excitation sources and t be the time interval. Then, for any two different signals s i (t) and s j (t) satisfies the following condition within the time period T:

[0082]

[0083] In the formula, s i (t) represents the probe excitation signal injected at the i-th position; s j (t) represents the probe excitation signal injected at the j-th position; T represents a complete cycle of the signal or an observation time window.

[0084] This orthogonality is generated by modulating different fundamental signals with pseudo-random sequences. For example, different m-sequences or Gold codes can be used to modulate a set of sinusoidal carriers. This modulation method gives each probe excitation signal a broadband spectral characteristic similar to white noise, while the signals are uncorrelated. This characteristic is a prerequisite for blind source separation calculations in subsequent causal graph construction because it ensures that the mixed thermal response generated by the system at any observation point can be uniquely decomposed mathematically back into the contribution of each independent excitation source.

[0085] Each signal generator in the distributed excitation network 10 generates a corresponding digital signal according to the instruction, and converts it into an electrical signal with the required power through a power amplifier, which is then precisely injected into multiple electrical nodes such as the battery cluster terminals and DC bus in the energy storage system.

[0086] Acquisition of global spatiotemporal response data:

[0087] At the same moment that the distributed excitation network 10 begins injecting the probe excitation signal, the AI ​​analysis and processing module 30 sends a synchronization trigger command to the global spatiotemporal sensing array 20. Upon receiving the command, the high-speed thermal imaging camera 21 performs continuous infrared thermal imaging scans of the energy storage system's surface at a preset high frame rate (e.g., 100 Hz or higher). This high frame rate is set to ensure that the rapid transient thermal effects, including high-frequency components, induced by the broadband probe excitation signal can be captured.

[0088] The acquisition process continues throughout the entire period of the probe excitation signal. The high-speed thermal imaging camera 21 outputs a series of two-dimensional thermal images arranged in chronological order. In the AI ​​analysis and processing module 30, the data preprocessing unit 31 stacks these consecutive two-dimensional image frames to form a three-dimensional data cube, namely, global spatiotemporal response data. Structurally, this data is a three-dimensional tensor D(x, y, t), where (x, y) represents the pixel spatial coordinates on the thermal image, t represents the discrete time step, and the values ​​in the tensor correspond to the temperature values ​​at specific locations at specific times.

[0089] Precise time synchronization is crucial in this process. The AI ​​analysis and processing module 30 acts as the master clock source, ensuring that the start time of the excitation injection is strictly aligned with the start time of image acquisition. Each acquired image frame is appended with a precise timestamp relative to the master clock source. This synchronization mechanism ensures that the system thermal response D(x,y,t) at any given time is accurate in subsequent analysis. k All of them can be coupled with the input excitation signals [s1(t)] at the same time. k ), s2(t k ), ...,s N (tk Establishing accurate causal relationships is the foundation for subsequent model comparisons and causal analysis.

[0090] Reference Figure 2 After completing the active exploration and global response acquisition, the obtained global spatiotemporal response data is sent to the AI ​​analysis and processing module 30 to perform model-based anomaly region identification.

[0091] Model-based anomaly region identification:

[0092] This step is performed by the suspicious area identification unit 32 within the AI ​​analysis and processing module 30. Its purpose is to quantitatively compare the real-time collected global spatiotemporal response data, which characterizes the current system state, with a pre-established health baseline model that characterizes the ideal health state, thereby accurately locating the spatial areas where the thermal response behavior deviates.

[0093] The calculation process first iterates through every spatial coordinate point (x, y) in the global spatiotemporal response data. For any given spatial coordinate point, the suspicious area identification unit 32 extracts its temperature change sequence over the entire time period from the three-dimensional data cube, forming a response time series vector v(x, y). Simultaneously, this unit loads a health baseline model from the data storage module 40 and retrieves the baseline time series vector v(x, y) corresponding to the current spatial coordinate point (x, y). base (x,y).

[0094] In obtaining the vector v(x,y) representing the actual response and the vector v representing the health baseline base After (x, y), the suspicious region identification unit 32 calculates the deviation between these two vectors and quantifies this deviation into a scalar value, namely the anomaly score A(x, y). The formula for calculating the anomaly score is:

[0095] A(x, y) = ||v(x, y) - v base (x,y)||2;

[0096] In the formula, A(x,y) is the anomaly score at the spatial coordinate point (x,y), a non-negative real number whose value directly reflects the degree to which the thermal response behavior at that point deviates from the healthy state; v(x,y) is the response time series vector collected at the spatial coordinate point (x,y); v base (x,y) is the baseline time series vector corresponding to the spatial coordinate point (x,y) in the health baseline model; ||·||2 is the Euclidean norm operation, used to calculate the straight-line distance between two time series vectors in multidimensional space.

[0097] After performing the above calculations on all spatial coordinate points, the suspicious area identification unit 32 generates a two-dimensional anomaly score map, in which the gray value or color value of each pixel is proportional to the anomaly score A(x,y) of that point.

[0098] Finally, to identify specific physical regions from the anomaly score map, the suspicious region identification unit 32 employs a threshold segmentation method. The anomaly scores A(x,y) of all spatial points are compared with a preset anomaly threshold. Any spatial coordinate point whose anomaly score is greater than the preset anomaly threshold is marked as an anomaly. The set of connected regions formed by all these anomaly points is identified as a suspicious region. The coordinate information of this suspicious region is output and transmitted to the resonance excitation generation unit 33 in the AI ​​analysis and processing module 30 for subsequent targeted testing.

[0099] Reference Figure 2 and Figure 3 After identifying suspicious areas, the method of the present invention changes from passive observation to active testing, that is, to perform adaptive generation of targeted resonant excitation.

[0100] Adaptive generation of targeted resonant excitation

[0101] This step is performed by the resonant excitation generation unit 33 in the AI ​​analysis and processing module 30. Its core objective is to adaptively design and generate an optimal excitation signal based on the specific response characteristics of the suspected area identified in the previous step. The design goal of this signal is to maximize the amplification of the physical response (i.e., thermal response) caused by the potential fault within the suspected area with minimal energy injection, thereby clearly separating weak fault features from the system background noise. Different fault types (such as internal micro-short circuits or excessive contact resistance) have different heat generation mechanisms and heat conduction paths when subjected to external electrical excitation, thus corresponding to different dynamic thermal response characteristics. Resonant excitation refers to the ability of the excitation signal parameters (such as frequency and waveform) to match the dynamic thermal response characteristics of a specific fault mode, thereby stimulating its thermal characteristics with maximum efficiency.

[0102] The generation process is structured as a closed-loop optimization problem, namely, finding an optimal excitation signal S within the physically permissible signal parameter space that maximizes a specific objective function value. opt The formula for calculating this optimization problem is:

[0103]

[0104] In the formula, S opt Ω is the optimal excitation signal, i.e., the final generated resonant excitation signal; S is the excitation signal to be optimized, which is defined by a set of parameters, such as frequency and duty cycle for a pulse width modulation (PWM) signal;S The physically permissible excitation signal space is defined by the range of values ​​for excitation signal parameters (such as voltage, current, and frequency) determined by the safety operation specifications of the energy storage system; R ROI (S) represents the amplified response data of the suspected region under the action of the excitation signal S, i.e., the dynamic temperature change of the region; J(R) ROI (S) is the objective function used to quantify the saliency of fault modes in the amplified response data. This objective function can be defined in various ways; in one specific embodiment, it is defined as the peak value of the temperature response within the suspected region to maximize the magnitude of temperature changes.

[0105]

[0106] In the formula, T(x, y, t|S) represents the coordinates (x, y) of a point in the region of suspicion (ROI) at observation time T under the influence of excitation S. obs The temperature value at time t.

[0107] To solve this optimization problem, the resonant excitation generation unit 33 employs a Bayesian optimization algorithm. (Refer to...) Figure 3 The execution flow of this algorithm is as follows:

[0108] First, the algorithm constructs an initial probabilistic surrogate model (e.g., a Gaussian process model) of the objective function J(S) based on existing prior knowledge or by performing several random samplings. This model can not only predict the possible objective function value generated by any excitation signal S, but also provide the uncertainty of this prediction.

[0109] Next, the algorithm enters an iterative optimization loop. In each iteration, the algorithm selects the next most valuable stimulus signal S to be tested based on a sampling function (e.g., an expected improvement function). This sampling function balances exploration (i.e., testing parameter regions with high uncertainty to improve the model) and exploitation (i.e., testing parameter regions with good model prediction performance to seek a better solution).

[0110] After selecting the excitation signal S, the resonant excitation generation unit 33 sends the parameter instructions of the signal to the distributed excitation network 10, which then precisely injects the signal into the suspected region. Simultaneously, the global spatiotemporal sensing array 20 focuses on this region for synchronous acquisition, obtaining response data R. ROI (S).

[0111] The resonance excitation generation unit 33 generates the resonant excitation based on the acquired R. ROI (S) Calculate the true objective function value J(R) for this test. ROI (S)). Then, this new data point (S,J(R)). ROI (S))) is used to update the probabilistic surrogate model, enabling it to make a more accurate estimate of the true form of the objective function.

[0112] This iterative process is repeated until a preset number of iterations is reached or the gain of the objective function converges to a very small threshold. Finally, the algorithm selects the signal that maximizes the objective function value from all tested excitation signals as the final resonant excitation signal S. opt The parameters of this signal will be used in the next step of amplification and acquisition.

[0113] Reference Figure 2 After adaptively generating targeted resonant excitation signals, the testing method enters the stage of fault amplification acquisition and precise diagnosis.

[0114] Amplified acquisition and precise diagnosis of faults:

[0115] This step begins with the AI ​​analysis and processing module 30 sending an instruction to the distributed excitation network 10 to optimize the optimal excitation signal S obtained in the previous step. opt The parameters (such as frequency, waveform, duty cycle, etc.) of the resonant excitation signal are sent to the network. Based on this instruction, the distributed excitation network 10 precisely injects this resonant excitation signal into the physical location corresponding to the locked suspicious area.

[0116] Simultaneously, the AI ​​analysis and processing module 30 controls the global spatiotemporal sensing array 20, focusing the field of view of its high-speed thermal imaging camera 21 on the suspected area for targeted data acquisition. Since the injected resonant excitation signal can efficiently excite the thermal response of the potential fault in the area, the amplitude and saliency of the thermal response signal acquired under this excitation are amplified. The high-resolution, high signal-to-noise ratio local spatiotemporal response data acquired in this process, covering only the suspected area, is the amplified response data.

[0117] The amplified response data is then transmitted to the fault diagnosis unit 34 in the AI ​​analysis and processing module 30. This unit contains a pre-trained deep convolutional neural network (DCNN) model. This model has been trained offline using a large number of amplified response data samples with clearly labeled faults, thereby learning the mapping relationship between different fault types (e.g., internal micro-short circuits, excessive contact resistance, electrolyte leakage, etc.) and their unique amplified thermal response patterns (i.e., specific spatiotemporal temperature distributions and evolution characteristics).

[0118] During diagnosis, the amplified response data is treated as a three-dimensional data tensor R. amp It is directly input into the deep convolutional neural network model F. cnn The model automatically processes the input R through its internal multiple convolutional layers, pooling layers, and fully connected layers. ampDeeper spatiotemporal features are extracted. The output layer of this network is processed by a softmax activation function, ultimately generating a confidence score vector C. This process can be represented by the following equation:

[0119] C = F cnn (R amp );

[0120] In the formula, C is the output confidence score vector, which has the form C = [c1, c2, ..., c K ];F cnn R represents the nonlinear transformation function of a pre-trained deep convolutional neural network model. amp The input amplified response data is a three-dimensional tensor with dimensions (width, height, time); c k The confidence score for determining whether the input data belongs to the k-th fault type is a real number between 0 and 1, and the sum of all elements in vector C is 1; K is the total number of fault types contained in the preset fault feature library.

[0121] Finally, the fault diagnosis unit 34 analyzes the output confidence score vector C. This unit determines the fault category corresponding to the component with the largest value in the vector as the final diagnosed fault type. The calculation formula for this decision-making process is as follows:

[0122]

[0123] In the formula, Fault diag The final diagnosed fault type; argmax k A function that takes the maximum value of the parameter, i.e., returns the value that makes c equal to the maximum value of the parameter. k The fault type corresponding to the index k with the largest value.

[0124] The diagnostic results will be used to generate a subsequent integrated diagnostic report.

[0125] Reference Figure 2 and Figure 4 After completing the accurate diagnosis of the fault, in order to reveal the propagation relationship of the fault at the system level and generate a final report, the test method performs system-level causal analysis and report generation.

[0126] System-level causal analysis and report generation:

[0127] This step is performed by the causal graph construction unit 35 in the AI ​​analysis and processing module 30. Its purpose is to go beyond the diagnosis of a single fault point and instead build a model that describes the energy propagation path and the mutual influence between components within the system, and integrate the specific fault types diagnosed into it.

[0128] The first part of exploring causal relationships:

[0129] The causal graph construction unit 35 first calls the global spatiotemporal response data acquired in step S2, as well as a known, injected set of mutually orthogonal probe excitation signals. In the linear time-invariant system of the energy storage system, the global spatiotemporal response data O(x, y, t) at any observation point (x, y) can be regarded as being composed of all N probe excitation signals s i (t) represents the linear superposition of the results after each propagation through its respective path. This physical process can be represented by the following convolutional model:

[0130]

[0131] In the formula, O(x, y, t) represents the temperature response observed at the spatial coordinate point (x, y) and time t; s i (t) represents the known probe excitation signal injected at the i-th position; h i (x, y, t) is the spatiotemporal impulse response function to be solved, which describes the change of temperature response at position (x, y) over time t as a unit impulse signal emitted from the i-th excitation source propagates through the system and finally occurs at position (x, y); N is the total number of excitation signal sources to be detected; * is the convolution operator; n(x, y, t) is the system and measurement noise.

[0132] Due to the injected probe excitation signal s i (t) have the property of mutual orthogonality. The causal graph construction unit 35 can take advantage of this prerequisite to solve for each independent spatiotemporal impulse response function h from the mixed observation signal O(x,y,t) in reverse through the blind source separation (BSS) algorithm. i (x,y,t). By performing this calculation on all spatial coordinate points (x,y) and all excitation sources i, a set of spatiotemporal impulse response functions is finally obtained. This set fully depicts the dynamic process of energy propagating from each excitation node to all other locations on the system surface.

[0133] Construction and integration of causal graphs:

[0134] After obtaining the set of spatiotemporal impulse response functions, the causal graph construction unit 35 further calculates the causal influence metric between components. Key components of the energy storage system (such as specific battery modules, inverters, etc.) have pre-defined spatial regions on the thermal image. The causal influence metric from component j to component k can be quantified by analyzing the impulse response energy generated by the excitation source associated with component j within the region of component k. Assuming that excitation source i is physically associated with component j, this causal influence metric M... j→k It can be calculated as:

[0135]

[0136] In the formula, M j→k A non-negative real number representing the causal effect from component j to component k; Area k h represents the spatial region occupied by component k on the thermal image. j (x,y,t) is the spatiotemporal impulse response function corresponding to the excitation source i associated with component j.

[0137] The causal graph construction unit 35 calculates the causal influence metric between all predefined component pairs and uses this to construct a weighted directed graph, i.e., a system-level causal graph. In this graph, each node represents an energy storage system component. If the causal influence metric M between two components is... j→k If the value is greater than a preset threshold, a directed edge is drawn between node j and node k, and the weight of the edge is M. j→k The value of .

[0138] Finally, this unit integrates the information. It combines the diagnostic results output by the fault diagnosis unit 34 (i.e., the specific fault type). diag This is associated with the graph node corresponding to the component that experienced the failure. (Refer to...) Figure 4 The resulting integrated diagnostic report presents the causal graph to the user in a visual format. In the graph, nodes diagnosed with faults are highlighted and their fault types are labeled. Weighted edges clearly show the energy transfer intensity and path between the fault point and other components, providing system-level decision-making support for root cause analysis and predictive maintenance. This integrated diagnostic report is stored in data storage module 40.

[0139] To further illustrate the technical solution of the present invention, a specific implementation scenario will be described below. This embodiment aims to illustrate the application of the present invention, and the specific equipment model, parameters, and processes described are merely examples and do not constitute a limitation on the scope of protection of the present invention.

[0140] Test system setup:

[0141] Reference Figure 1 Build a physical system for testing energy storage systems.

[0142] The distributed excitation network 10 consists of a programmable arbitrary waveform generator and a matching power amplifier. Each waveform generator is connected via an electrical interface to the positive busbar of a battery cluster in an energy storage system cabinet.

[0143] The global spatiotemporal sensing array 20 uses a cooled infrared camera 21 as its core component, a high-speed thermal imaging camera. This camera has an infrared resolution of 640x512 pixels, a thermal sensitivity (NETD) of less than 20mK, and a maximum sampling frame rate of 200 Hz. The camera is mounted on a three-axis precision guide rail that covers the surface of all battery modules inside the energy storage cabinet.

[0144] The AI ​​analysis and processing module 30 is deployed on a high-performance computing server, where all functional units run as software modules.

[0145] The data storage module 40 uses a 10TB RAID5 disk array.

[0146] User interaction module 50 is a graphical user interface (GUI) software installed on an industrial panel PC. All modules in the system communicate and transmit data via gigabit industrial Ethernet.

[0147] Test preparation:

[0148] Before performing the test, complete the following preparations:

[0149] Establishment of the health baseline model: A brand-new, factory-tested, and qualified industrial and commercial energy storage system cabinet of the same model is used as the benchmark system. Following the methods of the preparation phase, a standardized probing excitation signal is injected into the benchmark system, and its global spatiotemporal response data is collected by a high-speed thermal imaging camera 21. After processing by the data preprocessing unit 31, a health baseline model containing the baseline time series vectors of all spatial coordinate points is generated and stored in the data storage module 40.

[0150] Offline training of the diagnostic model: A dataset containing a large number of labeled fault samples is created using historically accumulated experimental and simulation data. Each sample is an amplified response data segment, precisely labeled with its corresponding fault type (e.g., internal micro-short circuit, excessive contact resistance at connection points, etc.). This dataset is used to perform supervised learning training on the deep convolutional neural network model in the fault diagnosis unit 34 until the model's classification accuracy on the validation set reaches a preset metric (e.g., 99%) or higher. The trained weight file is then embedded in the AI ​​analysis and processing module 30.

[0151] Test Implementation:

[0152] The test was performed on a commercial and industrial energy storage system cabinet of the same model that had been running online for six months.

[0153] The operator initiates the standard test procedure through the user interaction module 50. First, the system enters the active exploration and global response acquisition phase. The AI ​​analysis and processing module 30 instructs the distributed excitation network 10 to synchronously inject a set of mutually orthogonal pseudo-random sequence modulation signals with a duration of 60 seconds into the four battery cluster busbars. At the same time, the high-speed thermal imaging camera 21 scans the surface of all battery modules at a frame rate of 100 Hz, acquiring global spatiotemporal response data.

[0154] After data collection is complete, the system automatically enters the model-based anomaly region identification stage. The suspicious region identification unit 32 compares the real-time collected global spatiotemporal response data with the health baseline model stored in the data storage module 40. The generated anomaly score map shows that an area of ​​approximately 10cm x 5cm appears on the surface of the battery module in the third layer, and its anomaly score is significantly higher than other areas, exceeding the preset anomaly threshold. This area is identified as a suspicious region.

[0155] The system then enters the adaptive generation phase of targeted resonant excitation. The resonant excitation generation unit 33 identifies the suspected region and initiates a Bayesian optimization algorithm. After 15 iterations (each iteration taking approximately 10 seconds), the algorithm obtains an optimal resonant excitation signal, which is a square wave with a frequency of 15.5 Hz and a duty cycle of 50%. This parameter combination is determined to be the excitation parameter that maximizes the amplification of the thermal response characteristics of the suspected region.

[0156] The system then enters the fault amplification and precise diagnosis stage. The distributed excitation network 10 injects only the 15.5 Hz square wave signal generated in the previous step into the third battery cluster. The high-speed thermal imaging camera 21 focuses on the suspicious area on the 7th layer of the third battery rack, acquiring amplified response data for 20 seconds. This data is sent to the fault diagnosis unit 34. Through analysis by a pre-trained deep convolutional neural network model, the output confidence score vector shows that the confidence score for internal micro-short circuits is 0.982, the confidence score for excessive contact resistance at the connection point is 0.011, and the confidence scores for other fault types are all below 0.005. Based on this, the system diagnoses the fault type of this module as an internal micro-short circuit.

[0157] Finally, the system enters the system-level causal analysis and report generation stage. The causal graph construction unit 35 uses the earliest collected global data and probe excitation signals to reverse-engineer the system-level causal graph. The report is presented visually on the user interaction module 50: in the topology diagram representing the entire energy storage system, the module node in layer 3 is highlighted in red, with an annotation indicating an internal micro-short circuit (98.2% confidence level). Simultaneously, directed edges emanating from this node point to adjacent modules in layers 2 and 4; the edge weights indicate that the fault point had a moderate thermal conduction impact on adjacent modules. The test process ends.

[0158] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A testing method for energy storage systems based on AI cameras and AI analysis, characterized in that, The method includes the following steps: S1. Inject a set of mutually orthogonal detection excitation signals into multiple preset locations of the energy storage system; S2. A high-speed thermal imaging camera is used to collect global spatiotemporal response data characterizing the dynamic changes in the surface temperature of the energy storage system under a set of mutually orthogonal probing excitation signals. S3. Compare the global spatiotemporal response data with the preset health baseline model in the spatiotemporal dimension to identify suspicious areas where the response pattern deviates. S4. Define the amplified fault response characteristics of the suspected area as the optimization target, and generate a resonant excitation signal for the suspected area by iteratively adjusting the excitation signal parameters to solve the optimization target. S5. Inject a resonant excitation signal into the suspicious area, and use the high-speed thermal imaging camera to collect magnified response data of the suspicious area. Then, use the matching relationship between the magnified response data and the preset fault feature library to diagnose the fault type of the suspicious area.

2. The energy storage system testing method based on AI camera and AI analysis according to claim 1, characterized in that, Step S1, which involves injecting a set of mutually orthogonal detection excitation signals into multiple preset locations of the energy storage system, includes: A pseudo-random sequence modulation method is used to modulate multiple different sinusoidal carrier signals to generate a set of mutually orthogonal detection excitation signals; The pseudo-random sequence modulation method uses m-sequences or Gold codes; The generated probe excitation signal has broadband spectral characteristics.

3. The energy storage system testing method based on AI camera and AI analysis according to claim 1, characterized in that, Step S2, which involves using a high-speed thermal imaging camera to acquire global spatiotemporal response data characterizing the dynamic changes in the surface temperature of the energy storage system under a set of mutually orthogonal probing excitation signals, includes the following steps: The AI ​​analysis and processing module serves as the master clock source, sending synchronization trigger commands to the distributed excitation network and the high-speed thermal imaging camera to align the start time of excitation injection with the start time of image acquisition. Each frame of two-dimensional thermal image acquired by the high-speed thermal imaging camera is appended with a timestamp relative to the master clock source; A series of two-dimensional thermal images with the timestamps are combined in chronological order to form global spatiotemporal response data, which is a three-dimensional data structure.

4. The energy storage system testing method based on AI camera and AI analysis according to claim 1, characterized in that, Step S3, which involves comparing the global spatiotemporal response data with a preset health baseline model in the spatiotemporal dimension to identify suspicious areas where the response pattern deviates, includes: For any spatial coordinate point (x, y) in the global spatiotemporal response data, extract the response time series vector v(x, y); Based on the response time series vector v(x,y), calculate the baseline time series vector v(x,y) of the corresponding coordinate point in the health baseline model. base The deviation between (x, y) is used to obtain the outlier score A(x, y), and the formula for calculating the outlier score A(x, y) is as follows: A(x,y)=||v(x,y)-v base (x,y)||2; In the formula, A(x,y) is the anomaly score at the spatial coordinate point (x,y); v(x,y) is the response time series vector collected at the spatial coordinate point (x,y); v base (x, y) is the baseline time series vector corresponding to the spatial coordinate point (x, y) in the health baseline model; ||·||2 is the Euclidean norm operation; Based on the anomaly scores of all spatial coordinate points, a two-dimensional anomaly score map is generated. The set of spatial coordinate points where the abnormal score A(x, y) is greater than the preset abnormal threshold is determined as the suspicious area.

5. The energy storage system testing method based on AI camera and AI analysis according to claim 1, characterized in that, In step S4, the step of defining the amplified fault response characteristics of the suspected region as the optimization objective, and generating a resonant excitation signal for the suspected region by iteratively adjusting the excitation signal parameters to solve for the optimization objective includes: The optimization objective is to solve for the optimal excitation signal S. opt The optimization problem is defined by the following formula: In the formula, S opt Ω is the optimal excitation signal; S is the excitation signal to be optimized; S For the physically permissible excitation signal space; R ROI (S) represents the amplified response data of the suspicious region under the action of excitation signal S; J(R) ROI (S) is the objective function used to quantify and amplify the significance of fault modes in response data; The optimization problem is solved iteratively using a Bayesian optimization algorithm to obtain the resonant excitation signal.

6. The energy storage system testing method based on AI camera and AI analysis according to claim 1, characterized in that, Step S5, which involves injecting a resonant excitation signal into the suspected area and using the high-speed thermal imaging camera to acquire magnified response data of the suspected area, and then using the matching relationship between the magnified response data and a preset fault feature library to diagnose the fault type of the suspected area, includes the following steps: The field of view of the high-speed thermal imaging camera is focused on the suspicious area to acquire the magnified response data as a three-dimensional data tensor; The amplified response data is input into a deep convolutional neural network model trained offline; The deep convolutional neural network model outputs a confidence score vector based on the spatiotemporal features of the extracted amplified response data. The confidence score vector contains confidence scores corresponding to multiple preset fault types. The fault type corresponding to the highest confidence score in the confidence score vector is determined as the diagnosed fault type.

7. The energy storage system testing method based on AI camera and AI analysis according to claim 1, characterized in that, The method further includes the step of generating a system-level causal analysis report, the steps of which include: The global spatiotemporal response data at any spatial coordinate point is modeled as the result of the linear superposition of all exploration excitation signals after they have passed through their respective propagation paths; Based on the mutual orthogonality of the exploration excitation signals, a blind source separation algorithm is applied to the global spatiotemporal response data to separate the response components independently triggered by each exploration excitation signal. The set of spatiotemporal impulse response functions is obtained by reverse engineering. Each function in the set defines the heat propagation characteristics from the injection location of the probe excitation signal to any spatial coordinate point on the surface of the energy storage system.

8. The energy storage system testing method based on AI camera and AI analysis according to claim 7, characterized in that, Using the solved set of spatiotemporal impulse response functions as the basis for calculation, the causal influence between any two components in the energy storage system is quantified, and based on the causal influence measurement and a preset causal association threshold, a system-level causal graph characterizing the causal influence path between components within the energy storage system is constructed.

9. A testing method for an energy storage system based on an AI camera and AI analysis according to claim 8, characterized in that, The fault types of the diagnosed suspicious areas are used as node attributes and integrated with the constructed system-level causal graph to generate an integrated diagnostic report that includes both local fault information and fault propagation paths.

10. A test system for an energy storage system based on an AI camera and AI analysis, applied to the method described in any one of claims 1-9, characterized in that, The system includes: A distributed excitation network is used to inject a set of mutually orthogonal probing excitation signals into multiple preset locations of the energy storage system, and to inject resonant excitation signals into identified suspicious areas according to instructions. The global spatiotemporal sensing array includes a high-speed thermal imaging camera. The global spatiotemporal sensing array is used to acquire global spatiotemporal response data under the action of a set of mutually orthogonal probing excitation signals, as well as amplified response data of the suspicious area under the action of the resonant excitation signal. The AI ​​analysis and processing module is used to compare the global spatiotemporal response data with a preset health baseline model to identify the suspicious areas; to define the amplified fault response characteristics of the suspicious areas as optimization targets, to generate the resonant excitation signal through iterative solution, and to instruct the distributed excitation network to inject it; and to diagnose the fault type of the suspicious areas by using the matching relationship between the amplified response data and the preset fault feature library, and to generate an integrated diagnostic report. The data storage module is used to store the global spatiotemporal response data, amplified response data, and integrated diagnostic reports. The user interaction module is used to present the integrated diagnostic report in a visual manner.

Images